Add simple property test example

Co-authored-by: Tyler Neely <tyler.neely@memgraph.io>

src/common/errors.hpp

@@ -30,6 +30,8 @@ enum class ErrorCode : uint8_t {
   SCHEMA_VERTEX_UPDATE_PRIMARY_LABEL,
   SCHEMA_VERTEX_SECONDARY_LABEL_IS_PRIMARY,
   SCHEMA_VERTEX_PRIMARY_PROPERTIES_UNDEFINED,
+  // Distributed race conditions
+  STALE_SHARD_MAP,
 
   OBJECT_NOT_FOUND,
 };
@@ -62,6 +64,8 @@ constexpr std::string_view ErrorCodeToString(const ErrorCode code) {
       return "SCHEMA_VERTEX_PRIMARY_PROPERTIES_UNDEFINED";
     case ErrorCode::OBJECT_NOT_FOUND:
       return "OBJECT_NOT_FOUND";
+    case ErrorCode::STALE_SHARD_MAP:
+      return "STALE_SHARD_MAP";
   }
 }
 

src/coordinator/coordinator.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -9,21 +9,272 @@
 // by the Apache License, Version 2.0, included in the file
 // licenses/APL.txt.
 
-#include <coordinator/coordinator.hpp>
+#include "coordinator/coordinator.hpp"
+#include "query/v2/requests.hpp"
+#include "storage/v3/value_conversions.hpp"
 
 namespace memgraph::coordinator {
 
+// 1. try to begin any suggested splits
+// 2. mark all initialized RSMs as INITIALIZED in the ShardMap
+// 3. assign any valid underreplicated shards to the Heartbeat sender
+// 4. send any split requests that the Heartbeat sender should be applying
 CoordinatorWriteResponses Coordinator::ApplyWrite(HeartbeatRequest &&heartbeat_request) {
-  spdlog::info("Coordinator handling HeartbeatRequest");
+  spdlog::trace("Coordinator handling HeartbeatRequest");
 
-  // add this storage engine to any under-replicated shards that it is not already a part of
+  HeartbeatResponse ret{};
 
-  auto initializing_rsms_for_shard_manager =
-      shard_map_.AssignShards(heartbeat_request.from_storage_manager, heartbeat_request.initialized_rsms);
+  bool initiated_split = false;
 
-  return HeartbeatResponse{
-      .shards_to_initialize = initializing_rsms_for_shard_manager,
-  };
+  // 1. try to begin any suggested splits
+  for (const auto &suggested_split_info : heartbeat_request.suggested_splits) {
+    const LabelId label_id = suggested_split_info.label_id;
+    auto &label_space = shard_map_.label_spaces.at(label_id);
+
+    const PrimaryKey splitting_shard_low_key =
+        storage::conversions::ConvertPropertyVector(suggested_split_info.splitting_shard_low_key);
+    const PrimaryKey split_key = storage::conversions::ConvertPropertyVector(suggested_split_info.split_key);
+
+    auto &shard = label_space.shards.at(splitting_shard_low_key);
+    const ShardId splitting_shard_id = std::make_pair(label_id, splitting_shard_low_key);
+
+    if (shard.pending_split.has_value() || shard.version != suggested_split_info.shard_version) {
+      spdlog::debug("Coordinator skipping split, already splitting: {}, shard.version: {}, suggested shard_version: {}",
+                    shard.pending_split.has_value(), shard.version, suggested_split_info.shard_version);
+      continue;
+    }
+
+    MG_ASSERT(!label_space.shards.contains(split_key));
+
+    // begin the split process for this shard
+    const auto new_uuid_lhs = shard_map_.GetHlc();
+    const auto new_uuid_rhs = shard_map_.GetHlc();
+    spdlog::debug(
+        "Coordinator beginning new split process for shard {} after receiving a pending split. splitting into lhs: {} "
+        "and rhs: {} at split key {}",
+        shard.version.logical_id, new_uuid_lhs.logical_id, new_uuid_rhs.logical_id, split_key.back());
+
+    // bump current shard version and store pending split info
+    shard.version = new_uuid_lhs;
+    shard.pending_split = suggested_split_info;
+    MG_ASSERT(!splitting_shards_.contains(splitting_shard_id));
+    splitting_shards_.insert(splitting_shard_id);
+    initiated_split = true;
+
+    // copy this shard and store it in the ShardMap
+    ShardMetadata duplicated_shard{shard};
+    duplicated_shard.version = new_uuid_rhs;
+    duplicated_shard.pending_split.reset();
+    std::map<boost::uuids::uuid, boost::uuids::uuid> split_mapping = {};
+    for (auto &peer_metadata : duplicated_shard.peers) {
+      peer_metadata.status = Status::PENDING_SPLIT;
+      peer_metadata.split_from = peer_metadata.address.unique_id;
+
+      const auto new_uuid = shard_map_.NewShardUuid();
+
+      spdlog::debug("Coordinator allocating new rsm uuid: {} for new shard {} with low key {}", new_uuid, new_uuid_rhs,
+                    split_key.back());
+
+      // store new uuid for the right side of each shard
+      rsm_split_from_.insert({new_uuid, splitting_shard_id});
+      split_mapping.emplace(peer_metadata.address.unique_id, new_uuid);
+
+      peer_metadata.address.unique_id = new_uuid;
+    }
+
+    // TODO(tyler) fix this good assertion
+    // if (high_key) {
+    //   MG_ASSERT(converted_pk < *high_key, "Split point is beyond low key of the next shard");
+    // }
+    label_space.shards.insert({split_key, duplicated_shard});
+
+    // we will send the heartbeater its split request in stage 4 below
+  }
+
+  // 2. mark all initialized RSMs as INITIALIZED in the ShardMap
+  for (const auto &[initialized_rsm, shard_id] : heartbeat_request.initialized_rsms) {
+    auto [label_id, low_key] = shard_id;
+    auto &label_space = shard_map_.label_spaces.at(label_id);
+    auto &shard = label_space.shards.at(low_key);
+
+    // if even a single shard has been initialized after a split, its raft log has
+    // reached consensus and we can allow more splits again and remove the pending
+    // split state.
+    if (rsm_split_from_.contains(initialized_rsm)) {
+      auto split_shard_id = rsm_split_from_.at(initialized_rsm);
+
+      auto [split_label_id, split_low_key] = split_shard_id;
+      auto &split_shard = label_space.shards.at(split_low_key);
+
+      spdlog::debug("Coordinator clearing split for shard {}", split_shard.version);
+      MG_ASSERT(split_shard.pending_split.has_value());
+      MG_ASSERT(splitting_shards_.contains(split_shard_id));
+
+      split_shard.pending_split.reset();
+      splitting_shards_.erase(split_shard_id);
+      rsm_split_from_.erase(initialized_rsm);
+    }
+
+    size_t initialized_count = 0;
+    bool found = false;
+    for (auto &peer : shard.peers) {
+      if (peer.address.last_known_ip == heartbeat_request.from_storage_manager.last_known_ip &&
+          peer.address.last_known_port == heartbeat_request.from_storage_manager.last_known_port) {
+        // TODO(tyler) switch the conditional to match on rsm uuid and update the peer last_known_* address always
+        const int low_key_int = low_key[0].ValueInt();
+        // TODO(tyler) this is failing in simulation tests
+        spdlog::debug("Coordinator marking rsm {} for shard {} as initialized", initialized_rsm, shard.version);
+        MG_ASSERT(peer.address.unique_id == initialized_rsm,
+                  "expected Coordinator uuid {} to equal peer uuid {} for shard version {} with low key {}",
+                  peer.address.unique_id, initialized_rsm, shard.version, low_key_int);
+        peer.status = Status::CONSENSUS_PARTICIPANT;
+        ret.acknowledged_initialized_rsms.push_back(initialized_rsm);
+        found = true;
+      }
+
+      if (peer.status == Status::CONSENSUS_PARTICIPANT) {
+        initialized_count++;
+      }
+    }
+    MG_ASSERT(shard.peers.size() == 1);
+
+    MG_ASSERT(found, "did not find peer {} in shard version {}", heartbeat_request.from_storage_manager.unique_id,
+              shard.version);
+
+    if (initialized_count >= label_space.replication_factor) {
+      spdlog::debug("Coordinator clearing underreplicated shard with {} initialized peers", initialized_count);
+      underreplicated_shards_.erase(shard_id);
+    }
+  }
+
+  // 3. assign any valid underreplicated shards to the Heartbeat sender
+  for (auto &underreplicated_shard_id : underreplicated_shards_) {
+    auto [label_id, low_key] = underreplicated_shard_id;
+    auto &label_space = shard_map_.label_spaces.at(label_id);
+    auto &shard = label_space.shards.at(low_key);
+
+    bool needs_to_initialize = true;
+
+    // make sure we're not already a member
+    for (const auto &peer : shard.peers) {
+      if (peer.address.last_known_ip == heartbeat_request.from_storage_manager.last_known_ip &&
+          peer.address.last_known_port == heartbeat_request.from_storage_manager.last_known_port) {
+        needs_to_initialize = false;
+
+        if (peer.status == Status::INITIALIZING) {
+          spdlog::trace(
+              "Coordinator reminding ShardManager to initialize a shard that we have previously assigned it to");
+          Address address = heartbeat_request.from_storage_manager;
+          address.unique_id = peer.address.unique_id;
+
+          std::optional<PrimaryKey> high_key;
+          auto next = std::next(label_space.shards.find(low_key));
+          if (next != label_space.shards.end()) {
+            high_key = next->first;
+          }
+
+          ret.shards_to_initialize.push_back(ShardToInitialize{
+              .new_shard_version = shard.version,
+              .uuid = address.unique_id,
+              .label_id = label_id,
+              .min_key = low_key,
+              .max_key = high_key,
+              .schema = shard_map_.schemas[label_id],
+              .config = Config{.split =
+                                   Config::Split{
+                                       .max_shard_vertex_size = label_space.split_threshold,
+                                   }},
+              .id_to_names = shard_map_.IdToNames(),
+          });
+        }
+      }
+    }
+
+    if (!needs_to_initialize) {
+      // we are already a member of this shard. continue to next underreplicated shard to consider
+      continue;
+    }
+
+    // we are not already a member, so we can be assigned to this shard
+    Address address = heartbeat_request.from_storage_manager;
+    address.unique_id = shard_map_.NewShardUuid();
+
+    std::optional<PrimaryKey> high_key;
+    auto next = std::next(label_space.shards.find(low_key));
+    if (next != label_space.shards.end()) {
+      high_key = next->first;
+    }
+
+    spdlog::debug("Coordinator assigning new rsm uuid {} to shard version {} with low key {} to shard worker {}",
+                  address.unique_id, shard.version, low_key.back(), address);
+
+    ret.shards_to_initialize.push_back(ShardToInitialize{
+        .new_shard_version = shard.version,
+        .uuid = address.unique_id,
+        .label_id = label_id,
+        .min_key = low_key,
+        .max_key = high_key,
+        .schema = shard_map_.schemas[label_id],
+        .config = Config{.split =
+                             Config::Split{
+                                 .max_shard_vertex_size = label_space.split_threshold,
+                             }},
+        .id_to_names = shard_map_.IdToNames(),
+    });
+
+    PeerMetadata peer_metadata = {
+        .address = address,
+        .status = Status::INITIALIZING,
+    };
+
+    shard.peers.emplace_back(peer_metadata);
+  }
+
+  // 4. send any split requests that the Heartbeat sender should be applying
+  for (const auto &[label_id, low_key] : splitting_shards_) {
+    // see if this machine is a PENDING_SPLIT peer for any of the splitting shards
+    const auto &label_space = shard_map_.label_spaces.at(label_id);
+    const auto &splitting_shard = label_space.shards.at(low_key);
+    MG_ASSERT(splitting_shard.pending_split.has_value());
+    const auto &suggested_split_info = splitting_shard.pending_split.value();
+    const auto split_key = storage::conversions::ConvertPropertyVector(suggested_split_info.split_key);
+    const auto &new_shard = label_space.shards.at(split_key);
+
+    for (const auto &peer : new_shard.peers) {
+      if (peer.status != Status::PENDING_SPLIT ||
+          peer.address.last_known_ip != heartbeat_request.from_storage_manager.last_known_ip ||
+          peer.address.last_known_port != heartbeat_request.from_storage_manager.last_known_port) {
+        spdlog::debug("Coordinator not splitting peer: status is: {}, peer address: {} storage manager address: {}",
+                      peer.status, peer.address, heartbeat_request.from_storage_manager);
+        // not splitting or not us
+        continue;
+      }
+
+      spdlog::trace("Coordinator expecting heartbeating peer to split, so it will reply with a ShardToSplit message");
+
+      std::map<boost::uuids::uuid, boost::uuids::uuid> uuid_mapping{};
+
+      // need to iterate over all peers again to build the full uuid_mapping to
+      // send to this machine
+      for (const auto &peer_metadata2 : new_shard.peers) {
+        uuid_mapping.emplace(peer_metadata2.split_from, peer_metadata2.address.unique_id);
+      }
+
+      ret.shards_to_split.push_back(ShardToSplit{
+          .split_key = split_key,
+          .old_shard_version = suggested_split_info.shard_version,
+          .new_lhs_shard_version = splitting_shard.version,
+          .new_rhs_shard_version = new_shard.version,
+          .uuid_mapping = uuid_mapping,
+      });
+    }
+  }
+
+  if (initiated_split) {
+    MG_ASSERT(!ret.shards_to_split.empty(), "did not send a split request to the shard that suggested a split");
+  }
+
+  return ret;
 }
 
 CoordinatorWriteResponses Coordinator::ApplyWrite(HlcRequest &&hlc_request) {
@@ -37,6 +288,7 @@ CoordinatorWriteResponses Coordinator::ApplyWrite(HlcRequest &&hlc_request) {
       .logical_id = ++highest_allocated_timestamp_,
       // TODO(tyler) probably pass some more context to the Coordinator here
       // so that we can use our wall clock and enforce monotonicity.
+      // Check it to ensure it's +/- 1 day of the coordinator's io_.Now()
       // .coordinator_wall_clock = io_.Now(),
   };
 
@@ -61,22 +313,6 @@ CoordinatorWriteResponses Coordinator::ApplyWrite(AllocateEdgeIdBatchRequest &&a
   return res;
 }
 
-/// This splits the shard immediately beneath the provided
-/// split key, keeping the assigned peers identical for now,
-/// but letting them be gradually migrated over time.
-CoordinatorWriteResponses Coordinator::ApplyWrite(SplitShardRequest &&split_shard_request) {
-  SplitShardResponse res{};
-
-  if (split_shard_request.previous_shard_map_version != shard_map_.shard_map_version) {
-    res.success = false;
-  } else {
-    res.success = shard_map_.SplitShard(split_shard_request.previous_shard_map_version, split_shard_request.label_id,
-                                        split_shard_request.split_key);
-  }
-
-  return res;
-}
-
 /// This adds the provided storage engine to the standby storage engine pool,
 /// which can be used to rebalance storage over time.
 CoordinatorWriteResponses Coordinator::ApplyWrite(
@@ -102,12 +338,16 @@ CoordinatorWriteResponses Coordinator::ApplyWrite(InitializeLabelRequest &&initi
 
   std::optional<LabelId> new_label_id = shard_map_.InitializeNewLabel(
       initialize_label_request.label_name, initialize_label_request.schema, initialize_label_request.replication_factor,
-      initialize_label_request.last_shard_map_version);
+      initialize_label_request.split_threshold, initialize_label_request.last_shard_map_version);
 
   if (new_label_id) {
     res.new_label_id = new_label_id.value();
     res.fresher_shard_map = std::nullopt;
     res.success = true;
+
+    auto min_key = SchemaToMinKey(initialize_label_request.schema);
+    auto shard_id = std::make_pair(new_label_id.value(), min_key);
+    underreplicated_shards_.insert(shard_id);
   } else {
     res.fresher_shard_map = shard_map_;
     res.success = false;

src/coordinator/coordinator.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -20,13 +20,14 @@
 
 #include <boost/uuid/uuid.hpp>
 
-#include <coordinator/hybrid_logical_clock.hpp>
-#include <coordinator/shard_map.hpp>
-#include <io/simulator/simulator.hpp>
-#include <io/time.hpp>
-#include <io/transport.hpp>
-#include <storage/v3/id_types.hpp>
-#include <storage/v3/schemas.hpp>
+#include "coordinator/hybrid_logical_clock.hpp"
+#include "coordinator/shard_map.hpp"
+#include "io/simulator/simulator.hpp"
+#include "io/time.hpp"
+#include "io/transport.hpp"
+#include "query/v2/requests.hpp"
+#include "storage/v3/id_types.hpp"
+#include "storage/v3/schemas.hpp"
 
 namespace memgraph::coordinator {
 
@@ -37,6 +38,8 @@ using memgraph::storage::v3::SchemaProperty;
 using SimT = memgraph::io::simulator::SimulatorTransport;
 using PrimaryKey = std::vector<PropertyValue>;
 
+using ShardId = std::pair<LabelId, PrimaryKey>;
+
 struct HlcRequest {
   Hlc last_shard_map_version;
 };
@@ -112,6 +115,7 @@ struct InitializeLabelRequest {
   std::string label_name;
   std::vector<SchemaProperty> schema;
   size_t replication_factor;
+  uint64_t split_threshold;
   Hlc last_shard_map_version;
 };
 
@@ -121,28 +125,29 @@ struct InitializeLabelResponse {
   std::optional<ShardMap> fresher_shard_map;
 };
 
-struct HeartbeatRequest {
-  Address from_storage_manager;
-  std::set<boost::uuids::uuid> initialized_rsms;
-};
-
-struct HeartbeatResponse {
-  std::vector<ShardToInitialize> shards_to_initialize;
-};
-
 using CoordinatorWriteRequests =
-    std::variant<HlcRequest, AllocateEdgeIdBatchRequest, SplitShardRequest, RegisterStorageEngineRequest,
-                 DeregisterStorageEngineRequest, InitializeLabelRequest, AllocatePropertyIdsRequest, HeartbeatRequest>;
-using CoordinatorWriteResponses = std::variant<HlcResponse, AllocateEdgeIdBatchResponse, SplitShardResponse,
-                                               RegisterStorageEngineResponse, DeregisterStorageEngineResponse,
-                                               InitializeLabelResponse, AllocatePropertyIdsResponse, HeartbeatResponse>;
+    std::variant<HlcRequest, AllocateEdgeIdBatchRequest, RegisterStorageEngineRequest, DeregisterStorageEngineRequest,
+                 InitializeLabelRequest, AllocatePropertyIdsRequest, HeartbeatRequest>;
+using CoordinatorWriteResponses = std::variant<HlcResponse, AllocateEdgeIdBatchResponse, RegisterStorageEngineResponse,
+                                               DeregisterStorageEngineResponse, InitializeLabelResponse,
+                                               AllocatePropertyIdsResponse, HeartbeatResponse>;
 
 using CoordinatorReadRequests = std::variant<GetShardMapRequest>;
 using CoordinatorReadResponses = std::variant<GetShardMapResponse>;
 
 class Coordinator {
  public:
-  explicit Coordinator(ShardMap sm) : shard_map_{std::move(sm)} {}
+  explicit Coordinator(ShardMap sm) : shard_map_{std::move(sm)} {
+    // Populate underreplicated_shards_
+    for (const auto &[label_id, label_space] : shard_map_.label_spaces) {
+      for (const auto &[low_key, shard] : label_space.shards) {
+        if (shard.peers.size() < label_space.replication_factor) {
+          ShardId shard_id = std::make_pair(label_id, low_key);
+          underreplicated_shards_.insert(shard_id);
+        }
+      }
+    }
+  }
 
   // NOLINTNEXTLINE(readability-convert-member-functions-to-static
   CoordinatorReadResponses Read(CoordinatorReadRequests requests) {
@@ -160,6 +165,11 @@ class Coordinator {
   ShardMap shard_map_;
   uint64_t highest_allocated_timestamp_{0};
 
+  std::set<ShardId> underreplicated_shards_;
+  std::map<boost::uuids::uuid, ShardId> rsm_split_from_;
+  std::set<ShardId> splitting_shards_;
+  std::map<Address, std::set<ShardId>> assigned_shards_;
+
   /// Query engines need to periodically request batches of unique edge IDs.
   uint64_t highest_allocated_edge_id_{0};
 
@@ -175,11 +185,6 @@ class Coordinator {
 
   CoordinatorWriteResponses ApplyWrite(AllocateEdgeIdBatchRequest &&ahr);
 
-  /// This splits the shard immediately beneath the provided
-  /// split key, keeping the assigned peers identical for now,
-  /// but letting them be gradually migrated over time.
-  CoordinatorWriteResponses ApplyWrite(SplitShardRequest &&split_shard_request);
-
   /// This adds the provided storage engine to the standby storage engine pool,
   /// which can be used to rebalance storage over time.
   static CoordinatorWriteResponses ApplyWrite(RegisterStorageEngineRequest && /* register_storage_engine_request */);

src/coordinator/hybrid_logical_clock.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -23,17 +23,19 @@ namespace memgraph::coordinator {
 using Time = memgraph::io::Time;
 
 /// Hybrid-logical clock
-struct Hlc {
-  uint64_t logical_id = 0;
+struct Hlc final {
+  uint64_t logical_id{0};
   Time coordinator_wall_clock = Time::min();
 
   auto operator<=>(const Hlc &other) const { return logical_id <=> other.logical_id; }
 
-  bool operator==(const Hlc &other) const = default;
-  bool operator<(const Hlc &other) const = default;
-  bool operator==(const uint64_t other) const { return logical_id == other; }
-  bool operator<(const uint64_t other) const { return logical_id < other; }
-  bool operator>=(const uint64_t other) const { return logical_id >= other; }
+  bool operator==(const Hlc &other) const noexcept = default;
+  bool operator<(const Hlc &other) const noexcept = default;
+  bool operator==(const uint64_t other) const noexcept { return logical_id == other; }
+  bool operator<(const uint64_t other) const noexcept { return logical_id < other; }
+  bool operator>=(const uint64_t other) const noexcept { return logical_id >= other; }
+
+  Hlc operator++() { return {.logical_id = ++logical_id, .coordinator_wall_clock = Time::min()}; }
 
   friend std::ostream &operator<<(std::ostream &in, const Hlc &hlc) {
     auto wall_clock = std::chrono::system_clock::to_time_t(hlc.coordinator_wall_clock);

src/coordinator/shard_map.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -9,15 +9,21 @@
 // by the Apache License, Version 2.0, included in the file
 // licenses/APL.txt.
 
+#include <boost/uuid/uuid.hpp>
+#include <map>
 #include <optional>
 #include <unordered_map>
 #include <vector>
 
 #include "common/types.hpp"
+#include "coordinator/hybrid_logical_clock.hpp"
 #include "coordinator/shard_map.hpp"
+#include "query/v2/requests.hpp"
 #include "spdlog/spdlog.h"
+#include "storage/v3/config.hpp"
 #include "storage/v3/schemas.hpp"
 #include "storage/v3/temporal.hpp"
+#include "storage/v3/value_conversions.hpp"
 #include "utils/cast.hpp"
 #include "utils/exceptions.hpp"
 #include "utils/string.hpp"
@@ -68,6 +74,9 @@ PrimaryKey SchemaToMinKey(const std::vector<SchemaProperty> &schema) {
 }
 
 ShardMap ShardMap::Parse(std::istream &input_stream) {
+  const uint64_t default_replication_factor = 1;
+  const uint64_t default_split_threshold = 1'000'000;
+
   ShardMap shard_map;
   const auto read_size = [&input_stream] {
     size_t size{0};
@@ -84,23 +93,17 @@ ShardMap ShardMap::Parse(std::istream &input_stream) {
 
   const auto read_names = [&read_size, &read_word] {
     const auto number_of_names = read_size();
-    spdlog::trace("Reading {} names", number_of_names);
+    spdlog::trace("ShardMap::Parse reading {} names", number_of_names);
     std::vector<std::string> names;
     names.reserve(number_of_names);
 
     for (auto name_index = 0; name_index < number_of_names; ++name_index) {
       names.push_back(read_word());
-      spdlog::trace("Read '{}'", names.back());
+      spdlog::trace("ShardMap::Parse read '{}'", names.back());
     }
     return names;
   };
 
-  const auto read_line = [&input_stream] {
-    std::string line;
-    std::getline(input_stream, line);
-    return line;
-  };
-
   const auto parse_type = [](const std::string &type) {
     static const auto type_map = std::unordered_map<std::string, common::SchemaType>{
         {"string", common::SchemaType::STRING}, {"int", common::SchemaType::INT}, {"bool", common::SchemaType::BOOL}};
@@ -110,45 +113,32 @@ ShardMap ShardMap::Parse(std::istream &input_stream) {
     return it->second;
   };
 
-  const auto parse_property_value = [](std::string text, const common::SchemaType type) {
-    if (type == common::SchemaType::STRING) {
-      return storage::v3::PropertyValue{std::move(text)};
-    }
-    if (type == common::SchemaType::INT) {
-      size_t processed{0};
-      int64_t value = std::stoll(text, &processed);
-      MG_ASSERT(processed == text.size() || text[processed] == ' ', "Invalid integer format: '{}'", text);
-      return storage::v3::PropertyValue{value};
-    }
-    LOG_FATAL("Not supported type: {}", utils::UnderlyingCast(type));
-  };
-
-  spdlog::debug("Reading properties");
+  spdlog::trace("ShardMap::Parse reading properties");
   const auto properties = read_names();
   MG_ASSERT(shard_map.AllocatePropertyIds(properties).size() == properties.size(),
             "Unexpected number of properties created!");
 
-  spdlog::debug("Reading edge types");
+  spdlog::trace("ShardMap::Parse Reading edge types");
   const auto edge_types = read_names();
   MG_ASSERT(shard_map.AllocateEdgeTypeIds(edge_types).size() == edge_types.size(),
             "Unexpected number of properties created!");
 
-  spdlog::debug("Reading primary labels");
+  spdlog::trace("ShardMap::Parse reading primary labels");
   const auto number_of_primary_labels = read_size();
-  spdlog::debug("Reading {} primary labels", number_of_primary_labels);
+  spdlog::trace("ShardMap::Parse reading {} primary labels", number_of_primary_labels);
 
   for (auto label_index = 0; label_index < number_of_primary_labels; ++label_index) {
     const auto primary_label = read_word();
-    spdlog::debug("Reading primary label named '{}'", primary_label);
+    spdlog::trace("ShardMap::Parse reading primary label named '{}'", primary_label);
     const auto number_of_primary_properties = read_size();
-    spdlog::debug("Reading {} primary properties", number_of_primary_properties);
+    spdlog::trace("ShardMap::Parse reading {} primary properties", number_of_primary_properties);
     std::vector<std::string> pp_names;
     std::vector<common::SchemaType> pp_types;
     pp_names.reserve(number_of_primary_properties);
     pp_types.reserve(number_of_primary_properties);
     for (auto property_index = 0; property_index < number_of_primary_properties; ++property_index) {
       pp_names.push_back(read_word());
-      spdlog::debug("Reading primary property named '{}'", pp_names.back());
+      spdlog::trace("ShardMap::Parse reading primary property named '{}'", pp_names.back());
       pp_types.push_back(parse_type(read_word()));
     }
     auto pp_mapping = shard_map.AllocatePropertyIds(pp_names);
@@ -159,33 +149,10 @@ ShardMap ShardMap::Parse(std::istream &input_stream) {
       schema.push_back(storage::v3::SchemaProperty{pp_mapping.at(pp_names[property_index]), pp_types[property_index]});
     }
     const auto hlc = shard_map.GetHlc();
-    MG_ASSERT(shard_map.InitializeNewLabel(primary_label, schema, 1, hlc).has_value(),
-              "Cannot initialize new label: {}", primary_label);
-
-    const auto number_of_split_points = read_size();
-    spdlog::debug("Reading {} split points", number_of_split_points);
-
-    [[maybe_unused]] const auto remainder_from_last_line = read_line();
-    for (auto split_point_index = 0; split_point_index < number_of_split_points; ++split_point_index) {
-      const auto line = read_line();
-      spdlog::debug("Read split point '{}'", line);
-      MG_ASSERT(line.front() == '[', "Invalid split file format!");
-      MG_ASSERT(line.back() == ']', "Invalid split file format!");
-      std::string_view line_view{line};
-      line_view.remove_prefix(1);
-      line_view.remove_suffix(1);
-      static constexpr std::string_view kDelimiter{","};
-      auto pk_values_as_text = utils::Split(line_view, kDelimiter);
-      std::vector<PropertyValue> pk;
-      pk.reserve(number_of_primary_properties);
-      MG_ASSERT(pk_values_as_text.size() == number_of_primary_properties,
-                "Split point contains invalid number of values '{}'", line);
-
-      for (auto property_index = 0; property_index < number_of_primary_properties; ++property_index) {
-        pk.push_back(parse_property_value(std::move(pk_values_as_text[property_index]), schema[property_index].type));
-      }
-      shard_map.SplitShard(shard_map.GetHlc(), shard_map.labels.at(primary_label), pk);
-    }
+    MG_ASSERT(
+        shard_map.InitializeNewLabel(primary_label, schema, default_replication_factor, default_split_threshold, hlc)
+            .has_value(),
+        "Cannot initialize new label: {}", primary_label);
   }
 
   return shard_map;
@@ -246,9 +213,7 @@ std::unordered_map<uint64_t, std::string> ShardMap::IdToNames() {
   return id_to_names;
 }
 
-Hlc ShardMap::GetHlc() const noexcept { return shard_map_version; }
-
-boost::uuids::uuid NewShardUuid(uint64_t shard_id) {
+boost::uuids::uuid Uint64ToUuid(uint64_t u) {
   return boost::uuids::uuid{0,
                             0,
                             0,
@@ -257,124 +222,28 @@ boost::uuids::uuid NewShardUuid(uint64_t shard_id) {
                             0,
                             0,
                             0,
-                            static_cast<unsigned char>(shard_id >> 56U),
-                            static_cast<unsigned char>(shard_id >> 48U),
-                            static_cast<unsigned char>(shard_id >> 40U),
-                            static_cast<unsigned char>(shard_id >> 32U),
-                            static_cast<unsigned char>(shard_id >> 24U),
-                            static_cast<unsigned char>(shard_id >> 16U),
-                            static_cast<unsigned char>(shard_id >> 8U),
-                            static_cast<unsigned char>(shard_id)};
+                            static_cast<unsigned char>(u >> 56U),
+                            static_cast<unsigned char>(u >> 48U),
+                            static_cast<unsigned char>(u >> 40U),
+                            static_cast<unsigned char>(u >> 32U),
+                            static_cast<unsigned char>(u >> 24U),
+                            static_cast<unsigned char>(u >> 16U),
+                            static_cast<unsigned char>(u >> 8U),
+                            static_cast<unsigned char>(u)};
 }
 
-std::vector<ShardToInitialize> ShardMap::AssignShards(Address storage_manager,
-                                                      std::set<boost::uuids::uuid> initialized) {
-  std::vector<ShardToInitialize> ret{};
-
-  bool mutated = false;
-
-  for (auto &[label_id, label_space] : label_spaces) {
-    for (auto it = label_space.shards.begin(); it != label_space.shards.end(); it++) {
-      auto &[low_key, shard] = *it;
-      std::optional<PrimaryKey> high_key;
-      if (const auto next_it = std::next(it); next_it != label_space.shards.end()) {
-        high_key = next_it->first;
-      }
-      // TODO(tyler) avoid these triple-nested loops by having the heartbeat include better info
-      bool machine_contains_shard = false;
-
-      for (auto &aas : shard.peers) {
-        if (initialized.contains(aas.address.unique_id)) {
-          machine_contains_shard = true;
-          if (aas.status != Status::CONSENSUS_PARTICIPANT) {
-            mutated = true;
-            spdlog::info("marking shard as full consensus participant: {}", aas.address.unique_id);
-            aas.status = Status::CONSENSUS_PARTICIPANT;
-          }
-        } else {
-          const bool same_machine = aas.address.last_known_ip == storage_manager.last_known_ip &&
-                                    aas.address.last_known_port == storage_manager.last_known_port;
-          if (same_machine) {
-            machine_contains_shard = true;
-            spdlog::info("reminding shard manager that they should begin participating in shard");
-
-            ret.push_back(ShardToInitialize{
-                .uuid = aas.address.unique_id,
-                .label_id = label_id,
-                .min_key = low_key,
-                .max_key = high_key,
-                .schema = schemas[label_id],
-                .config = Config{},
-                .id_to_names = IdToNames(),
-            });
-          }
-        }
-      }
-
-      if (!machine_contains_shard && shard.peers.size() < label_space.replication_factor) {
-        // increment version for each new uuid for deterministic creation
-        IncrementShardMapVersion();
-
-        Address address = storage_manager;
-
-        // TODO(tyler) use deterministic UUID so that coordinators don't diverge here
-        address.unique_id = NewShardUuid(shard_map_version.logical_id);
-
-        spdlog::info("assigning shard manager to shard");
-
-        ret.push_back(ShardToInitialize{
-            .uuid = address.unique_id,
-            .label_id = label_id,
-            .min_key = low_key,
-            .max_key = high_key,
-            .schema = schemas[label_id],
-            .config = Config{},
-            .id_to_names = IdToNames(),
-        });
-
-        AddressAndStatus aas = {
-            .address = address,
-            .status = Status::INITIALIZING,
-        };
-
-        shard.peers.emplace_back(aas);
-      }
-    }
-  }
-
-  if (mutated) {
-    IncrementShardMapVersion();
-  }
-  return ret;
+boost::uuids::uuid ShardMap::NewShardUuid() {
+  uint64_t shard_id = GetHlc().logical_id;
+  return Uint64ToUuid(shard_id);
 }
 
-bool ShardMap::SplitShard(Hlc previous_shard_map_version, LabelId label_id, const PrimaryKey &key) {
-  if (previous_shard_map_version != shard_map_version) {
-    return false;
-  }
-
-  auto &label_space = label_spaces.at(label_id);
-  auto &shards_in_map = label_space.shards;
-
-  MG_ASSERT(!shards_in_map.empty());
-  MG_ASSERT(!shards_in_map.contains(key));
-  MG_ASSERT(label_spaces.contains(label_id));
-
-  // Finding the ShardMetadata that the new PrimaryKey should map to.
-  auto prev = std::prev(shards_in_map.upper_bound(key));
-  ShardMetadata duplicated_shard = prev->second;
-
-  // Apply the split
-  shards_in_map[key] = duplicated_shard;
-
-  IncrementShardMapVersion();
-
-  return true;
-}
+Hlc ShardMap::GetHlc() noexcept { return ++shard_map_version; }
 
 std::optional<LabelId> ShardMap::InitializeNewLabel(std::string label_name, std::vector<SchemaProperty> schema,
-                                                    size_t replication_factor, Hlc last_shard_map_version) {
+                                                    size_t replication_factor, uint64_t split_threshold,
+                                                    Hlc last_shard_map_version) {
   if (shard_map_version != last_shard_map_version || labels.contains(label_name)) {
+    MG_ASSERT(false, "failed to InitializeNewLabel");
     return std::nullopt;
   }
 
@@ -393,7 +262,9 @@ std::optional<LabelId> ShardMap::InitializeNewLabel(std::string label_name, std:
       .schema = schema,
       .shards = shards,
       .replication_factor = replication_factor,
+      .split_threshold = split_threshold,
   };
+
   schemas[label_id] = std::move(schema);
 
   label_spaces.emplace(label_id, label_space);
@@ -557,13 +428,13 @@ bool ShardMap::ClusterInitialized() const {
   for (const auto &[label_id, label_space] : label_spaces) {
     for (const auto &[low_key, shard] : label_space.shards) {
       if (shard.peers.size() < label_space.replication_factor) {
-        spdlog::info("label_space below desired replication factor");
+        spdlog::trace("ShardMap::ClusterInitialized label_space below desired replication factor");
         return false;
       }
 
-      for (const auto &aas : shard.peers) {
-        if (aas.status != Status::CONSENSUS_PARTICIPANT) {
-          spdlog::info("shard member not yet a CONSENSUS_PARTICIPANT");
+      for (const auto &peer_metadata : shard.peers) {
+        if (peer_metadata.status != Status::CONSENSUS_PARTICIPANT) {
+          spdlog::trace("ShardMap::ClusterInitialized shard member not yet a CONSENSUS_PARTICIPANT");
           return false;
         }
       }
@@ -573,4 +444,29 @@ bool ShardMap::ClusterInitialized() const {
   return true;
 }
 
+size_t ShardMap::InitializedShards() const {
+  size_t count = 0;
+
+  for (const auto &[label_id, label_space] : label_spaces) {
+    for (const auto &[low_key, shard] : label_space.shards) {
+      if (shard.peers.size() < label_space.replication_factor) {
+        continue;
+      }
+
+      bool all_initialized = true;
+      for (const auto &peer_metadata : shard.peers) {
+        if (peer_metadata.status != Status::CONSENSUS_PARTICIPANT) {
+          all_initialized = false;
+        }
+      }
+
+      if (all_initialized) {
+        count += 1;
+      }
+    }
+  }
+
+  return count;
+}
+
 }  // namespace memgraph::coordinator

src/coordinator/shard_map.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -14,6 +14,7 @@
 #include <algorithm>
 #include <limits>
 #include <map>
+#include <optional>
 #include <set>
 #include <vector>
 
@@ -23,6 +24,7 @@
 #include "common/types.hpp"
 #include "coordinator/hybrid_logical_clock.hpp"
 #include "io/address.hpp"
+#include "query/v2/requests.hpp"
 #include "storage/v3/config.hpp"
 #include "storage/v3/id_types.hpp"
 #include "storage/v3/name_id_mapper.hpp"
@@ -47,39 +49,59 @@ using memgraph::storage::v3::SchemaProperty;
 enum class Status : uint8_t {
   CONSENSUS_PARTICIPANT,
   INITIALIZING,
+  PENDING_SPLIT,
   // TODO(tyler) this will possibly have more states,
   // depending on the reconfiguration protocol that we
   // implement.
 };
 
-struct AddressAndStatus {
+struct PeerMetadata {
   memgraph::io::Address address;
   Status status;
+  boost::uuids::uuid split_from;
 
-  friend bool operator<(const AddressAndStatus &lhs, const AddressAndStatus &rhs) { return lhs.address < rhs.address; }
+  friend bool operator<(const PeerMetadata &lhs, const PeerMetadata &rhs) { return lhs.address < rhs.address; }
 
-  friend std::ostream &operator<<(std::ostream &in, const AddressAndStatus &address_and_status) {
-    in << "AddressAndStatus { address: ";
-    in << address_and_status.address;
-    if (address_and_status.status == Status::CONSENSUS_PARTICIPANT) {
-      in << ", status: CONSENSUS_PARTICIPANT }";
+  friend std::ostream &operator<<(std::ostream &in, const PeerMetadata &peer_metadata) {
+    in << "PeerMetadata { address: ";
+    in << peer_metadata.address;
+
+    if (peer_metadata.status == Status::CONSENSUS_PARTICIPANT) {
+      in << ", status: CONSENSUS_PARTICIPANT";
+    } else if (peer_metadata.status == Status::INITIALIZING) {
+      in << ", status: INITIALIZING";
+    } else if (peer_metadata.status == Status::PENDING_SPLIT) {
+      in << ", status: PENDING_SPLIT";
     } else {
-      in << ", status: INITIALIZING }";
+      MG_ASSERT(false, "failed to update the operator<< implementation for Status");
     }
 
+    in << ", split_from: " << peer_metadata.split_from << " }";
+
     return in;
   }
 
-  friend bool operator==(const AddressAndStatus &lhs, const AddressAndStatus &rhs) {
-    return lhs.address == rhs.address;
-  }
+  friend bool operator==(const PeerMetadata &lhs, const PeerMetadata &rhs) { return lhs.address == rhs.address; }
 };
 
 using PrimaryKey = std::vector<PropertyValue>;
+using ShardId = std::pair<LabelId, PrimaryKey>;
 
 struct ShardMetadata {
-  std::vector<AddressAndStatus> peers;
-  uint64_t version;
+  std::vector<PeerMetadata> peers;
+  Hlc version;
+  std::optional<msgs::SuggestedSplitInfo> pending_split;
+
+  bool Underreplicated(size_t replication_factor) { return peers.size() < replication_factor; }
+
+  bool ContainsPeer(io::PartialAddress machine) {
+    for (const auto &peer : peers) {
+      if (peer.address.last_known_ip == machine.ip && peer.address.last_known_port == machine.port) {
+        return true;
+      }
+    }
+    return false;
+  }
 
   friend std::ostream &operator<<(std::ostream &in, const ShardMetadata &shard) {
     using utils::print_helpers::operator<<;
@@ -93,7 +115,9 @@ struct ShardMetadata {
     return in;
   }
 
-  friend bool operator==(const ShardMetadata &lhs, const ShardMetadata &rhs) = default;
+  friend bool operator==(const ShardMetadata &lhs, const ShardMetadata &rhs) {
+    return lhs.version == rhs.version && lhs.peers == rhs.peers;
+  };
 
   friend bool operator<(const ShardMetadata &lhs, const ShardMetadata &rhs) {
     if (lhs.peers != rhs.peers) {
@@ -112,6 +136,7 @@ using PropertyMap = std::map<PropertyName, PropertyId>;
 using EdgeTypeIdMap = std::map<EdgeTypeName, EdgeTypeId>;
 
 struct ShardToInitialize {
+  Hlc new_shard_version;
   boost::uuids::uuid uuid;
   LabelId label_id;
   PrimaryKey min_key;
@@ -121,6 +146,26 @@ struct ShardToInitialize {
   std::unordered_map<uint64_t, std::string> id_to_names;
 };
 
+struct ShardToSplit {
+  PrimaryKey split_key;
+  Hlc old_shard_version;
+  Hlc new_lhs_shard_version;
+  Hlc new_rhs_shard_version;
+  std::map<boost::uuids::uuid, boost::uuids::uuid> uuid_mapping;
+};
+
+struct HeartbeatRequest {
+  Address from_storage_manager;
+  std::map<boost::uuids::uuid, ShardId> initialized_rsms;
+  std::set<msgs::SuggestedSplitInfo> suggested_splits;
+};
+
+struct HeartbeatResponse {
+  std::vector<ShardToInitialize> shards_to_initialize;
+  std::vector<ShardToSplit> shards_to_split;
+  std::vector<boost::uuids::uuid> acknowledged_initialized_rsms;
+};
+
 PrimaryKey SchemaToMinKey(const std::vector<SchemaProperty> &schema);
 
 struct LabelSpace {
@@ -128,6 +173,7 @@ struct LabelSpace {
   // Maps between the smallest primary key stored in the shard and the shard
   std::map<PrimaryKey, ShardMetadata> shards;
   size_t replication_factor;
+  uint64_t split_threshold;
 
   friend std::ostream &operator<<(std::ostream &in, const LabelSpace &label_space) {
     using utils::print_helpers::operator<<;
@@ -163,17 +209,15 @@ struct ShardMap {
   // TODO(gabor) later we will want to update the wallclock time with
   // the given Io<impl>'s time as well
   Hlc IncrementShardMapVersion() noexcept;
-  Hlc GetHlc() const noexcept;
+  Hlc GetHlc() noexcept;
 
   std::unordered_map<uint64_t, std::string> IdToNames();
 
-  // Returns the shard UUIDs that have been assigned but not yet acknowledged for this storage manager
-  std::vector<ShardToInitialize> AssignShards(Address storage_manager, std::set<boost::uuids::uuid> initialized);
-
-  bool SplitShard(Hlc previous_shard_map_version, LabelId label_id, const PrimaryKey &key);
-
   std::optional<LabelId> InitializeNewLabel(std::string label_name, std::vector<SchemaProperty> schema,
-                                            size_t replication_factor, Hlc last_shard_map_version);
+                                            size_t replication_factor, uint64_t split_threshold,
+                                            Hlc last_shard_map_version);
+
+  boost::uuids::uuid NewShardUuid();
 
   void AddServer(Address server_address);
 
@@ -199,6 +243,9 @@ struct ShardMap {
   /// the CONSENSUS_PARTICIPANT state. Note that this does not necessarily mean that
   /// there is also an active leader for each shard.
   bool ClusterInitialized() const;
+
+  /// Returns the current count of all shards that are fully initialized.
+  size_t InitializedShards() const;
 };
 
 }  // namespace memgraph::coordinator

src/expr/interpret/eval.hpp

@@ -120,6 +120,10 @@ class ExpressionEvaluator : public ExpressionVisitor<TypedValue> {
       case Error::SCHEMA_VERTEX_SECONDARY_LABEL_IS_PRIMARY:
       case Error::SCHEMA_VERTEX_PRIMARY_PROPERTIES_UNDEFINED:
         throw ExpressionRuntimeException("Unexpected schema violation when accessing {}.", accessed_object);
+      case Error::STALE_SHARD_MAP:
+        throw ExpressionRuntimeException(
+            "Cluster performed a Shard split or merge that invalidated the transaction's metadata. This should have "
+            "been handled in the RequestRouter.");
     }
   }
 

src/functions/awesome_memgraph_functions.hpp

@@ -457,6 +457,7 @@ TypedValueT Properties(const TypedValueT *args, int64_t nargs, const FunctionCon
           case common::ErrorCode::SCHEMA_VERTEX_SECONDARY_LABEL_IS_PRIMARY:
           case common::ErrorCode::SCHEMA_VERTEX_PRIMARY_PROPERTIES_UNDEFINED:
           case common::ErrorCode::OBJECT_NOT_FOUND:
+          case common::ErrorCode::STALE_SHARD_MAP:
             throw functions::FunctionRuntimeException("Unexpected error when getting properties.");
         }
       }
@@ -532,6 +533,7 @@ inline size_t UnwrapDegreeResult(storage::v3::ShardResult<size_t> maybe_degree)
       case common::ErrorCode::SCHEMA_VERTEX_SECONDARY_LABEL_IS_PRIMARY:
       case common::ErrorCode::SCHEMA_VERTEX_PRIMARY_PROPERTIES_UNDEFINED:
       case common::ErrorCode::OBJECT_NOT_FOUND:
+      case common::ErrorCode::STALE_SHARD_MAP:
         throw functions::FunctionRuntimeException("Unexpected error when getting node degree.");
     }
   }
@@ -719,6 +721,7 @@ TypedValueT Labels(const TypedValueT *args, int64_t nargs, const FunctionContext
         case common::ErrorCode::SCHEMA_VERTEX_SECONDARY_LABEL_IS_PRIMARY:
         case common::ErrorCode::SCHEMA_VERTEX_PRIMARY_PROPERTIES_UNDEFINED:
         case common::ErrorCode::OBJECT_NOT_FOUND:
+        case common::ErrorCode::STALE_SHARD_MAP:
           throw functions::FunctionRuntimeException("Unexpected error when getting labels.");
       }
     }

src/io/errors.hpp

@@ -14,7 +14,12 @@
 namespace memgraph::io {
 // Signifies that a retriable operation was unable to
 // complete after a configured number of retries.
-struct RetriesExhausted {};
+struct RetriesExhausted {
+  friend std::ostream &operator<<(std::ostream &in, const RetriesExhausted & /* retries_exhausted */) {
+    in << "RetriesExhausted {}";
+    return in;
+  }
+};
 
 // Signifies that a request was unable to receive a response
 // within some configured timeout duration. It is important
@@ -22,5 +27,21 @@ struct RetriesExhausted {};
 // not signify that a request was not received or processed.
 // It may be the case that the request was fully processed
 // but that the response was not received.
-struct TimedOut {};
+struct TimedOut {
+  friend std::ostream &operator<<(std::ostream &in, const TimedOut & /* timed_out */) {
+    in << "TimedOut {}";
+    return in;
+  }
+};
+
+// This error signifies that a shard has been contacted that does
+// not match our expected shard version, and that we should retry
+// the operation so that we can ensure that we are talking to the
+// correct shard.
+struct ShardVersionMismatch {
+  friend std::ostream &operator<<(std::ostream &in, const ShardVersionMismatch & /* shard_version_mismatch */) {
+    in << "ShardVersionMismatch {}";
+    return in;
+  }
+};
 };  // namespace memgraph::io

src/io/local_transport/local_transport.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -20,6 +20,7 @@
 #include "io/local_transport/local_transport_handle.hpp"
 #include "io/time.hpp"
 #include "io/transport.hpp"
+#include "utils/concrete_msg_sender.hpp"
 
 namespace memgraph::io::local_transport {
 
@@ -30,19 +31,19 @@ class LocalTransport {
   explicit LocalTransport(std::shared_ptr<LocalTransportHandle> local_transport_handle)
       : local_transport_handle_(std::move(local_transport_handle)) {}
 
-  template <Message RequestT, Message ResponseT>
+  template <utils::Message RequestT, utils::Message ResponseT>
   ResponseFuture<ResponseT> Request(Address to_address, Address from_address, RequestT request,
                                     std::function<void()> fill_notifier, Duration timeout) {
     return local_transport_handle_->template SubmitRequest<RequestT, ResponseT>(
         to_address, from_address, std::move(request), timeout, fill_notifier);
   }
 
-  template <Message... Ms>
+  template <utils::Message... Ms>
   requires(sizeof...(Ms) > 0) RequestResult<Ms...> Receive(Address receiver_address, Duration timeout) {
     return local_transport_handle_->template Receive<Ms...>(receiver_address, timeout);
   }
 
-  template <Message M>
+  template <utils::Message M>
   void Send(Address to_address, Address from_address, RequestId request_id, M &&message) {
     return local_transport_handle_->template Send<M>(to_address, from_address, request_id, std::forward<M>(message));
   }

src/io/local_transport/local_transport_handle.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -22,6 +22,7 @@
 #include "io/message_histogram_collector.hpp"
 #include "io/time.hpp"
 #include "io/transport.hpp"
+#include "utils/concrete_msg_sender.hpp"
 
 namespace memgraph::io::local_transport {
 
@@ -30,7 +31,7 @@ class LocalTransportHandle {
   mutable std::condition_variable cv_;
   bool should_shut_down_ = false;
   MessageHistogramCollector histograms_;
-  RequestId request_id_counter_ = 0;
+  RequestId request_id_counter_ = 1;
 
   // the responses to requests that are being waited on
   std::map<PromiseKey, DeadlineAndOpaquePromise> promises_;
@@ -67,7 +68,7 @@ class LocalTransportHandle {
     return std::chrono::time_point_cast<std::chrono::microseconds>(nano_time);
   }
 
-  template <Message... Ms>
+  template <utils::Message... Ms>
   requires(sizeof...(Ms) > 0) RequestResult<Ms...> Receive(Address /* receiver_address */, Duration timeout) {
     std::unique_lock lock(mu_);
 
@@ -103,7 +104,7 @@ class LocalTransportHandle {
     return std::move(m_opt).value();
   }
 
-  template <Message M>
+  template <utils::Message M>
   void Send(Address to_address, Address from_address, RequestId request_id, M &&message) {
     auto type_info = TypeInfoFor(message);
 
@@ -138,7 +139,7 @@ class LocalTransportHandle {
     cv_.notify_all();
   }
 
-  template <Message RequestT, Message ResponseT>
+  template <utils::Message RequestT, utils::Message ResponseT>
   ResponseFuture<ResponseT> SubmitRequest(Address to_address, Address from_address, RequestT &&request,
                                           Duration timeout, std::function<void()> fill_notifier) {
     auto [future, promise] = memgraph::io::FuturePromisePairWithNotifications<ResponseResult<ResponseT>>(
@@ -155,7 +156,7 @@ class LocalTransportHandle {
     const auto now = Now();
     const Time deadline = now + timeout;
 
-    RequestId request_id = 0;
+    RequestId request_id;
     {
       std::unique_lock<std::mutex> lock(mu_);
 

src/io/message_conversion.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -14,6 +14,7 @@
 #include <boost/core/demangle.hpp>
 
 #include "io/transport.hpp"
+#include "utils/concrete_msg_sender.hpp"
 #include "utils/type_info_ref.hpp"
 
 namespace memgraph::io {
@@ -48,7 +49,7 @@ struct OpaqueMessage {
   /// Return is the full std::variant<Ts...> type that holds the
   /// full parameter pack without interfering with recursive
   /// narrowing expansion.
-  template <typename Return, Message Head, Message... Rest>
+  template <typename Return, utils::Message Head, utils::Message... Rest>
   std::optional<Return> Unpack(std::any &&a) {
     if (typeid(Head) == a.type()) {
       Head concrete = std::any_cast<Head>(std::move(a));
@@ -67,12 +68,12 @@ struct OpaqueMessage {
   /// parameter pack for the types that they want to compare
   /// with the any and potentially include in the returned
   /// variant.
-  template <Message... Ms>
+  template <utils::Message... Ms>
   requires(sizeof...(Ms) > 0) std::optional<std::variant<Ms...>> VariantFromAny(std::any &&a) {
     return Unpack<std::variant<Ms...>, Ms...>(std::move(a));
   }
 
-  template <Message... Ms>
+  template <utils::Message... Ms>
   requires(sizeof...(Ms) > 0) std::optional<RequestEnvelope<Ms...>> Take() && {
     std::optional<std::variant<Ms...>> m_opt = VariantFromAny<Ms...>(std::move(message));
 

src/io/messages.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -13,8 +13,8 @@
 
 #include <variant>
 
-#include <coordinator/coordinator.hpp>
-#include <io/rsm/raft.hpp>
+#include "coordinator/coordinator.hpp"
+#include "io/rsm/raft.hpp"
 #include "query/v2/requests.hpp"
 #include "utils/concepts.hpp"
 
@@ -41,6 +41,6 @@ using CoordinatorMessages =
 using ShardMessages = std::variant<ReadRequest<StorageReadRequest>, AppendRequest<StorageWriteRequest>, AppendResponse,
                                    WriteRequest<StorageWriteRequest>, VoteRequest, VoteResponse>;
 
-using ShardManagerMessages = std::variant<WriteResponse<CoordinatorWriteResponses>>;
+using ShardManagerMessages = std::variant<msgs::SuggestedSplitInfo, msgs::InitializeSplitShard>;
 
 }  // namespace memgraph::io::messages

src/io/rsm/raft.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -19,17 +19,23 @@
 #include <map>
 #include <set>
 #include <thread>
+#include <type_traits>
 #include <vector>
 
 #include <boost/core/demangle.hpp>
 
 #include "io/message_conversion.hpp"
+#include "io/rsm/shard_rsm.hpp"
 #include "io/simulator/simulator.hpp"
 #include "io/transport.hpp"
+#include "query/v2/requests.hpp"
 #include "utils/concepts.hpp"
 
 namespace memgraph::io::rsm {
 
+template <typename Type>
+concept HasShouldSplit = std::is_member_function_pointer<decltype(&Type::ShouldSplit)>::value;
+
 /// Timeout and replication tunables
 using namespace std::chrono_literals;
 static constexpr auto kMinimumElectionTimeout = 100ms;
@@ -342,15 +348,21 @@ class Raft {
     }
   }
 
- private:
-  // Raft paper - 5.3
-  // When the entry has been safely replicated, the leader applies the
-  // entry to its state machine and returns the result of that
-  // execution to the client.
-  //
-  // "Safely replicated" is defined as being known to be present
-  // on at least a majority of all peers (inclusive of the Leader).
-  void BumpCommitIndexAndReplyToClients(Leader &leader) {
+  template <typename = void>
+  requires HasShouldSplit<ReplicatedState> std::optional<msgs::SuggestedSplitInfo> ShouldSplit() {
+    return replicated_state_.ShouldSplit();
+  }
+
+ private :
+     // Raft paper - 5.3
+     // When the entry has been safely replicated, the leader applies the
+     // entry to its state machine and returns the result of that
+     // execution to the client.
+     //
+     // "Safely replicated" is defined as being known to be present
+     // on at least a majority of all peers (inclusive of the Leader).
+     void
+     BumpCommitIndexAndReplyToClients(Leader &leader) {
     auto confirmed_log_sizes = std::vector<LogSize>{};
 
     // We include our own log size in the calculation of the log
@@ -402,13 +414,17 @@ class Raft {
         const PendingClientRequest client_request = std::move(leader.pending_client_requests.at(apply_index));
         leader.pending_client_requests.erase(apply_index);
 
-        const WriteResponse<WriteResponseValue> resp{
-            .success = true,
-            .write_return = std::move(write_return),
-            .raft_index = apply_index,
-        };
+        if (client_request.request_id == 0) {
+          Log("not replying to Raft request with explicit request ID of 0");
+        } else {
+          const WriteResponse<WriteResponseValue> resp{
+              .success = true,
+              .write_return = std::move(write_return),
+              .raft_index = apply_index,
+          };
 
-        io_.Send(client_request.address, client_request.request_id, std::move(resp));
+          io_.Send(client_request.address, client_request.request_id, std::move(resp));
+        }
       }
     }
 
@@ -624,7 +640,7 @@ class Raft {
       MG_ASSERT(std::max(req.term, state_.term) == req.term);
     }
 
-    const VoteResponse res{
+    VoteResponse res{
         .term = std::max(req.term, state_.term),
         .committed_log_size = state_.committed_log_size,
         .vote_granted = new_leader,

src/io/rsm/rsm_client.hpp

@@ -65,7 +65,7 @@ class RsmClient {
     size_t addr_index = io_.Rand(addr_distrib);
     leader_ = server_addrs_[addr_index];
 
-    spdlog::debug("selecting a random leader at index {} with address {}", addr_index, leader_.ToString());
+    spdlog::trace("RsmClient selecting a random leader at index {} with address {}", addr_index, leader_.ToString());
   }
 
   template <typename ResponseT>
@@ -73,7 +73,7 @@ class RsmClient {
     if (response.retry_leader) {
       MG_ASSERT(!response.success, "retry_leader should never be set for successful responses");
       leader_ = response.retry_leader.value();
-      spdlog::debug("client redirected to leader server {}", leader_.ToString());
+      spdlog::trace("RsmClient redirected to leader server {}", leader_.ToString());
     }
     if (!response.success) {
       SelectRandomLeader();
@@ -139,6 +139,10 @@ class RsmClient {
   }
 
   std::optional<BasicResult<TimedOut, ReadResponseT>> PollAsyncReadRequest(const ReadinessToken &readiness_token) {
+    if (!async_reads_.contains(readiness_token.GetId())) {
+      spdlog::debug("RsmClient async_reads_ does not contain Polled readiness token {}", readiness_token.GetId());
+      return std::nullopt;
+    }
     auto &async_request = async_reads_.at(readiness_token.GetId());
 
     if (!async_request.future.IsReady()) {
@@ -158,7 +162,7 @@ class RsmClient {
 
     if (result_has_error && past_time_out) {
       // TODO static assert the exact type of error.
-      spdlog::debug("client timed out while trying to communicate with leader server {}", leader_.ToString());
+      spdlog::debug("RsmClient timed out while trying to communicate with leader server {}", leader_.ToString());
       async_reads_.erase(readiness_token.GetId());
       return TimedOut{};
     }
@@ -171,7 +175,8 @@ class RsmClient {
 
       if (read_get_response.success) {
         async_reads_.erase(readiness_token.GetId());
-        spdlog::debug("returning read_return for RSM request");
+        spdlog::trace("RsmClient returning read_return and erasing async state for RSM request for token {}",
+                      readiness_token.GetId());
         return std::move(read_get_response.read_return);
       }
     } else {
@@ -209,6 +214,10 @@ class RsmClient {
   }
 
   std::optional<BasicResult<TimedOut, WriteResponseT>> PollAsyncWriteRequest(const ReadinessToken &readiness_token) {
+    if (!async_writes_.contains(readiness_token.GetId())) {
+      spdlog::debug("RsmClient async_writes_ does not contain Polled readiness token {}", readiness_token.GetId());
+      return std::nullopt;
+    }
     auto &async_request = async_writes_.at(readiness_token.GetId());
 
     if (!async_request.future.IsReady()) {
@@ -228,7 +237,7 @@ class RsmClient {
 
     if (result_has_error && past_time_out) {
       // TODO static assert the exact type of error.
-      spdlog::debug("client timed out while trying to communicate with leader server {}", leader_.ToString());
+      spdlog::debug("RsmClient timed out while trying to communicate with leader server {}", leader_.ToString());
       async_writes_.erase(readiness_token.GetId());
       return TimedOut{};
     }

src/io/simulator/simulator_handle.cpp

@@ -22,7 +22,30 @@ namespace memgraph::io::simulator {
 
 void SimulatorHandle::ShutDown() {
   std::unique_lock<std::mutex> lock(mu_);
+
+  if (should_shut_down_) {
+    spdlog::warn("Simulator's ShutDown method called multiple times.");
+    return;
+  }
+
+  while (true) {
+    const size_t blocked_servers = blocked_on_receive_.size();
+
+    const bool all_servers_blocked = blocked_servers == server_addresses_.size();
+
+    if (all_servers_blocked) {
+      spdlog::trace("quiescent state detected - {} out of {} servers now blocked on receive", blocked_servers,
+                    server_addresses_.size());
+      break;
+    }
+
+    spdlog::trace("not returning from quiescent because we see {} blocked out of {}", blocked_servers,
+                  server_addresses_.size());
+    cv_.wait(lock);
+  }
+
   should_shut_down_ = true;
+
   for (auto it = promises_.begin(); it != promises_.end();) {
     auto &[promise_key, dop] = *it;
     std::move(dop).promise.TimeOut();

src/io/simulator/simulator_handle.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -32,6 +32,7 @@
 #include "io/simulator/simulator_stats.hpp"
 #include "io/time.hpp"
 #include "io/transport.hpp"
+#include "utils/concrete_msg_sender.hpp"
 
 namespace memgraph::io::simulator {
 
@@ -58,15 +59,16 @@ class SimulatorHandle {
   std::uniform_int_distribution<int> drop_distrib_{0, 99};
   SimulatorConfig config_;
   MessageHistogramCollector histograms_;
-  RequestId request_id_counter_{0};
+  RequestId request_id_counter_{1};
 
   bool TimeoutPromisesPastDeadline() {
     bool timed_anything_out = false;
     const Time now = cluster_wide_time_microseconds_;
+
     for (auto it = promises_.begin(); it != promises_.end();) {
       auto &[promise_key, dop] = *it;
       if (dop.deadline < now && config_.perform_timeouts) {
-        spdlog::trace("timing out request from requester {}.", promise_key.requester_address.ToString());
+        spdlog::trace("Simulator timing out request from requester {}.", promise_key.requester_address.ToString());
         std::move(dop).promise.TimeOut();
         it = promises_.erase(it);
 
@@ -76,6 +78,7 @@ class SimulatorHandle {
         ++it;
       }
     }
+
     return timed_anything_out;
   }
 
@@ -105,13 +108,13 @@ class SimulatorHandle {
 
   bool ShouldShutDown() const;
 
-  template <Message Request, Message Response>
+  template <utils::Message Request, utils::Message Response>
   ResponseFuture<Response> SubmitRequest(Address to_address, Address from_address, Request &&request, Duration timeout,
                                          std::function<bool()> &&maybe_tick_simulator,
                                          std::function<void()> &&fill_notifier) {
     auto type_info = TypeInfoFor(request);
     std::string demangled_name = boost::core::demangle(type_info.get().name());
-    spdlog::trace("simulator sending request {} to {}", demangled_name, to_address);
+    spdlog::trace("Simulator sending request {} to {}", demangled_name, to_address);
 
     auto [future, promise] = memgraph::io::FuturePromisePairWithNotifications<ResponseResult<Response>>(
         // set notifier for when the Future::Wait is called
@@ -155,7 +158,7 @@ class SimulatorHandle {
     return std::move(future);
   }
 
-  template <Message... Ms>
+  template <utils::Message... Ms>
   requires(sizeof...(Ms) > 0) RequestResult<Ms...> Receive(const Address &receiver, Duration timeout) {
     std::unique_lock<std::mutex> lock(mu_);
 
@@ -182,20 +185,21 @@ class SimulatorHandle {
       if (!should_shut_down_) {
         if (!blocked_on_receive_.contains(receiver)) {
           blocked_on_receive_.emplace(receiver);
-          spdlog::trace("blocking receiver {}", receiver.ToPartialAddress().port);
+          spdlog::trace("Simulator blocking receiver until it receives something or a timeout happens {}",
+                        receiver.ToPartialAddress().port);
           cv_.notify_all();
         }
         cv_.wait(lock);
       }
     }
-    spdlog::trace("timing out receiver {}", receiver.ToPartialAddress().port);
+    spdlog::trace("Simulator timing out receiver {}", receiver.ToPartialAddress().port);
 
     return TimedOut{};
   }
 
-  template <Message M>
+  template <utils::Message M>
   void Send(Address to_address, Address from_address, RequestId request_id, M message) {
-    spdlog::trace("sending message from {} to {}", from_address.last_known_port, to_address.last_known_port);
+    spdlog::trace("Simulator sending message from {} to {}", from_address.last_known_port, to_address.last_known_port);
     auto type_info = TypeInfoFor(message);
     {
       std::unique_lock<std::mutex> lock(mu_);

src/io/simulator/simulator_transport.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -18,6 +18,7 @@
 #include "io/notifier.hpp"
 #include "io/simulator/simulator_handle.hpp"
 #include "io/time.hpp"
+#include "utils/concrete_msg_sender.hpp"
 
 namespace memgraph::io::simulator {
 
@@ -33,7 +34,7 @@ class SimulatorTransport {
   SimulatorTransport(std::shared_ptr<SimulatorHandle> simulator_handle, Address address, uint64_t seed)
       : simulator_handle_(simulator_handle), address_(address), rng_(std::mt19937{seed}) {}
 
-  template <Message RequestT, Message ResponseT>
+  template <utils::Message RequestT, utils::Message ResponseT>
   ResponseFuture<ResponseT> Request(Address to_address, Address from_address, RequestT request,
                                     std::function<void()> notification, Duration timeout) {
     std::function<bool()> tick_simulator = [handle_copy = simulator_handle_] {
@@ -44,12 +45,12 @@ class SimulatorTransport {
         to_address, from_address, std::move(request), timeout, std::move(tick_simulator), std::move(notification));
   }
 
-  template <Message... Ms>
+  template <utils::Message... Ms>
   requires(sizeof...(Ms) > 0) RequestResult<Ms...> Receive(Address receiver_address, Duration timeout) {
     return simulator_handle_->template Receive<Ms...>(receiver_address, timeout);
   }
 
-  template <Message M>
+  template <utils::Message M>
   void Send(Address to_address, Address from_address, uint64_t request_id, M message) {
     return simulator_handle_->template Send<M>(to_address, from_address, request_id, message);
   }

src/io/transport.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -22,21 +22,16 @@
 #include "io/message_histogram_collector.hpp"
 #include "io/notifier.hpp"
 #include "io/time.hpp"
+#include "utils/concrete_msg_sender.hpp"
 #include "utils/result.hpp"
 
 namespace memgraph::io {
 
 using memgraph::utils::BasicResult;
 
-// TODO(tyler) ensure that Message continues to represent
-// reasonable constraints around message types over time,
-// as we adapt things to use Thrift-generated message types.
-template <typename T>
-concept Message = std::same_as<T, std::decay_t<T>>;
-
 using RequestId = uint64_t;
 
-template <Message M>
+template <utils::Message M>
 struct ResponseEnvelope {
   M message;
   RequestId request_id;
@@ -45,16 +40,16 @@ struct ResponseEnvelope {
   Duration response_latency;
 };
 
-template <Message M>
+template <utils::Message M>
 using ResponseResult = BasicResult<TimedOut, ResponseEnvelope<M>>;
 
-template <Message M>
-using ResponseFuture = memgraph::io::Future<ResponseResult<M>>;
+template <utils::Message M>
+using ResponseFuture = Future<ResponseResult<M>>;
 
-template <Message M>
-using ResponsePromise = memgraph::io::Promise<ResponseResult<M>>;
+template <utils::Message M>
+using ResponsePromise = Promise<ResponseResult<M>>;
 
-template <Message... Ms>
+template <utils::Message... Ms>
 struct RequestEnvelope {
   std::variant<Ms...> message;
   RequestId request_id;
@@ -62,7 +57,7 @@ struct RequestEnvelope {
   Address from_address;
 };
 
-template <Message... Ms>
+template <utils::Message... Ms>
 using RequestResult = BasicResult<TimedOut, RequestEnvelope<Ms...>>;
 
 template <typename I>
@@ -82,7 +77,7 @@ class Io {
   Duration GetDefaultTimeout() { return default_timeout_; }
 
   /// Issue a request with an explicit timeout in microseconds provided. This tends to be used by clients.
-  template <Message RequestT, Message ResponseT>
+  template <utils::Message RequestT, utils::Message ResponseT>
   ResponseFuture<ResponseT> RequestWithTimeout(Address address, RequestT request, Duration timeout) {
     const Address from_address = address_;
     std::function<void()> fill_notifier = nullptr;
@@ -92,7 +87,7 @@ class Io {
 
   /// Issue a request that times out after the default timeout. This tends
   /// to be used by clients.
-  template <Message RequestT, Message ResponseT>
+  template <utils::Message RequestT, utils::Message ResponseT>
   ResponseFuture<ResponseT> Request(Address to_address, RequestT request) {
     const Duration timeout = default_timeout_;
     const Address from_address = address_;
@@ -102,7 +97,7 @@ class Io {
   }
 
   /// Issue a request that will notify a Notifier when it is filled or times out.
-  template <Message RequestT, Message ResponseT>
+  template <utils::Message RequestT, utils::Message ResponseT>
   ResponseFuture<ResponseT> RequestWithNotification(Address to_address, RequestT request, Notifier notifier,
                                                     ReadinessToken readiness_token) {
     const Duration timeout = default_timeout_;
@@ -113,7 +108,7 @@ class Io {
   }
 
   /// Issue a request that will notify a Notifier when it is filled or times out.
-  template <Message RequestT, Message ResponseT>
+  template <utils::Message RequestT, utils::Message ResponseT>
   ResponseFuture<ResponseT> RequestWithNotificationAndTimeout(Address to_address, RequestT request, Notifier notifier,
                                                               ReadinessToken readiness_token, Duration timeout) {
     const Address from_address = address_;
@@ -124,14 +119,14 @@ class Io {
 
   /// Wait for an explicit number of microseconds for a request of one of the
   /// provided types to arrive. This tends to be used by servers.
-  template <Message... Ms>
+  template <utils::Message... Ms>
   RequestResult<Ms...> ReceiveWithTimeout(Duration timeout) {
     return implementation_.template Receive<Ms...>(address_, timeout);
   }
 
   /// Wait the default number of microseconds for a request of one of the
   /// provided types to arrive. This tends to be used by servers.
-  template <Message... Ms>
+  template <utils::Message... Ms>
   requires(sizeof...(Ms) > 0) RequestResult<Ms...> Receive() {
     const Duration timeout = default_timeout_;
     return implementation_.template Receive<Ms...>(address_, timeout);
@@ -140,7 +135,7 @@ class Io {
   /// Send a message in a best-effort fashion. This is used for messaging where
   /// responses are not necessarily expected, and for servers to respond to requests.
   /// If you need reliable delivery, this must be built on-top. TCP is not enough for most use cases.
-  template <Message M>
+  template <utils::Message M>
   void Send(Address to_address, RequestId request_id, M message) {
     Address from_address = address_;
     return implementation_.template Send<M>(to_address, from_address, request_id, std::move(message));
@@ -161,6 +156,7 @@ class Io {
   }
 
   Address GetAddress() { return address_; }
+
   void SetAddress(Address address) { address_ = address; }
 
   Io<I> ForkLocal(boost::uuids::uuid uuid) {
@@ -173,5 +169,16 @@ class Io {
   }
 
   LatencyHistogramSummaries ResponseLatencies() { return implementation_.ResponseLatencies(); }
+
+  template <utils::Message M>
+  utils::Sender<M> GetSender(Address address) {
+    Io<I> io_copy = Io(implementation_, address_);
+
+    std::function<void(M)> sender = [address, io_copy](M &&message) mutable {
+      io_copy.template Send<M>(address, 0, std::forward<M>(message));
+    };
+
+    return utils::Sender{sender};
+  }
 };
 };  // namespace memgraph::io

src/machine_manager/machine_manager.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -80,7 +80,7 @@ class MachineManager {
         config_(config),
         coordinator_address_(io.GetAddress().ForkLocalCoordinator()),
         shard_manager_{io.ForkLocal(io.GetAddress().ForkLocalShardManager().unique_id), config.shard_worker_threads,
-                       coordinator_address_} {
+                       coordinator_address_, config_.sync_message_handling} {
     auto coordinator_io = io.ForkLocal(coordinator_address_.unique_id);
     CoordinatorWorker coordinator_worker{coordinator_io, coordinator_queue_, coordinator};
     coordinator_handle_ = std::jthread([coordinator = std::move(coordinator_worker)]() mutable { coordinator.Run(); });
@@ -93,6 +93,7 @@ class MachineManager {
 
   ~MachineManager() {
     if (coordinator_handle_.joinable()) {
+      MaybeBlockOnSyncHandling();
       coordinator_queue_.Push(coordinator::coordinator_worker::ShutDown{});
       coordinator_handle_.join();
     }
@@ -114,10 +115,7 @@ class MachineManager {
       uint64_t next_us = next_cron_.time_since_epoch().count();
 
       if (now >= next_cron_) {
-        spdlog::info("now {} >= next_cron_ {}", now_us, next_us);
         next_cron_ = Cron();
-      } else {
-        spdlog::info("now {} < next_cron_ {}", now_us, next_us);
       }
 
       Duration receive_timeout = std::max(next_cron_, now) - now;
@@ -127,16 +125,17 @@ class MachineManager {
           std::variant<ReadRequest<CoordinatorReadRequests>, AppendRequest<CoordinatorWriteRequests>, AppendResponse,
                        WriteRequest<CoordinatorWriteRequests>, VoteRequest, VoteResponse,
                        WriteResponse<CoordinatorWriteResponses>, ReadRequest<StorageReadRequest>,
-                       AppendRequest<StorageWriteRequest>, WriteRequest<StorageWriteRequest>>;
+                       AppendRequest<StorageWriteRequest>, WriteRequest<StorageWriteRequest>, msgs::SuggestedSplitInfo,
+                       msgs::InitializeSplitShard>;
 
-      spdlog::info("MM waiting on Receive on address {}", io_.GetAddress().ToString());
+      spdlog::trace("MachineManager waiting on Receive on address {}", io_.GetAddress().ToString());
 
       // Note: this parameter pack must be kept in-sync with the AllMessages parameter pack above
       auto request_result = io_.template ReceiveWithTimeout<
           ReadRequest<CoordinatorReadRequests>, AppendRequest<CoordinatorWriteRequests>, AppendResponse,
           WriteRequest<CoordinatorWriteRequests>, VoteRequest, VoteResponse, WriteResponse<CoordinatorWriteResponses>,
-          ReadRequest<StorageReadRequest>, AppendRequest<StorageWriteRequest>, WriteRequest<StorageWriteRequest>>(
-          receive_timeout);
+          ReadRequest<StorageReadRequest>, AppendRequest<StorageWriteRequest>, WriteRequest<StorageWriteRequest>,
+          msgs::SuggestedSplitInfo, msgs::InitializeSplitShard>(receive_timeout);
 
       if (request_result.HasError()) {
         // time to do Cron
@@ -145,11 +144,13 @@ class MachineManager {
 
       auto &&request_envelope = std::move(request_result.GetValue());
 
-      spdlog::info("MM got message to {}", request_envelope.to_address.ToString());
+      spdlog::trace("MachineManager received message addressed to {}", request_envelope.to_address.ToString());
 
       // If message is for the coordinator, cast it to subset and pass it to the coordinator
       bool to_coordinator = coordinator_address_ == request_envelope.to_address;
       if (to_coordinator) {
+        spdlog::trace("MachineManager got message addressed to the coordinator");
+
         std::optional<CoordinatorMessages> conversion_attempt =
             ConvertVariant<AllMessages, ReadRequest<CoordinatorReadRequests>, AppendRequest<CoordinatorWriteRequests>,
                            AppendResponse, WriteRequest<CoordinatorWriteRequests>, VoteRequest, VoteResponse>(
@@ -157,8 +158,6 @@ class MachineManager {
 
         MG_ASSERT(conversion_attempt.has_value(), "coordinator message conversion failed");
 
-        spdlog::info("got coordinator message");
-
         CoordinatorMessages &&cm = std::move(conversion_attempt.value());
 
         CoordinatorRouteMessage route_message{
@@ -172,14 +171,20 @@ class MachineManager {
       }
 
       bool to_sm = shard_manager_.GetAddress() == request_envelope.to_address;
-      spdlog::info("smm: {}", shard_manager_.GetAddress().ToString());
       if (to_sm) {
+        spdlog::trace("MachineManager got shard manager message, addressed to {}", request_envelope.to_address);
+
         std::optional<ShardManagerMessages> conversion_attempt =
-            ConvertVariant<AllMessages, WriteResponse<CoordinatorWriteResponses>>(std::move(request_envelope.message));
+            ConvertVariant<AllMessages, msgs::SuggestedSplitInfo, msgs::InitializeSplitShard>(
+                std::move(request_envelope.message));
 
-        MG_ASSERT(conversion_attempt.has_value(), "shard manager message conversion failed");
-
-        spdlog::info("got shard manager message");
+        if (!conversion_attempt.has_value()) {
+          spdlog::debug(
+              "MachineManager dropping message with an unexpected type addressed to the ShardManager, assuming it was "
+              "a HeartbeatResponse that "
+              "timed out");
+          continue;
+        }
 
         ShardManagerMessages &&smm = std::move(conversion_attempt.value());
         shard_manager_.Receive(std::forward<ShardManagerMessages>(smm), request_envelope.request_id,
@@ -197,7 +202,7 @@ class MachineManager {
       MG_ASSERT(conversion_attempt.has_value(), "shard rsm message conversion failed for {} - incorrect message type",
                 request_envelope.to_address.ToString());
 
-      spdlog::info("got shard rsm message");
+      spdlog::trace("MachineManager forwarding message to shard rsm");
 
       ShardMessages &&sm = std::move(conversion_attempt.value());
       shard_manager_.Route(std::forward<ShardMessages>(sm), request_envelope.request_id, request_envelope.to_address,
@@ -221,7 +226,7 @@ class MachineManager {
   }
 
   Time Cron() {
-    spdlog::info("running MachineManager::Cron, address {}", io_.GetAddress().ToString());
+    spdlog::trace("MachineManager running Cron, address {}", io_.GetAddress().ToString());
     coordinator_queue_.Push(coordinator::coordinator_worker::Cron{});
     MaybeBlockOnSyncHandling();
     Time ret = shard_manager_.Cron();

src/memgraph.cpp

@@ -622,13 +622,16 @@ int main(int argc, char **argv) {
   if (FLAGS_split_file.empty()) {
     const std::string property{"property"};
     const std::string label{"label"};
+    // TODO(tyler) make this more easily configurable in the short-term
+    const auto split_threshold = 4;
+
     auto prop_map = sm.AllocatePropertyIds(std::vector<std::string>{property});
     auto edge_type_map = sm.AllocateEdgeTypeIds(std::vector<std::string>{"TO"});
     std::vector<memgraph::storage::v3::SchemaProperty> schema{
         {prop_map.at(property), memgraph::common::SchemaType::INT}};
-    sm.InitializeNewLabel(label, schema, 1, sm.shard_map_version);
-    sm.SplitShard(sm.GetHlc(), *sm.GetLabelId(label),
-                  std::vector<memgraph::storage::v3::PropertyValue>{memgraph::storage::v3::PropertyValue{2}});
+
+    // TODO(tyler) remove pre-initialization completely along with split files
+    sm.InitializeNewLabel(label, schema, 1, split_threshold, sm.shard_map_version);
   } else {
     std::ifstream input{FLAGS_split_file, std::ios::in};
     MG_ASSERT(input.is_open(), "Cannot open split file to read: {}", FLAGS_split_file);

src/query/v2/interpreter.hpp

@@ -295,6 +295,7 @@ class Interpreter final {
   void Abort();
 
   const RequestRouterInterface *GetRequestRouter() const { return request_router_.get(); }
+
   void InstallSimulatorTicker(std::function<bool()> &&tick_simulator) {
     request_router_->InstallSimulatorTicker(tick_simulator);
   }

src/query/v2/request_router.hpp

@@ -99,6 +99,7 @@ class RequestRouterInterface {
 
   virtual ~RequestRouterInterface() = default;
 
+  virtual coordinator::Hlc RefreshShardMap() = 0;
   virtual void StartTransaction() = 0;
   virtual void Commit() = 0;
   virtual std::vector<VertexAccessor> ScanVertices(std::optional<std::string> label) = 0;
@@ -150,77 +151,92 @@ class RequestRouter : public RequestRouterInterface {
     notifier_.InstallSimulatorTicker(tick_simulator);
   }
 
-  void StartTransaction() override {
-    coordinator::HlcRequest req{.last_shard_map_version = shards_map_.GetHlc()};
+  coordinator::Hlc RefreshShardMap() override {
+    coordinator::HlcRequest req{.last_shard_map_version = shard_map_.GetHlc()};
     CoordinatorWriteRequests write_req = req;
-    spdlog::trace("sending hlc request to start transaction");
+    spdlog::trace("RequestRouter sending hlc request to get an HLC and refresh the ShardMap");
     auto write_res = coord_cli_.SendWriteRequest(write_req);
-    spdlog::trace("received hlc response to start transaction");
-    if (write_res.HasError()) {
-      throw std::runtime_error("HLC request failed");
+    spdlog::trace("RequestRouter received hlc response to start transaction");
+
+    // TODO(tyler) enforce max retries here
+    while (write_res.HasError()) {
+      spdlog::debug("RequestRouter retrying HlcRequest to coordinator after timeout");
+      write_res = coord_cli_.SendWriteRequest(write_req);
     }
+
     auto coordinator_write_response = write_res.GetValue();
     auto hlc_response = std::get<coordinator::HlcResponse>(coordinator_write_response);
 
-    // Transaction ID to be used later...
-    transaction_id_ = hlc_response.new_hlc;
-
     if (hlc_response.fresher_shard_map) {
-      shards_map_ = hlc_response.fresher_shard_map.value();
+      shard_map_ = hlc_response.fresher_shard_map.value();
       SetUpNameIdMappers();
     }
+
+    return hlc_response.new_hlc;
   }
 
-  void Commit() override {
-    coordinator::HlcRequest req{.last_shard_map_version = shards_map_.GetHlc()};
-    CoordinatorWriteRequests write_req = req;
-    spdlog::trace("sending hlc request before committing transaction");
-    auto write_res = coord_cli_.SendWriteRequest(write_req);
-    spdlog::trace("received hlc response before committing transaction");
-    if (write_res.HasError()) {
-      throw std::runtime_error("HLC request for commit failed");
-    }
-    auto coordinator_write_response = write_res.GetValue();
-    auto hlc_response = std::get<coordinator::HlcResponse>(coordinator_write_response);
-
-    if (hlc_response.fresher_shard_map) {
-      shards_map_ = hlc_response.fresher_shard_map.value();
-      SetUpNameIdMappers();
-    }
-    auto commit_timestamp = hlc_response.new_hlc;
+  void StartTransaction() override {
+    // Transaction ID to be used later...
+    transaction_id_ = RefreshShardMap();
+  }
 
+  bool CommitInner(coordinator::Hlc commit_timestamp) {
     msgs::CommitRequest commit_req{.transaction_id = transaction_id_, .commit_timestamp = commit_timestamp};
 
-    for (const auto &[label, space] : shards_map_.label_spaces) {
+    for (const auto &[label, space] : shard_map_.label_spaces) {
       for (const auto &[key, shard] : space.shards) {
         auto &storage_client = GetStorageClientForShard(shard);
         // TODO(kostasrim) Currently requests return the result directly. Adjust this when the API works MgFuture
         // instead.
         auto commit_response = storage_client.SendWriteRequest(commit_req);
-        // RETRY on timeouts?
-        // Sometimes this produces a timeout. Temporary solution is to use a while(true) as was done in shard_map test
-        if (commit_response.HasError()) {
-          throw std::runtime_error("Commit request timed out");
+        while (commit_response.HasError()) {
+          spdlog::debug("RequestRouter retrying Commit request due to timeout");
+          commit_response = storage_client.SendWriteRequest(commit_req);
         }
         msgs::WriteResponses write_response_variant = commit_response.GetValue();
         auto &response = std::get<msgs::CommitResponse>(write_response_variant);
         if (response.error) {
-          throw std::runtime_error("Commit request did not succeed");
+          if (response.error->code == common::ErrorCode::STALE_SHARD_MAP) {
+            RefreshShardMap();
+
+            // signal to caller that we should retry
+            return false;
+          } else {
+            spdlog::warn("RequestRouter throwing because of unhandled commit failure 3: {}",
+                         common::ErrorCodeToString(response.error->code));
+            throw std::runtime_error("Commit request did not succeed");
+          }
         }
       }
     }
+
+    return true;
+  }
+
+  void Commit() override {
+    spdlog::trace("sending hlc request before committing transaction");
+    auto commit_timestamp = RefreshShardMap();
+
+    // TODO(tyler) enforce max commit retries here when we go distributed
+    while (true) {
+      // Commit is idempotent, so it's fine to retry it to the same shards if it fails
+      bool success = CommitInner(commit_timestamp);
+      if (success) {
+        return;
+      }
+    }
   }
 
   storage::v3::EdgeTypeId NameToEdgeType(const std::string &name) const override {
-    return shards_map_.GetEdgeTypeId(name).value();
+    return shard_map_.GetEdgeTypeId(name).value();
   }
 
   storage::v3::PropertyId NameToProperty(const std::string &name) const override {
-    return shards_map_.GetPropertyId(name).value();
+    return shard_map_.GetPropertyId(name).value();
   }
 
   storage::v3::LabelId NameToLabel(const std::string &name) const override {
-    return shards_map_.GetLabelId(name).value();
+    return shard_map_.GetLabelId(name).value();
   }
 
   const std::string &PropertyToName(storage::v3::PropertyId id) const override {
@@ -232,8 +248,8 @@ class RequestRouter : public RequestRouterInterface {
   }
 
   bool IsPrimaryProperty(storage::v3::LabelId primary_label, storage::v3::PropertyId property) const override {
-    const auto schema_it = shards_map_.schemas.find(primary_label);
-    MG_ASSERT(schema_it != shards_map_.schemas.end(), "Invalid primary label id: {}", primary_label.AsUint());
+    const auto schema_it = shard_map_.schemas.find(primary_label);
+    MG_ASSERT(schema_it != shard_map_.schemas.end(), "Invalid primary label id: {}", primary_label.AsUint());
 
     return std::find_if(schema_it->second.begin(), schema_it->second.end(), [property](const auto &schema_prop) {
              return schema_prop.property_id == property;
@@ -241,94 +257,136 @@ class RequestRouter : public RequestRouterInterface {
   }
 
   const std::vector<coordinator::SchemaProperty> &GetSchemaForLabel(storage::v3::LabelId label) const override {
-    return shards_map_.schemas.at(label);
+    return shard_map_.schemas.at(label);
   }
 
-  bool IsPrimaryLabel(storage::v3::LabelId label) const override { return shards_map_.label_spaces.contains(label); }
+  bool IsPrimaryLabel(storage::v3::LabelId label) const override { return shard_map_.label_spaces.contains(label); }
 
   // TODO(kostasrim) Simplify return result
   std::vector<VertexAccessor> ScanVertices(std::optional<std::string> label) override {
-    // create requests
-    auto requests_to_be_sent = RequestsForScanVertices(label);
+    // TODO(tyler) enforce max commit retries here when we go distributed
+    while (true) {
+      // create requests
+      auto requests_to_be_sent = RequestsForScanVertices(label);
 
-    spdlog::trace("created {} ScanVertices requests", requests_to_be_sent.size());
+      spdlog::trace("created {} ScanVertices requests", requests_to_be_sent.size());
 
-    // begin all requests in parallel
-    RunningRequests<msgs::ScanVerticesRequest> running_requests = {};
-    running_requests.reserve(requests_to_be_sent.size());
-    for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
-      auto &request = requests_to_be_sent[i];
-      io::ReadinessToken readiness_token{i};
-      auto &storage_client = GetStorageClientForShard(request.shard);
-      storage_client.SendAsyncReadRequest(request.request, notifier_, readiness_token);
-      running_requests.emplace(readiness_token.GetId(), request);
-    }
-    spdlog::trace("sent {} ScanVertices requests in parallel", running_requests.size());
-
-    // drive requests to completion
-    auto responses = DriveReadResponses<msgs::ScanVerticesRequest, msgs::ScanVerticesResponse>(running_requests);
-    spdlog::trace("got back {} ScanVertices responses after driving to completion", responses.size());
-
-    // convert responses into VertexAccessor objects to return
-    std::vector<VertexAccessor> accessors;
-    accessors.reserve(responses.size());
-    for (auto &response : responses) {
-      for (auto &result_row : response.results) {
-        accessors.emplace_back(VertexAccessor(std::move(result_row.vertex), std::move(result_row.props), this));
+      // begin all requests in parallel
+      RunningRequests<msgs::ScanVerticesRequest> running_requests = {};
+      running_requests.reserve(requests_to_be_sent.size());
+      for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
+        auto &request = requests_to_be_sent[i];
+        io::ReadinessToken readiness_token{i};
+        auto &storage_client = GetStorageClientForShard(request.shard);
+        storage_client.SendAsyncReadRequest(request.request, notifier_, readiness_token);
+        running_requests.emplace(readiness_token.GetId(), request);
       }
-    }
+      spdlog::trace("sent {} ScanVertices requests in parallel", running_requests.size());
 
-    return accessors;
+      // drive requests to completion
+      auto responses_result =
+          DriveReadResponses<msgs::ScanVerticesRequest, msgs::ScanVerticesResponse>(running_requests);
+
+      if (responses_result.HasError()) {
+        spdlog::debug(
+            "RequestRouter refreshing ShardMap and re-sending requests for ScanVertices after Shard version mismatch");
+        RefreshShardMap();
+        continue;
+      }
+
+      auto responses = responses_result.GetValue();
+
+      spdlog::trace("got back {} ScanVertices responses after driving to completion", responses.size());
+
+      // convert responses into VertexAccessor objects to return
+      std::vector<VertexAccessor> accessors;
+      accessors.reserve(responses.size());
+      for (auto &response : responses) {
+        for (auto &result_row : response.results) {
+          accessors.emplace_back(VertexAccessor(std::move(result_row.vertex), std::move(result_row.props), this));
+        }
+      }
+
+      return accessors;
+    }
   }
 
   std::vector<msgs::CreateVerticesResponse> CreateVertices(std::vector<msgs::NewVertex> new_vertices) override {
     MG_ASSERT(!new_vertices.empty());
 
-    // create requests
-    std::vector<ShardRequestState<msgs::CreateVerticesRequest>> requests_to_be_sent =
-        RequestsForCreateVertices(new_vertices);
-    spdlog::trace("created {} CreateVertices requests", requests_to_be_sent.size());
-
-    // begin all requests in parallel
-    RunningRequests<msgs::CreateVerticesRequest> running_requests = {};
-    running_requests.reserve(requests_to_be_sent.size());
-    for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
-      auto &request = requests_to_be_sent[i];
-      io::ReadinessToken readiness_token{i};
-      for (auto &new_vertex : request.request.new_vertices) {
-        new_vertex.label_ids.erase(new_vertex.label_ids.begin());
-      }
-      auto &storage_client = GetStorageClientForShard(request.shard);
-      storage_client.SendAsyncWriteRequest(request.request, notifier_, readiness_token);
-      running_requests.emplace(readiness_token.GetId(), request);
+    for (auto &new_vertex : new_vertices) {
+      new_vertex.idempotency_token = idempotency_token_generator_++;
     }
-    spdlog::trace("sent {} CreateVertices requests in parallel", running_requests.size());
 
-    // drive requests to completion
-    return DriveWriteResponses<msgs::CreateVerticesRequest, msgs::CreateVerticesResponse>(running_requests);
+    // TODO(tyler) enforce max commit retries here when we go distributed
+    while (true) {
+      // create requests
+      std::vector<ShardRequestState<msgs::CreateVerticesRequest>> requests_to_be_sent =
+          RequestsForCreateVertices(new_vertices);
+      spdlog::trace("created {} CreateVertices requests", requests_to_be_sent.size());
+
+      // begin all requests in parallel
+      RunningRequests<msgs::CreateVerticesRequest> running_requests = {};
+      running_requests.reserve(requests_to_be_sent.size());
+      for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
+        auto &request = requests_to_be_sent[i];
+        io::ReadinessToken readiness_token{i};
+        for (auto &new_vertex : request.request.new_vertices) {
+          new_vertex.label_ids.erase(new_vertex.label_ids.begin());
+        }
+        auto &storage_client = GetStorageClientForShard(request.shard);
+        storage_client.SendAsyncWriteRequest(request.request, notifier_, readiness_token);
+        running_requests.emplace(readiness_token.GetId(), request);
+      }
+      spdlog::trace("sent {} CreateVertices requests in parallel", running_requests.size());
+
+      // drive requests to completion
+      auto result = DriveWriteResponses<msgs::CreateVerticesRequest, msgs::CreateVerticesResponse>(running_requests);
+
+      if (result.HasValue()) {
+        return result.GetValue();
+      } else {
+        // retry request that failed due to outdated ShardMap
+        RefreshShardMap();
+      }
+    }
   }
 
   std::vector<msgs::CreateExpandResponse> CreateExpand(std::vector<msgs::NewExpand> new_edges) override {
     MG_ASSERT(!new_edges.empty());
 
-    // create requests
-    std::vector<ShardRequestState<msgs::CreateExpandRequest>> requests_to_be_sent =
-        RequestsForCreateExpand(std::move(new_edges));
-
-    // begin all requests in parallel
-    RunningRequests<msgs::CreateExpandRequest> running_requests = {};
-    running_requests.reserve(requests_to_be_sent.size());
-    for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
-      auto &request = requests_to_be_sent[i];
-      io::ReadinessToken readiness_token{i};
-      auto &storage_client = GetStorageClientForShard(request.shard);
-      msgs::WriteRequests req = request.request;
-      storage_client.SendAsyncWriteRequest(req, notifier_, readiness_token);
-      running_requests.emplace(readiness_token.GetId(), request);
+    for (auto &new_edge : new_edges) {
+      new_edge.idempotency_token = idempotency_token_generator_++;
     }
 
-    // drive requests to completion
-    return DriveWriteResponses<msgs::CreateExpandRequest, msgs::CreateExpandResponse>(running_requests);
+    // TODO(tyler) enforce max commit retries here when we go distributed
+    while (true) {
+      // create requests
+      std::vector<ShardRequestState<msgs::CreateExpandRequest>> requests_to_be_sent =
+          RequestsForCreateExpand(new_edges);
+
+      // begin all requests in parallel
+      RunningRequests<msgs::CreateExpandRequest> running_requests = {};
+      running_requests.reserve(requests_to_be_sent.size());
+      for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
+        auto &request = requests_to_be_sent[i];
+        io::ReadinessToken readiness_token{i};
+        auto &storage_client = GetStorageClientForShard(request.shard);
+        msgs::WriteRequests req = request.request;
+        storage_client.SendAsyncWriteRequest(req, notifier_, readiness_token);
+        running_requests.emplace(readiness_token.GetId(), request);
+      }
+
+      // drive requests to completion
+      auto result = DriveWriteResponses<msgs::CreateExpandRequest, msgs::CreateExpandResponse>(running_requests);
+
+      if (result.HasValue()) {
+        return result.GetValue();
+      } else {
+        // retry request that failed due to outdated ShardMap
+        RefreshShardMap();
+      }
+    }
   }
 
   std::vector<msgs::ExpandOneResultRow> ExpandOne(msgs::ExpandOneRequest request) override {
@@ -338,81 +396,107 @@ class RequestRouter : public RequestRouterInterface {
     // For each vertex U, the ExpandOne will result in <U, Edges>. The destination vertex and its properties
     // must be fetched again with an ExpandOne(Edges.dst)
 
-    // create requests
-    std::vector<ShardRequestState<msgs::ExpandOneRequest>> requests_to_be_sent = RequestsForExpandOne(request);
+    // TODO(tyler) enforce max commit retries here when we go distributed
+    while (true) {
+      // create requests
+      std::vector<ShardRequestState<msgs::ExpandOneRequest>> requests_to_be_sent = RequestsForExpandOne(request);
 
-    // begin all requests in parallel
-    RunningRequests<msgs::ExpandOneRequest> running_requests = {};
-    running_requests.reserve(requests_to_be_sent.size());
-    for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
-      auto &request = requests_to_be_sent[i];
-      io::ReadinessToken readiness_token{i};
-      auto &storage_client = GetStorageClientForShard(request.shard);
-      msgs::ReadRequests req = request.request;
-      storage_client.SendAsyncReadRequest(req, notifier_, readiness_token);
-      running_requests.emplace(readiness_token.GetId(), request);
+      // begin all requests in parallel
+      RunningRequests<msgs::ExpandOneRequest> running_requests = {};
+      running_requests.reserve(requests_to_be_sent.size());
+      for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
+        auto &request = requests_to_be_sent[i];
+        io::ReadinessToken readiness_token{i};
+        auto &storage_client = GetStorageClientForShard(request.shard);
+        msgs::ReadRequests req = request.request;
+        storage_client.SendAsyncReadRequest(req, notifier_, readiness_token);
+        running_requests.emplace(readiness_token.GetId(), request);
+      }
+
+      // drive requests to completion
+      auto responses_result = DriveReadResponses<msgs::ExpandOneRequest, msgs::ExpandOneResponse>(running_requests);
+
+      if (responses_result.HasError()) {
+        spdlog::debug(
+            "RequestRouter refreshing ShardMap and re-sending requests for ExpandOne after Shard version mismatch");
+        RefreshShardMap();
+        continue;
+      }
+
+      auto responses = responses_result.GetValue();
+
+      // post-process responses
+      std::vector<msgs::ExpandOneResultRow> result_rows;
+      const auto total_row_count = std::accumulate(
+          responses.begin(), responses.end(), 0, [](const int64_t partial_count, const msgs::ExpandOneResponse &resp) {
+            return partial_count + resp.result.size();
+          });
+      result_rows.reserve(total_row_count);
+
+      for (auto &response : responses) {
+        result_rows.insert(result_rows.end(), std::make_move_iterator(response.result.begin()),
+                           std::make_move_iterator(response.result.end()));
+      }
+
+      return result_rows;
     }
-
-    // drive requests to completion
-    auto responses = DriveReadResponses<msgs::ExpandOneRequest, msgs::ExpandOneResponse>(running_requests);
-
-    // post-process responses
-    std::vector<msgs::ExpandOneResultRow> result_rows;
-    const auto total_row_count = std::accumulate(responses.begin(), responses.end(), 0,
-                                                 [](const int64_t partial_count, const msgs::ExpandOneResponse &resp) {
-                                                   return partial_count + resp.result.size();
-                                                 });
-    result_rows.reserve(total_row_count);
-
-    for (auto &response : responses) {
-      result_rows.insert(result_rows.end(), std::make_move_iterator(response.result.begin()),
-                         std::make_move_iterator(response.result.end()));
-    }
-
-    return result_rows;
   }
 
   std::vector<msgs::GetPropertiesResultRow> GetProperties(msgs::GetPropertiesRequest requests) override {
     requests.transaction_id = transaction_id_;
-    // create requests
-    std::vector<ShardRequestState<msgs::GetPropertiesRequest>> requests_to_be_sent =
-        RequestsForGetProperties(std::move(requests));
 
-    // begin all requests in parallel
-    RunningRequests<msgs::GetPropertiesRequest> running_requests = {};
-    running_requests.reserve(requests_to_be_sent.size());
-    for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
-      auto &request = requests_to_be_sent[i];
-      io::ReadinessToken readiness_token{i};
-      auto &storage_client = GetStorageClientForShard(request.shard);
-      msgs::ReadRequests req = request.request;
-      storage_client.SendAsyncReadRequest(req, notifier_, readiness_token);
-      running_requests.emplace(readiness_token.GetId(), request);
+    // TODO(tyler) enforce max commit retries here when we go distributed
+    while (true) {
+      // create requests
+      std::vector<ShardRequestState<msgs::GetPropertiesRequest>> requests_to_be_sent =
+          RequestsForGetProperties(requests);
+
+      // begin all requests in parallel
+      RunningRequests<msgs::GetPropertiesRequest> running_requests = {};
+      running_requests.reserve(requests_to_be_sent.size());
+      for (size_t i = 0; i < requests_to_be_sent.size(); i++) {
+        auto &request = requests_to_be_sent[i];
+        io::ReadinessToken readiness_token{i};
+        auto &storage_client = GetStorageClientForShard(request.shard);
+        msgs::ReadRequests req = request.request;
+        storage_client.SendAsyncReadRequest(req, notifier_, readiness_token);
+        running_requests.emplace(readiness_token.GetId(), request);
+      }
+
+      // drive requests to completion
+      auto responses_result =
+          DriveReadResponses<msgs::GetPropertiesRequest, msgs::GetPropertiesResponse>(running_requests);
+
+      if (responses_result.HasError()) {
+        spdlog::debug(
+            "RequestRouter refreshing ShardMap and re-sending requests for GetProperties after Shard version mismatch");
+        RefreshShardMap();
+        continue;
+      }
+
+      auto responses = responses_result.GetValue();
+
+      // post-process responses
+      std::vector<msgs::GetPropertiesResultRow> result_rows;
+
+      for (auto &&response : responses) {
+        std::move(response.result_row.begin(), response.result_row.end(), std::back_inserter(result_rows));
+      }
+
+      return result_rows;
     }
-
-    // drive requests to completion
-    auto responses = DriveReadResponses<msgs::GetPropertiesRequest, msgs::GetPropertiesResponse>(running_requests);
-
-    // post-process responses
-    std::vector<msgs::GetPropertiesResultRow> result_rows;
-
-    for (auto &&response : responses) {
-      std::move(response.result_row.begin(), response.result_row.end(), std::back_inserter(result_rows));
-    }
-
-    return result_rows;
   }
 
   std::optional<storage::v3::PropertyId> MaybeNameToProperty(const std::string &name) const override {
-    return shards_map_.GetPropertyId(name);
+    return shard_map_.GetPropertyId(name);
   }
 
   std::optional<storage::v3::EdgeTypeId> MaybeNameToEdgeType(const std::string &name) const override {
-    return shards_map_.GetEdgeTypeId(name);
+    return shard_map_.GetEdgeTypeId(name);
   }
 
   std::optional<storage::v3::LabelId> MaybeNameToLabel(const std::string &name) const override {
-    return shards_map_.GetLabelId(name);
+    return shard_map_.GetLabelId(name);
   }
 
  private:
@@ -420,12 +504,13 @@ class RequestRouter : public RequestRouterInterface {
       const std::vector<msgs::NewVertex> &new_vertices) {
     std::map<ShardMetadata, msgs::CreateVerticesRequest> per_shard_request_table;
 
-    for (auto &new_vertex : new_vertices) {
+    for (const auto &new_vertex : new_vertices) {
       MG_ASSERT(!new_vertex.label_ids.empty(), "No label_ids provided for new vertex in RequestRouter::CreateVertices");
-      auto shard = shards_map_.GetShardForKey(new_vertex.label_ids[0].id,
-                                              storage::conversions::ConvertPropertyVector(new_vertex.primary_key));
+      auto shard = shard_map_.GetShardForKey(new_vertex.label_ids[0].id,
+                                             storage::conversions::ConvertPropertyVector(new_vertex.primary_key));
       if (!per_shard_request_table.contains(shard)) {
-        msgs::CreateVerticesRequest create_v_rqst{.transaction_id = transaction_id_};
+        msgs::CreateVerticesRequest create_v_rqst{.transaction_id = transaction_id_,
+                                                  .shard_map_version = shard_map_.shard_map_version};
         per_shard_request_table.insert(std::pair(shard, std::move(create_v_rqst)));
       }
       per_shard_request_table[shard].new_vertices.push_back(std::move(new_vertex));
@@ -445,7 +530,7 @@ class RequestRouter : public RequestRouterInterface {
   }
 
   std::vector<ShardRequestState<msgs::CreateExpandRequest>> RequestsForCreateExpand(
-      std::vector<msgs::NewExpand> new_expands) {
+      const std::vector<msgs::NewExpand> &new_expands) {
     std::map<ShardMetadata, msgs::CreateExpandRequest> per_shard_request_table;
     auto ensure_shard_exists_in_table = [&per_shard_request_table,
                                          transaction_id = transaction_id_](const ShardMetadata &shard) {
@@ -456,14 +541,14 @@ class RequestRouter : public RequestRouterInterface {
     };
 
     for (auto &new_expand : new_expands) {
-      const auto shard_src_vertex = shards_map_.GetShardForKey(
+      const auto shard_src_vertex = shard_map_.GetShardForKey(
           new_expand.src_vertex.first.id, storage::conversions::ConvertPropertyVector(new_expand.src_vertex.second));
-      const auto shard_dest_vertex = shards_map_.GetShardForKey(
+      const auto shard_dest_vertex = shard_map_.GetShardForKey(
           new_expand.dest_vertex.first.id, storage::conversions::ConvertPropertyVector(new_expand.dest_vertex.second));
 
       ensure_shard_exists_in_table(shard_src_vertex);
 
-      if (shard_src_vertex != shard_dest_vertex) {
+      if (shard_src_vertex.peers != shard_dest_vertex.peers) {
         ensure_shard_exists_in_table(shard_dest_vertex);
         per_shard_request_table[shard_dest_vertex].new_expands.push_back(new_expand);
       }
@@ -487,12 +572,12 @@ class RequestRouter : public RequestRouterInterface {
       const std::optional<std::string> &label) {
     std::vector<coordinator::Shards> multi_shards;
     if (label) {
-      const auto label_id = shards_map_.GetLabelId(*label);
+      const auto label_id = shard_map_.GetLabelId(*label);
       MG_ASSERT(label_id);
       MG_ASSERT(IsPrimaryLabel(*label_id));
-      multi_shards = {shards_map_.GetShardsForLabel(*label)};
+      multi_shards = {shard_map_.GetShardsForLabel(*label)};
     } else {
-      multi_shards = shards_map_.GetAllShards();
+      multi_shards = shard_map_.GetAllShards();
     }
 
     std::vector<ShardRequestState<msgs::ScanVerticesRequest>> requests = {};
@@ -503,6 +588,7 @@ class RequestRouter : public RequestRouterInterface {
 
         msgs::ScanVerticesRequest request;
         request.transaction_id = transaction_id_;
+        request.shard_map_version = shard_map_.shard_map_version;
         request.start_id.second = storage::conversions::ConvertValueVector(key);
 
         ShardRequestState<msgs::ScanVerticesRequest> shard_request_state{
@@ -525,7 +611,7 @@ class RequestRouter : public RequestRouterInterface {
 
     for (auto &vertex : request.src_vertices) {
       auto shard =
-          shards_map_.GetShardForKey(vertex.first.id, storage::conversions::ConvertPropertyVector(vertex.second));
+          shard_map_.GetShardForKey(vertex.first.id, storage::conversions::ConvertPropertyVector(vertex.second));
       if (!per_shard_request_table.contains(shard)) {
         per_shard_request_table.insert(std::pair(shard, top_level_rqst_template));
       }
@@ -547,7 +633,7 @@ class RequestRouter : public RequestRouterInterface {
   }
 
   std::vector<ShardRequestState<msgs::GetPropertiesRequest>> RequestsForGetProperties(
-      msgs::GetPropertiesRequest &&request) {
+      const msgs::GetPropertiesRequest &request) {
     std::map<ShardMetadata, msgs::GetPropertiesRequest> per_shard_request_table;
     auto top_level_rqst_template = request;
     top_level_rqst_template.transaction_id = transaction_id_;
@@ -556,7 +642,7 @@ class RequestRouter : public RequestRouterInterface {
 
     for (auto &&vertex : request.vertex_ids) {
       auto shard =
-          shards_map_.GetShardForKey(vertex.first.id, storage::conversions::ConvertPropertyVector(vertex.second));
+          shard_map_.GetShardForKey(vertex.first.id, storage::conversions::ConvertPropertyVector(vertex.second));
       if (!per_shard_request_table.contains(shard)) {
         per_shard_request_table.insert(std::pair(shard, top_level_rqst_template));
       }
@@ -565,7 +651,7 @@ class RequestRouter : public RequestRouterInterface {
 
     for (auto &[vertex, maybe_edge] : request.vertices_and_edges) {
       auto shard =
-          shards_map_.GetShardForKey(vertex.first.id, storage::conversions::ConvertPropertyVector(vertex.second));
+          shard_map_.GetShardForKey(vertex.first.id, storage::conversions::ConvertPropertyVector(vertex.second));
       if (!per_shard_request_table.contains(shard)) {
         per_shard_request_table.insert(std::pair(shard, top_level_rqst_template));
       }
@@ -594,7 +680,7 @@ class RequestRouter : public RequestRouterInterface {
   }
 
   StorageClient &GetStorageClientForShard(const std::string &label, const CompoundKey &key) {
-    auto shard = shards_map_.GetShardForKey(label, key);
+    auto shard = shard_map_.GetShardForKey(label, key);
     return GetStorageClientForShard(std::move(shard));
   }
 
@@ -611,16 +697,21 @@ class RequestRouter : public RequestRouterInterface {
   }
 
   template <typename RequestT, typename ResponseT>
-  std::vector<ResponseT> DriveReadResponses(RunningRequests<RequestT> &running_requests) {
+  utils::BasicResult<io::ShardVersionMismatch, std::vector<ResponseT>> DriveReadResponses(
+      RunningRequests<RequestT> &running_requests) {
     // Store responses in a map based on the corresponding request
     // offset, so that they can be reassembled in the correct order
     // even if they came back in randomized orders.
     std::map<size_t, ResponseT> response_map;
 
-    spdlog::trace("waiting on readiness for token");
+    spdlog::trace("waiting on readiness for token in DriveReadResponses");
+    size_t polls = 0;
     while (response_map.size() < running_requests.size()) {
       auto ready = notifier_.Await();
       spdlog::trace("got readiness for token {}", ready.GetId());
+
+      MG_ASSERT(polls++ / running_requests.size() < 1000,
+                "polled over 1000 times per operation (almost certainly due to a bug) when performing request");
       auto &request = running_requests.at(ready.GetId());
       auto &storage_client = GetStorageClientForShard(request.shard);
 
@@ -632,13 +723,23 @@ class RequestRouter : public RequestRouterInterface {
       }
 
       if (poll_result->HasError()) {
-        throw std::runtime_error("RequestRouter Read request timed out");
+        storage_client.SendAsyncReadRequest(request.request, notifier_, ready);
+        continue;
       }
 
       msgs::ReadResponses response_variant = poll_result->GetValue();
       auto response = std::get<ResponseT>(response_variant);
       if (response.error) {
-        throw std::runtime_error("RequestRouter Read request did not succeed");
+        if (response.error->code == common::ErrorCode::STALE_SHARD_MAP) {
+          RefreshShardMap();
+
+          // signal to caller that we should retry
+          return io::ShardVersionMismatch{};
+        } else {
+          spdlog::warn("throwing in DriveReadResponses because of unhandled error: {}",
+                       common::ErrorCodeToString(response.error->code));
+          throw std::runtime_error("RequestRouter Read request did not succeed");
+        }
       }
 
       // the readiness token has an ID based on the request vector offset
@@ -659,7 +760,8 @@ class RequestRouter : public RequestRouterInterface {
   }
 
   template <typename RequestT, typename ResponseT>
-  std::vector<ResponseT> DriveWriteResponses(RunningRequests<RequestT> &running_requests) {
+  utils::BasicResult<io::ShardVersionMismatch, std::vector<ResponseT>> DriveWriteResponses(
+      RunningRequests<RequestT> &running_requests) {
     // Store responses in a map based on the corresponding request
     // offset, so that they can be reassembled in the correct order
     // even if they came back in randomized orders.
@@ -678,13 +780,23 @@ class RequestRouter : public RequestRouterInterface {
       }
 
       if (poll_result->HasError()) {
-        throw std::runtime_error("RequestRouter Write request timed out");
+        storage_client.SendAsyncWriteRequest(request.request, notifier_, ready);
+        continue;
       }
 
       msgs::WriteResponses response_variant = poll_result->GetValue();
       auto response = std::get<ResponseT>(response_variant);
       if (response.error) {
-        throw std::runtime_error("RequestRouter Write request did not succeed");
+        if (response.error->code == common::ErrorCode::STALE_SHARD_MAP) {
+          RefreshShardMap();
+
+          // signal to caller that we should retry
+          return io::ShardVersionMismatch{};
+        } else {
+          spdlog::warn("throwing in DriveWriteResponses because of unhandled error: {}",
+                       common::ErrorCodeToString(response.error->code));
+          throw std::runtime_error("RequestRouter Write request did not succeed");
+        }
       }
 
       // the readiness token has an ID based on the request vector offset
@@ -706,17 +818,17 @@ class RequestRouter : public RequestRouterInterface {
 
   void SetUpNameIdMappers() {
     std::unordered_map<uint64_t, std::string> id_to_name;
-    for (const auto &[name, id] : shards_map_.labels) {
+    for (const auto &[name, id] : shard_map_.labels) {
       id_to_name.emplace(id.AsUint(), name);
     }
     labels_.StoreMapping(std::move(id_to_name));
     id_to_name.clear();
-    for (const auto &[name, id] : shards_map_.properties) {
+    for (const auto &[name, id] : shard_map_.properties) {
       id_to_name.emplace(id.AsUint(), name);
     }
     properties_.StoreMapping(std::move(id_to_name));
     id_to_name.clear();
-    for (const auto &[name, id] : shards_map_.edge_types) {
+    for (const auto &[name, id] : shard_map_.edge_types) {
       id_to_name.emplace(id.AsUint(), name);
     }
     edge_types_.StoreMapping(std::move(id_to_name));
@@ -739,7 +851,7 @@ class RequestRouter : public RequestRouterInterface {
     return {};
   }
 
-  ShardMap shards_map_;
+  ShardMap shard_map_;
   storage::v3::NameIdMapper properties_;
   storage::v3::NameIdMapper edge_types_;
   storage::v3::NameIdMapper labels_;
@@ -748,6 +860,7 @@ class RequestRouter : public RequestRouterInterface {
   io::Io<TTransport> io_;
   coordinator::Hlc transaction_id_;
   io::Notifier notifier_ = {};
+  std::atomic<uint64_t> idempotency_token_generator_;
   // TODO(kostasrim) Add batch prefetching
 };
 
@@ -800,15 +913,19 @@ class SimulatedRequestRouterFactory : public RequestRouterFactory {
     io::Address unique_local_addr_query;
 
     // The simulated RR should not introduce stochastic behavior.
-    random_uuid = boost::uuids::uuid{3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
-    unique_local_addr_query = {.unique_id = boost::uuids::uuid{4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}};
+    random_uuid = boost::uuids::uuid{3, 3, 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+    unique_local_addr_query = {.unique_id = boost::uuids::uuid{4, 4, 4, 4, 4, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}};
 
     auto io = simulator_->Register(unique_local_addr_query);
     auto query_io = io.ForkLocal(random_uuid);
 
-    return std::make_unique<RequestRouter<TransportType>>(
+    auto ret = std::make_unique<RequestRouter<TransportType>>(
         coordinator::CoordinatorClient<TransportType>(query_io, coordinator_address, {coordinator_address}),
         std::move(io));
+
+    ret->InstallSimulatorTicker(simulator_->GetSimulatorTickClosure());
+
+    return ret;
   }
 };
 

src/query/v2/requests.hpp

@@ -12,6 +12,7 @@
 #pragma once
 
 #include <chrono>
+#include <cstdint>
 #include <iostream>
 #include <map>
 #include <memory>
@@ -21,10 +22,13 @@
 #include <variant>
 #include <vector>
 
+#include <boost/uuid/uuid.hpp>
+
 #include "coordinator/hybrid_logical_clock.hpp"
 #include "storage/v3/id_types.hpp"
 #include "storage/v3/property_value.hpp"
 #include "storage/v3/result.hpp"
+#include "storage/v3/shard.hpp"
 #include "utils/fnv.hpp"
 
 namespace memgraph::msgs {
@@ -340,6 +344,7 @@ enum class StorageView { OLD = 0, NEW = 1 };
 
 struct ScanVerticesRequest {
   Hlc transaction_id;
+  Hlc shard_map_version;
   // This should be optional
   VertexId start_id;
   //  The empty optional means return all of the properties, while an empty list means do not return any properties
@@ -487,6 +492,7 @@ struct UpdateEdgeProp {
  * Vertices
  */
 struct NewVertex {
+  uint64_t idempotency_token;
   std::vector<Label> label_ids;
   PrimaryKey primary_key;
   // This should be a map
@@ -495,6 +501,7 @@ struct NewVertex {
 
 struct CreateVerticesRequest {
   Hlc transaction_id;
+  Hlc shard_map_version;
   std::vector<NewVertex> new_vertices;
 };
 
@@ -528,6 +535,7 @@ struct UpdateVerticesResponse {
 // No need for specifying direction since it has to be in one, and src and dest
 // vertices clearly communicate the direction
 struct NewExpand {
+  uint64_t idempotency_token;
   EdgeId id;
   EdgeType type;
   VertexId src_vertex;
@@ -571,13 +579,65 @@ struct CommitResponse {
   std::optional<ShardError> error;
 };
 
+struct SuggestedSplitInfo {
+  boost::uuids::uuid shard_to_split_uuid;
+  LabelId label_id;
+  PrimaryKey splitting_shard_low_key;
+  PrimaryKey split_key;
+  Hlc shard_version;
+
+  friend bool operator<(const SuggestedSplitInfo &lhs, const SuggestedSplitInfo &rhs) {
+    if (lhs.shard_to_split_uuid != rhs.shard_to_split_uuid) {
+      return lhs.shard_to_split_uuid < rhs.shard_to_split_uuid;
+    }
+
+    /*
+    TODO(tyler) fix this
+    if (lhs.split_key != rhs.split_key) {
+      return lhs.split_key < rhs.split_key;
+    }
+    */
+
+    return lhs.shard_version < rhs.shard_version;
+  }
+};
+
+struct SplitInfo {
+  PrimaryKey split_key;
+  boost::uuids::uuid right_side_uuid;
+  Hlc shard_version;
+};
+
+struct SplitRequest {
+  PrimaryKey split_key;
+  Hlc old_shard_version;
+  Hlc new_lhs_shard_version;
+  Hlc new_rhs_shard_version;
+  std::map<boost::uuids::uuid, boost::uuids::uuid> uuid_mapping;
+};
+
+struct InitializeSplitShard {
+  // TODO(jbajic) Make it unique by solving that is not copyable int std::any
+  std::shared_ptr<storage::v3::Shard> shard;
+  std::map<boost::uuids::uuid, boost::uuids::uuid> uuid_mapping;
+};
+
+struct InitializeSplitShardByUUID {
+  std::shared_ptr<storage::v3::Shard> shard;
+  boost::uuids::uuid shard_uuid;
+};
+
+struct SplitResponse {};
+
 using ReadRequests = std::variant<ExpandOneRequest, GetPropertiesRequest, ScanVerticesRequest>;
 using ReadResponses = std::variant<ExpandOneResponse, GetPropertiesResponse, ScanVerticesResponse>;
 
-using WriteRequests = std::variant<CreateVerticesRequest, DeleteVerticesRequest, UpdateVerticesRequest,
-                                   CreateExpandRequest, DeleteEdgesRequest, UpdateEdgesRequest, CommitRequest>;
-using WriteResponses = std::variant<CreateVerticesResponse, DeleteVerticesResponse, UpdateVerticesResponse,
-                                    CreateExpandResponse, DeleteEdgesResponse, UpdateEdgesResponse, CommitResponse>;
+using WriteRequests =
+    std::variant<CreateVerticesRequest, DeleteVerticesRequest, UpdateVerticesRequest, CreateExpandRequest,
+                 DeleteEdgesRequest, UpdateEdgesRequest, CommitRequest, SplitRequest>;
+using WriteResponses =
+    std::variant<CreateVerticesResponse, DeleteVerticesResponse, UpdateVerticesResponse, CreateExpandResponse,
+                 DeleteEdgesResponse, UpdateEdgesResponse, CommitResponse, SplitResponse>;
 
 }  // namespace memgraph::msgs
 

src/storage/v3/CMakeLists.txt

@@ -18,6 +18,7 @@ set(storage_v3_src_files
     bindings/typed_value.cpp
     expr.cpp
     vertex.cpp
+    splitter.cpp
     request_helper.cpp)
 
 # ######################

src/storage/v3/bindings/db_accessor.hpp

@@ -81,8 +81,10 @@ class DbAccessor final {
   storage::v3::ShardResult<EdgeAccessor> InsertEdge(VertexAccessor *from, VertexAccessor *to,
                                                     const storage::v3::EdgeTypeId &edge_type) {
     static constexpr auto kDummyGid = storage::v3::Gid::FromUint(0);
-    auto maybe_edge = accessor_->CreateEdge(from->Id(storage::v3::View::NEW).GetValue(),
-                                            to->Id(storage::v3::View::NEW).GetValue(), edge_type, kDummyGid);
+    const uint64_t dummy_idempotency_token = 0;
+    auto maybe_edge =
+        accessor_->CreateEdge(from->Id(storage::v3::View::NEW).GetValue(), to->Id(storage::v3::View::NEW).GetValue(),
+                              edge_type, kDummyGid, dummy_idempotency_token);
     if (maybe_edge.HasError()) return {maybe_edge.GetError()};
     return EdgeAccessor(*maybe_edge);
   }

src/storage/v3/config.hpp

@@ -30,6 +30,10 @@ struct Config {
     io::Duration reclamation_interval{};
   } gc;
 
+  struct Split {
+    uint64_t max_shard_vertex_size{500'000};
+  } split;
+
   struct Items {
     bool properties_on_edges{true};
   } items;

src/storage/v3/delta.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -13,12 +13,14 @@
 
 #include <cstdint>
 #include <memory>
+
 #include "storage/v3/edge_ref.hpp"
 #include "storage/v3/id_types.hpp"
 #include "storage/v3/property_value.hpp"
 #include "storage/v3/vertex.hpp"
 #include "storage/v3/vertex_id.hpp"
 #include "utils/logging.hpp"
+#include "utils/synchronized.hpp"
 
 namespace memgraph::storage::v3 {
 
@@ -27,6 +29,11 @@ struct Edge;
 struct Delta;
 struct CommitInfo;
 
+inline uint64_t GetNextDeltaId() noexcept {
+  static utils::Synchronized<uint64_t, utils::SpinLock> delta_id{0};
+  return delta_id.WithLock([](auto &id) { return id++; });
+}
+
 // This class stores one of three pointers (`Delta`, `Vertex` and `Edge`)
 // without using additional memory for storing the type. The type is stored in
 // the pointer itself in the lower bits. All of those structures contain large
@@ -158,46 +165,59 @@ struct Delta {
   struct RemoveInEdgeTag {};
   struct RemoveOutEdgeTag {};
 
-  Delta(DeleteObjectTag /*unused*/, CommitInfo *commit_info, uint64_t command_id)
-      : action(Action::DELETE_OBJECT), commit_info(commit_info), command_id(command_id) {}
+  Delta(DeleteObjectTag /*unused*/, CommitInfo *commit_info, uint64_t delta_id, uint64_t command_id,
+        uint64_t idempotency_token)
+      : action(Action::DELETE_OBJECT),
+        id(delta_id),
+        commit_info(commit_info),
+        command_id(command_id),
+        idempotency_token(idempotency_token) {}
 
-  Delta(RecreateObjectTag /*unused*/, CommitInfo *commit_info, uint64_t command_id)
-      : action(Action::RECREATE_OBJECT), commit_info(commit_info), command_id(command_id) {}
+  Delta(RecreateObjectTag /*unused*/, CommitInfo *commit_info, uint64_t delta_id, uint64_t command_id)
+      : action(Action::RECREATE_OBJECT), id(delta_id), commit_info(commit_info), command_id(command_id) {}
 
-  Delta(AddLabelTag /*unused*/, LabelId label, CommitInfo *commit_info, uint64_t command_id)
-      : action(Action::ADD_LABEL), commit_info(commit_info), command_id(command_id), label(label) {}
+  Delta(AddLabelTag /*unused*/, LabelId label, CommitInfo *commit_info, uint64_t delta_id, uint64_t command_id)
+      : action(Action::ADD_LABEL), id(delta_id), commit_info(commit_info), command_id(command_id), label(label) {}
 
-  Delta(RemoveLabelTag /*unused*/, LabelId label, CommitInfo *commit_info, uint64_t command_id)
-      : action(Action::REMOVE_LABEL), commit_info(commit_info), command_id(command_id), label(label) {}
+  Delta(RemoveLabelTag /*unused*/, LabelId label, CommitInfo *commit_info, uint64_t delta_id, uint64_t command_id)
+      : action(Action::REMOVE_LABEL), id(delta_id), commit_info(commit_info), command_id(command_id), label(label) {}
 
   Delta(SetPropertyTag /*unused*/, PropertyId key, const PropertyValue &value, CommitInfo *commit_info,
-        uint64_t command_id)
-      : action(Action::SET_PROPERTY), commit_info(commit_info), command_id(command_id), property({key, value}) {}
+        uint64_t delta_id, uint64_t command_id)
+      : action(Action::SET_PROPERTY),
+        id(delta_id),
+        commit_info(commit_info),
+        command_id(command_id),
+        property({key, value}) {}
 
   Delta(AddInEdgeTag /*unused*/, EdgeTypeId edge_type, VertexId vertex_id, EdgeRef edge, CommitInfo *commit_info,
-        uint64_t command_id)
+        uint64_t delta_id, uint64_t command_id)
       : action(Action::ADD_IN_EDGE),
+        id(delta_id),
         commit_info(commit_info),
         command_id(command_id),
         vertex_edge({edge_type, std::move(vertex_id), edge}) {}
 
   Delta(AddOutEdgeTag /*unused*/, EdgeTypeId edge_type, VertexId vertex_id, EdgeRef edge, CommitInfo *commit_info,
-        uint64_t command_id)
+        uint64_t delta_id, uint64_t command_id)
       : action(Action::ADD_OUT_EDGE),
+        id(delta_id),
         commit_info(commit_info),
         command_id(command_id),
         vertex_edge({edge_type, std::move(vertex_id), edge}) {}
 
   Delta(RemoveInEdgeTag /*unused*/, EdgeTypeId edge_type, VertexId vertex_id, EdgeRef edge, CommitInfo *commit_info,
-        uint64_t command_id)
+        uint64_t delta_id, uint64_t command_id)
       : action(Action::REMOVE_IN_EDGE),
+        id(delta_id),
         commit_info(commit_info),
         command_id(command_id),
         vertex_edge({edge_type, std::move(vertex_id), edge}) {}
 
   Delta(RemoveOutEdgeTag /*unused*/, EdgeTypeId edge_type, VertexId vertex_id, EdgeRef edge, CommitInfo *commit_info,
-        uint64_t command_id)
+        uint64_t delta_id, uint64_t command_id)
       : action(Action::REMOVE_OUT_EDGE),
+        id(delta_id),
         commit_info(commit_info),
         command_id(command_id),
         vertex_edge({edge_type, std::move(vertex_id), edge}) {}
@@ -226,13 +246,16 @@ struct Delta {
     }
   }
 
-  Action action;
+  friend bool operator==(const Delta &lhs, const Delta &rhs) noexcept { return lhs.id == rhs.id; }
 
+  Action action;
+  uint64_t id;
   // TODO: optimize with in-place copy
   CommitInfo *commit_info;
   uint64_t command_id;
   PreviousPtr prev;
   Delta *next{nullptr};
+  uint64_t idempotency_token;
 
   union {
     LabelId label;

src/storage/v3/indices.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -325,7 +325,7 @@ void LabelIndex::RemoveObsoleteEntries(const uint64_t clean_up_before_timestamp)
   }
 }
 
-LabelIndex::Iterable::Iterator::Iterator(Iterable *self, LabelIndexContainer::iterator index_iterator)
+LabelIndex::Iterable::Iterator::Iterator(Iterable *self, IndexContainer::iterator index_iterator)
     : self_(self),
       index_iterator_(index_iterator),
       current_vertex_accessor_(nullptr, nullptr, nullptr, self_->config_, *self_->vertex_validator_),
@@ -353,7 +353,7 @@ void LabelIndex::Iterable::Iterator::AdvanceUntilValid() {
   }
 }
 
-LabelIndex::Iterable::Iterable(LabelIndexContainer &index_container, LabelId label, View view, Transaction *transaction,
+LabelIndex::Iterable::Iterable(IndexContainer &index_container, LabelId label, View view, Transaction *transaction,
                                Indices *indices, Config::Items config, const VertexValidator &vertex_validator)
     : index_container_(&index_container),
       label_(label),
@@ -465,7 +465,7 @@ void LabelPropertyIndex::RemoveObsoleteEntries(const uint64_t clean_up_before_ti
   }
 }
 
-LabelPropertyIndex::Iterable::Iterator::Iterator(Iterable *self, LabelPropertyIndexContainer::iterator index_iterator)
+LabelPropertyIndex::Iterable::Iterator::Iterator(Iterable *self, IndexContainer::iterator index_iterator)
     : self_(self),
       index_iterator_(index_iterator),
       current_vertex_accessor_(nullptr, nullptr, nullptr, self_->config_, *self_->vertex_validator_),
@@ -526,7 +526,7 @@ const PropertyValue kSmallestMap = PropertyValue(std::map<std::string, PropertyV
 const PropertyValue kSmallestTemporalData =
     PropertyValue(TemporalData{static_cast<TemporalType>(0), std::numeric_limits<int64_t>::min()});
 
-LabelPropertyIndex::Iterable::Iterable(LabelPropertyIndexContainer &index_container, LabelId label, PropertyId property,
+LabelPropertyIndex::Iterable::Iterable(IndexContainer &index_container, LabelId label, PropertyId property,
                                        const std::optional<utils::Bound<PropertyValue>> &lower_bound,
                                        const std::optional<utils::Bound<PropertyValue>> &upper_bound, View view,
                                        Transaction *transaction, Indices *indices, Config::Items config,

src/storage/v3/indices.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -18,6 +18,8 @@
 #include <utility>
 
 #include "storage/v3/config.hpp"
+#include "storage/v3/id_types.hpp"
+#include "storage/v3/key_store.hpp"
 #include "storage/v3/property_value.hpp"
 #include "storage/v3/transaction.hpp"
 #include "storage/v3/vertex_accessor.hpp"
@@ -40,12 +42,18 @@ class LabelIndex {
     bool operator==(const Entry &rhs) const { return vertex == rhs.vertex && timestamp == rhs.timestamp; }
   };
 
+  using IndexType = LabelId;
+
  public:
-  using LabelIndexContainer = std::set<Entry>;
+  using IndexContainer = std::set<Entry>;
 
   LabelIndex(Indices *indices, Config::Items config, const VertexValidator &vertex_validator)
       : indices_(indices), config_(config), vertex_validator_{&vertex_validator} {}
 
+  LabelIndex(Indices *indices, Config::Items config, const VertexValidator &vertex_validator,
+             std::map<LabelId, IndexContainer> &data)
+      : index_{std::move(data)}, indices_(indices), config_(config), vertex_validator_{&vertex_validator} {}
+
   /// @throw std::bad_alloc
   void UpdateOnAddLabel(LabelId label, Vertex *vertex, const Transaction &tx);
 
@@ -63,12 +71,12 @@ class LabelIndex {
 
   class Iterable {
    public:
-    Iterable(LabelIndexContainer &index_container, LabelId label, View view, Transaction *transaction, Indices *indices,
+    Iterable(IndexContainer &index_container, LabelId label, View view, Transaction *transaction, Indices *indices,
              Config::Items config, const VertexValidator &vertex_validator);
 
     class Iterator {
      public:
-      Iterator(Iterable *self, LabelIndexContainer::iterator index_iterator);
+      Iterator(Iterable *self, IndexContainer::iterator index_iterator);
 
       VertexAccessor operator*() const { return current_vertex_accessor_; }
 
@@ -81,7 +89,7 @@ class LabelIndex {
       void AdvanceUntilValid();
 
       Iterable *self_;
-      LabelIndexContainer::iterator index_iterator_;
+      IndexContainer::iterator index_iterator_;
       VertexAccessor current_vertex_accessor_;
       Vertex *current_vertex_;
     };
@@ -90,7 +98,7 @@ class LabelIndex {
     Iterator end() { return {this, index_container_->end()}; }
 
    private:
-    LabelIndexContainer *index_container_;
+    IndexContainer *index_container_;
     LabelId label_;
     View view_;
     Transaction *transaction_;
@@ -114,8 +122,35 @@ class LabelIndex {
 
   void Clear() { index_.clear(); }
 
+  std::map<IndexType, IndexContainer> SplitIndexEntries(const PrimaryKey &split_key) {
+    if (index_.empty()) {
+      return {};
+    }
+
+    // Cloned index entries will contain new index entry iterators, but old
+    // vertices address which need to be adjusted after extracting vertices
+    std::map<IndexType, IndexContainer> cloned_indices;
+    for (auto &[index_type_val, index] : index_) {
+      auto entry_it = index.begin();
+      auto &cloned_indices_container = cloned_indices[index_type_val];
+      while (entry_it != index.end()) {
+        // We need to save the next pointer since the current one will be
+        // invalidated after extract
+        auto next_entry_it = std::next(entry_it);
+        if (entry_it->vertex->first > split_key) {
+          [[maybe_unused]] const auto &[inserted_entry_it, inserted, node] =
+              cloned_indices_container.insert(index.extract(entry_it));
+          MG_ASSERT(inserted, "Failed to extract index entry!");
+        }
+        entry_it = next_entry_it;
+      }
+    }
+
+    return cloned_indices;
+  }
+
  private:
-  std::map<LabelId, LabelIndexContainer> index_;
+  std::map<LabelId, IndexContainer> index_;
   Indices *indices_;
   Config::Items config_;
   const VertexValidator *vertex_validator_;
@@ -133,9 +168,10 @@ class LabelPropertyIndex {
     bool operator<(const PropertyValue &rhs) const;
     bool operator==(const PropertyValue &rhs) const;
   };
+  using IndexType = std::pair<LabelId, PropertyId>;
 
  public:
-  using LabelPropertyIndexContainer = std::set<Entry>;
+  using IndexContainer = std::set<Entry>;
 
   LabelPropertyIndex(Indices *indices, Config::Items config, const VertexValidator &vertex_validator)
       : indices_(indices), config_(config), vertex_validator_{&vertex_validator} {}
@@ -159,14 +195,14 @@ class LabelPropertyIndex {
 
   class Iterable {
    public:
-    Iterable(LabelPropertyIndexContainer &index_container, LabelId label, PropertyId property,
+    Iterable(IndexContainer &index_container, LabelId label, PropertyId property,
              const std::optional<utils::Bound<PropertyValue>> &lower_bound,
              const std::optional<utils::Bound<PropertyValue>> &upper_bound, View view, Transaction *transaction,
              Indices *indices, Config::Items config, const VertexValidator &vertex_validator);
 
     class Iterator {
      public:
-      Iterator(Iterable *self, LabelPropertyIndexContainer::iterator index_iterator);
+      Iterator(Iterable *self, IndexContainer::iterator index_iterator);
 
       VertexAccessor operator*() const { return current_vertex_accessor_; }
 
@@ -179,7 +215,7 @@ class LabelPropertyIndex {
       void AdvanceUntilValid();
 
       Iterable *self_;
-      LabelPropertyIndexContainer::iterator index_iterator_;
+      IndexContainer::iterator index_iterator_;
       VertexAccessor current_vertex_accessor_;
       Vertex *current_vertex_;
     };
@@ -188,7 +224,7 @@ class LabelPropertyIndex {
     Iterator end();
 
    private:
-    LabelPropertyIndexContainer *index_container_;
+    IndexContainer *index_container_;
     LabelId label_;
     PropertyId property_;
     std::optional<utils::Bound<PropertyValue>> lower_bound_;
@@ -229,8 +265,35 @@ class LabelPropertyIndex {
 
   void Clear() { index_.clear(); }
 
+  std::map<IndexType, IndexContainer> SplitIndexEntries(const PrimaryKey &split_key) {
+    if (index_.empty()) {
+      return {};
+    }
+
+    // Cloned index entries will contain new index entry iterators, but old
+    // vertices address which need to be adjusted after extracting vertices
+    std::map<IndexType, IndexContainer> cloned_indices;
+    for (auto &[index_type_val, index] : index_) {
+      auto entry_it = index.begin();
+      auto &cloned_index_container = cloned_indices[index_type_val];
+      while (entry_it != index.end()) {
+        // We need to save the next pointer since the current one will be
+        // invalidated after extract
+        auto next_entry_it = std::next(entry_it);
+        if (entry_it->vertex->first > split_key) {
+          [[maybe_unused]] const auto &[inserted_entry_it, inserted, node] =
+              cloned_index_container.insert(index.extract(entry_it));
+          MG_ASSERT(inserted, "Failed to extract index entry!");
+        }
+        entry_it = next_entry_it;
+      }
+    }
+
+    return cloned_indices;
+  }
+
  private:
-  std::map<std::pair<LabelId, PropertyId>, LabelPropertyIndexContainer> index_;
+  std::map<std::pair<LabelId, PropertyId>, IndexContainer> index_;
   Indices *indices_;
   Config::Items config_;
   const VertexValidator *vertex_validator_;

src/storage/v3/mvcc.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -107,9 +107,9 @@ inline bool PrepareForWrite(Transaction *transaction, TObj *object) {
 /// and is primarily used to create the first delta for an object (that must be
 /// a `DELETE_OBJECT` delta).
 /// @throw std::bad_alloc
-inline Delta *CreateDeleteObjectDelta(Transaction *transaction) {
-  return &transaction->deltas.emplace_back(Delta::DeleteObjectTag(), transaction->commit_info.get(),
-                                           transaction->command_id);
+inline Delta *CreateDeleteObjectDelta(Transaction *transaction, uint64_t idempotency_token) {
+  return &transaction->deltas.emplace_back(Delta::DeleteObjectTag(), transaction->commit_info.get(), GetNextDeltaId(),
+                                           transaction->command_id, idempotency_token);
 }
 
 /// This function creates a delta in the transaction for the object and links
@@ -119,7 +119,7 @@ template <typename TObj, class... Args>
 requires utils::SameAsAnyOf<TObj, Edge, Vertex>
 inline void CreateAndLinkDelta(Transaction *transaction, TObj *object, Args &&...args) {
   auto delta = &transaction->deltas.emplace_back(std::forward<Args>(args)..., transaction->commit_info.get(),
-                                                 transaction->command_id);
+                                                 GetNextDeltaId(), transaction->command_id);
   auto *delta_holder = GetDeltaHolder(object);
 
   // The operations are written in such order so that both `next` and `prev`

src/storage/v3/shard.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -18,14 +18,14 @@
 #include <memory>
 #include <mutex>
 #include <optional>
-#include <variant>
 
 #include <bits/ranges_algo.h>
-#include <gflags/gflags.h>
 #include <spdlog/spdlog.h>
 
 #include "io/network/endpoint.hpp"
 #include "io/time.hpp"
+#include "storage/v3/delta.hpp"
+#include "storage/v3/edge.hpp"
 #include "storage/v3/edge_accessor.hpp"
 #include "storage/v3/id_types.hpp"
 #include "storage/v3/indices.hpp"
@@ -36,6 +36,7 @@
 #include "storage/v3/result.hpp"
 #include "storage/v3/schema_validator.hpp"
 #include "storage/v3/transaction.hpp"
+#include "storage/v3/value_conversions.hpp"
 #include "storage/v3/vertex.hpp"
 #include "storage/v3/vertex_accessor.hpp"
 #include "storage/v3/view.hpp"
@@ -323,30 +324,90 @@ bool VerticesIterable::Iterator::operator==(const Iterator &other) const {
 }
 
 Shard::Shard(const LabelId primary_label, const PrimaryKey min_primary_key,
-             const std::optional<PrimaryKey> max_primary_key, std::vector<SchemaProperty> schema, Config config,
-             std::unordered_map<uint64_t, std::string> id_to_name)
+             const std::optional<PrimaryKey> max_primary_key, std::vector<SchemaProperty> schema, Hlc shard_version,
+             Config config, std::unordered_map<uint64_t, std::string> id_to_name)
     : primary_label_{primary_label},
       min_primary_key_{min_primary_key},
       max_primary_key_{max_primary_key},
+      shard_version_{shard_version},
       schema_validator_{schemas_, name_id_mapper_},
       vertex_validator_{schema_validator_, primary_label},
       indices_{config.items, vertex_validator_},
       isolation_level_{config.transaction.isolation_level},
       config_{config},
-      uuid_{utils::GenerateUUID()},
-      epoch_id_{utils::GenerateUUID()},
-      global_locker_{file_retainer_.AddLocker()} {
+      shard_splitter_(primary_label, vertices_, edges_, start_logical_id_to_transaction_, indices_, config_, schema,
+                      name_id_mapper_) {
+  spdlog::trace("Shard constructed with low key {} (path 1)", min_primary_key_.back());
   CreateSchema(primary_label_, schema);
   StoreMapping(std::move(id_to_name));
 }
 
+Shard::Shard(LabelId primary_label, PrimaryKey min_primary_key, std::optional<PrimaryKey> max_primary_key,
+             std::vector<SchemaProperty> schema, VertexContainer &&vertices, EdgeContainer &&edges,
+             std::map<uint64_t, std::unique_ptr<Transaction>> &&start_logical_id_to_transaction, const Config &config,
+             const std::unordered_map<uint64_t, std::string> &id_to_name, const Hlc shard_version)
+    : primary_label_{primary_label},
+      min_primary_key_{min_primary_key},
+      max_primary_key_{max_primary_key},
+      vertices_(std::move(vertices)),
+      edges_(std::move(edges)),
+      shard_version_(shard_version),
+      schema_validator_{schemas_, name_id_mapper_},
+      vertex_validator_{schema_validator_, primary_label},
+      indices_{config.items, vertex_validator_},
+      isolation_level_{config.transaction.isolation_level},
+      config_{config},
+      start_logical_id_to_transaction_(std::move(start_logical_id_to_transaction)),
+      shard_splitter_(primary_label, vertices_, edges_, start_logical_id_to_transaction_, indices_, config_, schema,
+                      name_id_mapper_) {
+  spdlog::trace("Shard constructed with low key {} (path 2)", min_primary_key_.back());
+  CreateSchema(primary_label_, schema);
+  StoreMapping(id_to_name);
+}
+
+Shard::Shard(LabelId primary_label, PrimaryKey min_primary_key, std::optional<PrimaryKey> max_primary_key,
+             std::vector<SchemaProperty> schema, VertexContainer &&vertices,
+             std::map<uint64_t, std::unique_ptr<Transaction>> &&start_logical_id_to_transaction, const Config &config,
+             const std::unordered_map<uint64_t, std::string> &id_to_name, const Hlc shard_version)
+    : primary_label_{primary_label},
+      min_primary_key_{min_primary_key},
+      max_primary_key_{max_primary_key},
+      vertices_(std::move(vertices)),
+      shard_version_(shard_version),
+      schema_validator_{schemas_, name_id_mapper_},
+      vertex_validator_{schema_validator_, primary_label},
+      indices_{config.items, vertex_validator_},
+      isolation_level_{config.transaction.isolation_level},
+      config_{config},
+      start_logical_id_to_transaction_(std::move(start_logical_id_to_transaction)),
+      shard_splitter_(primary_label, vertices_, edges_, start_logical_id_to_transaction_, indices_, config_, schema,
+                      name_id_mapper_) {
+  spdlog::trace("Shard constructed with low key {} (path 3)", min_primary_key_.back());
+  CreateSchema(primary_label_, schema);
+  StoreMapping(id_to_name);
+}
+
 Shard::~Shard() {}
 
+std::unique_ptr<Shard> Shard::FromSplitData(SplitData &&split_data) {
+  if (split_data.config.items.properties_on_edges) [[likely]] {
+    return std::make_unique<Shard>(split_data.primary_label, split_data.min_primary_key, split_data.max_primary_key,
+                                   split_data.schema, std::move(split_data.vertices), std::move(*split_data.edges),
+                                   std::move(split_data.transactions), split_data.config, split_data.id_to_name,
+                                   split_data.shard_version);
+  }
+  const auto low_key_int = split_data.min_primary_key.back();
+  spdlog::trace("Shard constructed FromSplitData with low key {}", low_key_int);
+  return std::make_unique<Shard>(split_data.primary_label, split_data.min_primary_key, split_data.max_primary_key,
+                                 split_data.schema, std::move(split_data.vertices), std::move(split_data.transactions),
+                                 split_data.config, split_data.id_to_name, split_data.shard_version);
+}
+
 Shard::Accessor::Accessor(Shard &shard, Transaction &transaction)
     : shard_(&shard), transaction_(&transaction), config_(shard_->config_.items) {}
 
-ShardResult<VertexAccessor> Shard::Accessor::CreateVertexAndValidate(
-    const std::vector<LabelId> &labels, const PrimaryKey &primary_properties,
+std::optional<ShardError> Shard::Accessor::CreateVertexAndValidate(
+    const uint64_t idempotency_token, const std::vector<LabelId> &labels, const PrimaryKey &primary_properties,
     const std::vector<std::pair<PropertyId, PropertyValue>> &properties) {
   OOMExceptionEnabler oom_exception;
   const auto schema = shard_->GetSchema(shard_->primary_label_)->second;
@@ -357,14 +418,19 @@ ShardResult<VertexAccessor> Shard::Accessor::CreateVertexAndValidate(
     return {std::move(maybe_schema_violation.GetError())};
   }
 
-  auto *delta = CreateDeleteObjectDelta(transaction_);
+  auto *delta = CreateDeleteObjectDelta(transaction_, idempotency_token);
   auto [it, inserted] = shard_->vertices_.emplace(primary_properties, VertexData{delta});
-  delta->prev.Set(&*it);
 
-  VertexAccessor vertex_acc{&*it, transaction_, &shard_->indices_, config_, shard_->vertex_validator_};
   if (!inserted) {
+    if (it->second.delta->idempotency_token == idempotency_token) {
+      // this is a benign race condition due to request retries - signal success
+      return std::nullopt;
+    }
     return SHARD_ERROR(ErrorCode::VERTEX_ALREADY_INSERTED);
   }
+
+  delta->prev.Set(&*it);
+  VertexAccessor vertex_acc{&*it, transaction_, &shard_->indices_, config_, shard_->vertex_validator_};
   MG_ASSERT(it != shard_->vertices_.end(), "Invalid Vertex accessor!");
 
   // TODO(jbajic) Improve, maybe delay index update
@@ -381,7 +447,8 @@ ShardResult<VertexAccessor> Shard::Accessor::CreateVertexAndValidate(
       return {err.GetError()};
     }
   }
-  return vertex_acc;
+
+  return std::nullopt;
 }
 
 std::optional<VertexAccessor> Shard::Accessor::FindVertex(std::vector<PropertyValue> primary_key, View view) {
@@ -436,7 +503,7 @@ ShardResult<std::optional<std::pair<VertexAccessor, std::vector<EdgeAccessor>>>>
   }
 
   std::vector<EdgeAccessor> deleted_edges;
-  const VertexId vertex_id{shard_->primary_label_, *vertex->PrimaryKey(View::OLD)};  // TODO Replace
+  const VertexId vertex_id{shard_->primary_label_, *vertex->PrimaryKey(View::OLD)};
   for (const auto &item : in_edges) {
     auto [edge_type, from_vertex, edge] = item;
     EdgeAccessor e(edge, edge_type, from_vertex, vertex_id, transaction_, &shard_->indices_, config_);
@@ -481,7 +548,8 @@ ShardResult<std::optional<std::pair<VertexAccessor, std::vector<EdgeAccessor>>>>
 }
 
 ShardResult<EdgeAccessor> Shard::Accessor::CreateEdge(VertexId from_vertex_id, VertexId to_vertex_id,
-                                                      const EdgeTypeId edge_type, const Gid gid) {
+                                                      const EdgeTypeId edge_type, const Gid gid,
+                                                      const uint64_t idempotency_token) {
   OOMExceptionEnabler oom_exception;
   Vertex *from_vertex{nullptr};
   Vertex *to_vertex{nullptr};
@@ -515,7 +583,7 @@ ShardResult<EdgeAccessor> Shard::Accessor::CreateEdge(VertexId from_vertex_id, V
 
   EdgeRef edge(gid);
   if (config_.properties_on_edges) {
-    auto *delta = CreateDeleteObjectDelta(transaction_);
+    auto *delta = CreateDeleteObjectDelta(transaction_, idempotency_token);
     auto [it, inserted] = shard_->edges_.emplace(gid, Edge{gid, delta});
     MG_ASSERT(inserted, "The edge must be inserted here!");
     MG_ASSERT(it != shard_->edges_.end(), "Invalid Edge accessor!");
@@ -664,9 +732,16 @@ void Shard::Accessor::Commit(coordinator::Hlc commit_timestamp) {
   MG_ASSERT(!transaction_->must_abort, "The transaction can't be committed!");
   MG_ASSERT(transaction_->start_timestamp.logical_id < commit_timestamp.logical_id,
             "Commit timestamp must be older than start timestamp!");
-  MG_ASSERT(!transaction_->commit_info->is_locally_committed, "The transaction is already committed!");
-  transaction_->commit_info->start_or_commit_timestamp = commit_timestamp;
-  transaction_->commit_info->is_locally_committed = true;
+  if (transaction_->commit_info->is_locally_committed) {
+    MG_ASSERT(transaction_->commit_info->start_or_commit_timestamp == commit_timestamp);
+    spdlog::debug(
+        "Shard::Accessor::Commit called for already-committed transaction, probably due to a retry request that timed "
+        "out for the RequestRouter");
+  } else {
+    MG_ASSERT(transaction_->commit_info->start_or_commit_timestamp != commit_timestamp);
+    transaction_->commit_info->start_or_commit_timestamp = commit_timestamp;
+    transaction_->commit_info->is_locally_committed = true;
+  }
 }
 
 void Shard::Accessor::Abort() {
@@ -897,6 +972,8 @@ StorageInfo Shard::GetInfo() const {
   return {vertex_count, edge_count_, average_degree, utils::GetMemoryUsage()};
 }
 
+Hlc Shard::Version() const { return shard_version_; }
+
 VerticesIterable Shard::Accessor::Vertices(LabelId label, View view) {
   return VerticesIterable(shard_->indices_.label_index.Vertices(label, view, transaction_));
 }
@@ -1048,6 +1125,47 @@ void Shard::StoreMapping(std::unordered_map<uint64_t, std::string> id_to_name) {
   name_id_mapper_.StoreMapping(std::move(id_to_name));
 }
 
+std::optional<ShardSuggestedSplitInfo> Shard::ShouldSplit() const noexcept {
+  if (vertices_.size() >= config_.split.max_shard_vertex_size) {
+    spdlog::debug("Shard::ShouldSplit is signalling that the split process may begin");
+    auto mid_elem = vertices_.begin();
+
+    // TODO(tyler) the first time we calculate the split point, we should store it so that we don't have to
+    // iterate over half of the entire index each time Cron is run until the split succeeds.
+    std::ranges::advance(mid_elem, static_cast<VertexContainer::difference_type>(vertices_.size() / 2));
+
+    spdlog::debug("Shard::ShouldSplit returning split key with end Value {} for shard version {}",
+                  mid_elem->first.back(), shard_version_.logical_id);
+
+    return ShardSuggestedSplitInfo{
+        .label_id = PrimaryLabel(),
+        .splitting_shard_low_key = min_primary_key_,
+        .split_key = mid_elem->first,
+        .shard_version = shard_version_,
+    };
+  }
+
+  spdlog::trace("Shard not splitting because we have {} vertices, lower than split threshold of {}", vertices_.size(),
+                config_.split.max_shard_vertex_size);
+  return std::nullopt;
+}
+
+std::optional<SplitData> Shard::PerformSplit(const PrimaryKey &split_key, const Hlc old_shard_version,
+                                             const Hlc new_lhs_shard_version, const Hlc new_rhs_shard_version) {
+  if (old_shard_version != shard_version_) {
+    spdlog::debug("Shard::PerformSplit - Curent shard version {} does not match given {}", shard_version_,
+                  old_shard_version);
+    return std::nullopt;
+  }
+  spdlog::debug("Shard::PerformSplit - splitting from shard version {} to versions {} and {}", old_shard_version,
+                new_lhs_shard_version, new_rhs_shard_version);
+
+  shard_version_ = new_lhs_shard_version;
+  const auto old_max_key = max_primary_key_;
+  max_primary_key_ = split_key;
+  return shard_splitter_.SplitShard(split_key, old_max_key, new_rhs_shard_version);
+}
+
 bool Shard::IsVertexBelongToShard(const VertexId &vertex_id) const {
   return vertex_id.primary_label == primary_label_ && vertex_id.primary_key >= min_primary_key_ &&
          (!max_primary_key_.has_value() || vertex_id.primary_key < *max_primary_key_);

src/storage/v3/shard.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -14,12 +14,15 @@
 #include <cstdint>
 #include <filesystem>
 #include <map>
+#include <memory>
 #include <numeric>
 #include <optional>
 #include <shared_mutex>
 #include <variant>
 #include <vector>
 
+#include <boost/uuid/uuid.hpp>
+
 #include "coordinator/hybrid_logical_clock.hpp"
 #include "io/network/endpoint.hpp"
 #include "io/time.hpp"
@@ -37,11 +40,13 @@
 #include "storage/v3/result.hpp"
 #include "storage/v3/schema_validator.hpp"
 #include "storage/v3/schemas.hpp"
+#include "storage/v3/splitter.hpp"
 #include "storage/v3/transaction.hpp"
 #include "storage/v3/vertex.hpp"
 #include "storage/v3/vertex_accessor.hpp"
 #include "storage/v3/vertex_id.hpp"
 #include "storage/v3/view.hpp"
+#include "utils/concrete_msg_sender.hpp"
 #include "utils/exceptions.hpp"
 #include "utils/file_locker.hpp"
 #include "utils/on_scope_exit.hpp"
@@ -53,6 +58,8 @@
 
 namespace memgraph::storage::v3 {
 
+using coordinator::Hlc;
+
 // The storage is based on this paper:
 // https://db.in.tum.de/~muehlbau/papers/mvcc.pdf
 // The paper implements a fully serializable storage, in our implementation we
@@ -182,13 +189,31 @@ struct StorageInfo {
   uint64_t memory_usage;
 };
 
+struct ShardSuggestedSplitInfo {
+  boost::uuids::uuid shard_to_split_uuid;
+  LabelId label_id;
+  PrimaryKey splitting_shard_low_key;
+  PrimaryKey split_key;
+  Hlc shard_version;
+};
+
 class Shard final {
  public:
   /// @throw std::system_error
   /// @throw std::bad_alloc
-  explicit Shard(LabelId primary_label, PrimaryKey min_primary_key, std::optional<PrimaryKey> max_primary_key,
-                 std::vector<SchemaProperty> schema, Config config = Config(),
-                 std::unordered_map<uint64_t, std::string> id_to_name = {});
+  Shard(LabelId primary_label, PrimaryKey min_primary_key, std::optional<PrimaryKey> max_primary_key,
+        std::vector<SchemaProperty> schema, Hlc shard_version, Config config = Config(),
+        std::unordered_map<uint64_t, std::string> id_to_name = {});
+
+  Shard(LabelId primary_label, PrimaryKey min_primary_key, std::optional<PrimaryKey> max_primary_key,
+        std::vector<SchemaProperty> schema, VertexContainer &&vertices, EdgeContainer &&edges,
+        std::map<uint64_t, std::unique_ptr<Transaction>> &&start_logical_id_to_transaction, const Config &config,
+        const std::unordered_map<uint64_t, std::string> &id_to_name, Hlc shard_version);
+
+  Shard(LabelId primary_label, PrimaryKey min_primary_key, std::optional<PrimaryKey> max_primary_key,
+        std::vector<SchemaProperty> schema, VertexContainer &&vertices,
+        std::map<uint64_t, std::unique_ptr<Transaction>> &&start_logical_id_to_transaction, const Config &config,
+        const std::unordered_map<uint64_t, std::string> &id_to_name, Hlc shard_version);
 
   Shard(const Shard &) = delete;
   Shard(Shard &&) noexcept = delete;
@@ -196,6 +221,8 @@ class Shard final {
   Shard operator=(Shard &&) noexcept = delete;
   ~Shard();
 
+  static std::unique_ptr<Shard> FromSplitData(SplitData &&split_data);
+
   class Accessor final {
    private:
     friend class Shard;
@@ -204,8 +231,9 @@ class Shard final {
 
    public:
     /// @throw std::bad_alloc
-    ShardResult<VertexAccessor> CreateVertexAndValidate(
-        const std::vector<LabelId> &labels, const std::vector<PropertyValue> &primary_properties,
+    std::optional<ShardError> CreateVertexAndValidate(
+        const uint64_t idempotency_token, const std::vector<LabelId> &labels,
+        const std::vector<PropertyValue> &primary_properties,
         const std::vector<std::pair<PropertyId, PropertyValue>> &properties);
 
     std::optional<VertexAccessor> FindVertex(std::vector<PropertyValue> primary_key, View view);
@@ -266,7 +294,8 @@ class Shard final {
         VertexAccessor *vertex);
 
     /// @throw std::bad_alloc
-    ShardResult<EdgeAccessor> CreateEdge(VertexId from_vertex_id, VertexId to_vertex_id, EdgeTypeId edge_type, Gid gid);
+    ShardResult<EdgeAccessor> CreateEdge(VertexId from_vertex_id, VertexId to_vertex_id, EdgeTypeId edge_type, Gid gid,
+                                         const uint64_t idempotency_token);
 
     /// Accessor to the deleted edge if a deletion took place, std::nullopt otherwise
     /// @throw std::bad_alloc
@@ -329,6 +358,10 @@ class Shard final {
 
   LabelId PrimaryLabel() const;
 
+  Hlc Version() const;
+
+  PrimaryKey LowKey() const { return min_primary_key_; }
+
   [[nodiscard]] bool IsVertexBelongToShard(const VertexId &vertex_id) const;
 
   /// @throw std::bad_alloc
@@ -360,6 +393,11 @@ class Shard final {
 
   void StoreMapping(std::unordered_map<uint64_t, std::string> id_to_name);
 
+  std::optional<ShardSuggestedSplitInfo> ShouldSplit() const noexcept;
+
+  std::optional<SplitData> PerformSplit(const PrimaryKey &split_key, Hlc old_shard_version,
+                                        const Hlc new_lhs_shard_version, const Hlc new_rhs_shard_version);
+
  private:
   Transaction &GetTransaction(coordinator::Hlc start_timestamp, IsolationLevel isolation_level);
 
@@ -377,6 +415,7 @@ class Shard final {
   // list is used only when properties are enabled for edges. Because of that we
   // keep a separate count of edges that is always updated.
   uint64_t edge_count_{0};
+  Hlc shard_version_{};
 
   SchemaValidator schema_validator_;
   VertexValidator vertex_validator_;
@@ -396,41 +435,10 @@ class Shard final {
   // storage.
   std::list<Gid> deleted_edges_;
 
-  // UUID used to distinguish snapshots and to link snapshots to WALs
-  std::string uuid_;
-  // Sequence number used to keep track of the chain of WALs.
-  uint64_t wal_seq_num_{0};
-
-  // UUID to distinguish different main instance runs for replication process
-  // on SAME storage.
-  // Multiple instances can have same storage UUID and be MAIN at the same time.
-  // We cannot compare commit timestamps of those instances if one of them
-  // becomes the replica of the other so we use epoch_id_ as additional
-  // discriminating property.
-  // Example of this:
-  // We have 2 instances of the same storage, S1 and S2.
-  // S1 and S2 are MAIN and accept their own commits and write them to the WAL.
-  // At the moment when S1 commited a transaction with timestamp 20, and S2
-  // a different transaction with timestamp 15, we change S2's role to REPLICA
-  // and register it on S1.
-  // Without using the epoch_id, we don't know that S1 and S2 have completely
-  // different transactions, we think that the S2 is behind only by 5 commits.
-  std::string epoch_id_;
-  // History of the previous epoch ids.
-  // Each value consists of the epoch id along the last commit belonging to that
-  // epoch.
-  std::deque<std::pair<std::string, uint64_t>> epoch_history_;
-
-  uint64_t wal_unsynced_transactions_{0};
-
-  utils::FileRetainer file_retainer_;
-
-  // Global locker that is used for clients file locking
-  utils::FileRetainer::FileLocker global_locker_;
-
   // Holds all of the (in progress, committed and aborted) transactions that are read or write to this shard, but
   // haven't been cleaned up yet
   std::map<uint64_t, std::unique_ptr<Transaction>> start_logical_id_to_transaction_{};
+  Splitter shard_splitter_;
   bool has_any_transaction_aborted_since_last_gc{false};
 };
 

src/storage/v3/shard_manager.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -12,8 +12,9 @@
 #pragma once
 
 #include <queue>
-#include <set>
 #include <unordered_map>
+#include <variant>
+#include <vector>
 
 #include <boost/functional/hash.hpp>
 #include <boost/uuid/uuid.hpp>
@@ -31,6 +32,8 @@
 #include "storage/v3/shard.hpp"
 #include "storage/v3/shard_rsm.hpp"
 #include "storage/v3/shard_worker.hpp"
+#include "storage/v3/value_conversions.hpp"
+#include "utils/variant_helpers.hpp"
 
 namespace memgraph::storage::v3 {
 
@@ -42,7 +45,6 @@ using coordinator::HeartbeatRequest;
 using coordinator::HeartbeatResponse;
 using io::Address;
 using io::Duration;
-using io::Message;
 using io::RequestId;
 using io::ResponseFuture;
 using io::Time;
@@ -74,14 +76,17 @@ static_assert(kMinimumCronInterval < kMaximumCronInterval,
 /// * reconciling the storage engine's local configuration with the Coordinator's
 ///   intentions for how it should participate in multiple raft clusters
 /// * replying to heartbeat requests to the Coordinator
-/// * routing incoming messages to the appropriate sRSM
+/// * routing incoming messages to the appropriate RSM
 ///
 /// Every storage engine has exactly one RsmEngine.
 template <typename IoImpl>
 class ShardManager {
  public:
-  ShardManager(io::Io<IoImpl> io, size_t shard_worker_threads, Address coordinator_leader)
-      : io_(io), coordinator_leader_(coordinator_leader) {
+  ShardManager(io::Io<IoImpl> io, size_t shard_worker_threads, Address coordinator_leader,
+               bool serialize_shard_splits_for_determinisim = false)
+      : io_(io),
+        coordinator_leader_(coordinator_leader),
+        serialize_shard_splits_for_determinisim_(serialize_shard_splits_for_determinisim) {
     MG_ASSERT(shard_worker_threads >= 1);
 
     for (int i = 0; i < shard_worker_threads; i++) {
@@ -149,7 +154,7 @@ class ShardManager {
   /// Periodic protocol maintenance. Returns the time that Cron should be called again
   /// in the future.
   Time Cron() {
-    spdlog::info("running ShardManager::Cron, address {}", io_.GetAddress().ToString());
+    spdlog::trace("ShardManager running Cron, address {}", io_.GetAddress().ToString());
     Time now = io_.Now();
 
     if (now >= next_reconciliation_) {
@@ -174,7 +179,39 @@ class ShardManager {
   /// Returns the Address for our underlying Io implementation
   Address GetAddress() { return io_.GetAddress(); }
 
-  void Receive(ShardManagerMessages &&smm, RequestId request_id, Address from) {}
+  void InitializeSplitShard(msgs::InitializeSplitShard &&init_split_shard) {
+    spdlog::trace("ShardManager received InitializeSplitShard message");
+    for (const auto &[from_uuid, new_uuid] : init_split_shard.uuid_mapping) {
+      bool has_source = rsm_worker_mapping_.contains(from_uuid);
+      if (has_source) {
+        const auto low_key = init_split_shard.shard->LowKey();
+        coordinator::ShardId new_shard_id = std::make_pair(init_split_shard.shard->PrimaryLabel(), low_key);
+
+        spdlog::debug("ShardManager initialized split shard {} with uuid {} and low key {}",
+                      init_split_shard.shard->Version().logical_id, new_uuid, low_key.back());
+        msgs::InitializeSplitShardByUUID msg{.shard = std::move(init_split_shard.shard), .shard_uuid = new_uuid};
+        SendToWorkerByUuid(new_uuid, std::move(msg));
+
+        initialized_but_not_confirmed_rsm_.emplace(new_uuid, new_shard_id);
+        break;
+      }
+    }
+  }
+
+  void Receive(ShardManagerMessages &&smm, RequestId request_id, Address from) {
+    std::visit(utils::Overloaded{[this](msgs::SuggestedSplitInfo &&split_info) {
+                                   spdlog::debug(
+                                       "ShardManager adding new suggested split info to the pending_splits_ structure");
+                                   pending_splits_.emplace(std::move(split_info));
+                                 },
+                                 [this](msgs::InitializeSplitShard &&init_split_shard) {
+                                   // TODO(jbajic) remove pending split for this completed split
+                                   // TODO(jbajic) Add new shard to initialized but not confirmed rsm
+                                   spdlog::debug("ShardManager received a new split shard that it will now initialize");
+                                   InitializeSplitShard(std::move(init_split_shard));
+                                 }},
+               std::move(smm));
+  }
 
   void Route(ShardMessages &&sm, RequestId request_id, Address to, Address from) {
     Address address = io_.GetAddress();
@@ -201,14 +238,13 @@ class ShardManager {
   std::vector<shard_worker::Queue> workers_;
   std::vector<std::jthread> worker_handles_;
   std::vector<size_t> worker_rsm_counts_;
+  std::set<msgs::SuggestedSplitInfo> pending_splits_;
   std::unordered_map<uuid, size_t, boost::hash<boost::uuids::uuid>> rsm_worker_mapping_;
   Time next_reconciliation_ = Time::min();
   Address coordinator_leader_;
   std::optional<ResponseFuture<WriteResponse<CoordinatorWriteResponses>>> heartbeat_res_;
-
-  // TODO(tyler) over time remove items from initialized_but_not_confirmed_rsm_
-  // after the Coordinator is clearly aware of them
-  std::set<boost::uuids::uuid> initialized_but_not_confirmed_rsm_;
+  std::map<boost::uuids::uuid, coordinator::ShardId> initialized_but_not_confirmed_rsm_;
+  bool serialize_shard_splits_for_determinisim_;
 
   void Reconciliation() {
     if (heartbeat_res_.has_value()) {
@@ -218,18 +254,18 @@ class ShardManager {
         heartbeat_res_.reset();
 
         if (response_result.HasError()) {
-          spdlog::error("SM timed out while trying to reach C");
+          spdlog::info("ShardManager timed out while trying to reach Coordinator");
         } else {
           auto response_envelope = response_result.GetValue();
           WriteResponse<CoordinatorWriteResponses> wr = response_envelope.message;
 
           if (wr.retry_leader.has_value()) {
-            spdlog::info("SM redirected to new C leader");
+            spdlog::info("ShardManager redirected to new Coordinator leader");
             coordinator_leader_ = wr.retry_leader.value();
           } else if (wr.success) {
             CoordinatorWriteResponses cwr = wr.write_return;
             HeartbeatResponse hr = std::get<HeartbeatResponse>(cwr);
-            spdlog::info("SM received heartbeat response from C");
+            spdlog::info("ShardManager received heartbeat response from Coordinator");
 
             EnsureShardsInitialized(hr);
           }
@@ -242,28 +278,35 @@ class ShardManager {
     HeartbeatRequest req{
         .from_storage_manager = GetAddress(),
         .initialized_rsms = initialized_but_not_confirmed_rsm_,
+        .suggested_splits = std::move(pending_splits_),
     };
 
     CoordinatorWriteRequests cwr = req;
     WriteRequest<CoordinatorWriteRequests> ww;
     ww.operation = cwr;
 
-    spdlog::info("SM sending heartbeat to coordinator {}", coordinator_leader_.ToString());
+    spdlog::info("ShardManager sending heartbeat to coordinator {} with {} initialized rsms",
+                 coordinator_leader_.ToString(), initialized_but_not_confirmed_rsm_.size());
     heartbeat_res_.emplace(std::move(
         io_.template Request<WriteRequest<CoordinatorWriteRequests>, WriteResponse<CoordinatorWriteResponses>>(
             coordinator_leader_, ww)));
-    spdlog::info("SM sent heartbeat");
+    spdlog::info("ShardManager sent heartbeat");
   }
 
   void EnsureShardsInitialized(HeartbeatResponse hr) {
+    for (const auto &acknowledged_rsm : hr.acknowledged_initialized_rsms) {
+      initialized_but_not_confirmed_rsm_.erase(acknowledged_rsm);
+    }
+
     for (const auto &to_init : hr.shards_to_initialize) {
-      initialized_but_not_confirmed_rsm_.emplace(to_init.uuid);
+      coordinator::ShardId new_shard_id = std::make_pair(to_init.label_id, to_init.min_key);
+      spdlog::trace("ShardManager has been told to initialize shard rsm with uuid {}", to_init.uuid);
+      initialized_but_not_confirmed_rsm_.emplace(to_init.uuid, new_shard_id);
 
       if (rsm_worker_mapping_.contains(to_init.uuid)) {
-        // it's not a bug for the coordinator to send us UUIDs that we have
-        // already created, because there may have been lag that caused
-        // the coordinator not to hear back from us.
-        return;
+        spdlog::info(
+            "ShardManager forwarding shard intialization request to worker despite a mapping already existing. This "
+            "can happen due to benign race conditions.");
       }
 
       size_t worker_index = UuidToWorkerIndex(to_init.uuid);
@@ -272,6 +315,54 @@ class ShardManager {
 
       rsm_worker_mapping_.emplace(to_init.uuid, worker_index);
     }
+
+    for (const auto &to_split : hr.shards_to_split) {
+      for (const auto &[source, destination] : to_split.uuid_mapping) {
+        if (rsm_worker_mapping_.contains(source)) {
+          // Create the proper layered request providing a Raft write
+          // request to the local shard rsm, under the guess that it is
+          // the current leader. Most of the time this will be an incorrect
+          // guess, but it will eventually succeed when the right ShardManager
+          // happens to send the message to the leader shard that is local.
+          // This is done to avoid blocking on an RsmClient or maintaining
+          // complex async request logic. It's fine to fire-and-forget because
+          // it is rare and will eventually succeed.
+          msgs::WriteRequests split_request_1 =
+              msgs::SplitRequest{.split_key = conversions::ConvertValueVector(to_split.split_key),
+                                 .old_shard_version = to_split.old_shard_version,
+                                 .new_lhs_shard_version = to_split.new_lhs_shard_version,
+                                 .new_rhs_shard_version = to_split.new_rhs_shard_version,
+                                 .uuid_mapping = to_split.uuid_mapping};
+
+          WriteRequest<msgs::WriteRequests> split_request_2;
+          split_request_2.operation = split_request_1;
+          ShardMessages split_request_3 = split_request_2;
+
+          const Address our_address = io_.GetAddress();
+          Address shard_address = our_address;
+          shard_address.unique_id = source;
+
+          shard_worker::RouteMessage shard_worker_message = {
+              .message = split_request_3,
+              .request_id = 0,
+              .to = shard_address,
+              .from = our_address,
+          };
+
+          SendToWorkerByUuid(source, shard_worker_message);
+
+          if (serialize_shard_splits_for_determinisim_) {
+            // This is only serialized during simulation for determinism
+            // purposes, and has been tested for correctness without the
+            // imposed determinism.
+            size_t worker_index = UuidToWorkerIndex(source);
+            workers_[worker_index].BlockOnQuiescence();
+          }
+        } else {
+          MG_ASSERT(false, "bad split source: {}", source);
+        }
+      }
+    }
   }
 };
 

src/storage/v3/shard_rsm.cpp

@@ -17,7 +17,6 @@
 #include <optional>
 #include <unordered_set>
 #include <utility>
-#include <variant>
 
 #include "common/errors.hpp"
 #include "parser/opencypher/parser.hpp"
@@ -61,6 +60,8 @@ using conversions::FromMap;
 using conversions::FromPropertyValueToValue;
 using conversions::ToMsgsVertexId;
 using conversions::ToPropertyValue;
+using io::messages::ShardManagerMessages;
+using msgs::SplitResponse;
 
 auto CreateErrorResponse(const ShardError &shard_error, const auto transaction_id, const std::string_view action) {
   msgs::ShardError message_shard_error{shard_error.code, shard_error.message};
@@ -70,6 +71,16 @@ auto CreateErrorResponse(const ShardError &shard_error, const auto transaction_i
 }
 
 msgs::WriteResponses ShardRsm::ApplyWrite(msgs::CreateVerticesRequest &&req) {
+  if (req.shard_map_version < shard_->Version()) {
+    spdlog::debug("ShardRsm Rejecting client request with stale ShardMap version, so that they retry");
+    return msgs::CreateVerticesResponse{
+        .error = msgs::ShardError{
+            .code = common::ErrorCode::STALE_SHARD_MAP,
+            .message = "Shard has a higher version than the requestor's ShardMap due to a Shard split or merge, so the "
+                       "requestor must refresh their ShardMap and try again",
+        }};
+  }
+
   auto acc = shard_->Access(req.transaction_id);
 
   std::optional<msgs::ShardError> shard_error;
@@ -90,10 +101,12 @@ msgs::WriteResponses ShardRsm::ApplyWrite(msgs::CreateVerticesRequest &&req) {
     PrimaryKey transformed_pk;
     std::transform(new_vertex.primary_key.begin(), new_vertex.primary_key.end(), std::back_inserter(transformed_pk),
                    [](msgs::Value &val) { return ToPropertyValue(std::move(val)); });
-    auto result_schema = acc.CreateVertexAndValidate(converted_label_ids, transformed_pk, converted_property_map);
 
-    if (result_schema.HasError()) {
-      shard_error.emplace(CreateErrorResponse(result_schema.GetError(), req.transaction_id, "creating vertices"));
+    auto error_opt = acc.CreateVertexAndValidate(new_vertex.idempotency_token, converted_label_ids, transformed_pk,
+                                                 converted_property_map);
+
+    if (error_opt) {
+      shard_error.emplace(CreateErrorResponse(error_opt.value(), req.transaction_id, "creating vertices"));
       break;
     }
   }
@@ -199,7 +212,8 @@ msgs::WriteResponses ShardRsm::ApplyWrite(msgs::CreateExpandRequest &&req) {
       break;
     }
 
-    auto edge_acc = acc.CreateEdge(from_vertex_id, to_vertex_id, new_expand.type.id, Gid::FromUint(new_expand.id.gid));
+    auto edge_acc = acc.CreateEdge(from_vertex_id, to_vertex_id, new_expand.type.id, Gid::FromUint(new_expand.id.gid),
+                                   new_expand.idempotency_token);
     if (edge_acc.HasValue()) {
       auto edge = edge_acc.GetValue();
       if (!new_expand.properties.empty()) {
@@ -315,7 +329,32 @@ msgs::WriteResponses ShardRsm::ApplyWrite(msgs::UpdateEdgesRequest &&req) {
   return msgs::UpdateEdgesResponse{std::move(shard_error)};
 }
 
+msgs::WriteResponses ShardRsm::ApplyWrite(msgs::SplitRequest &&req) {
+  auto converted_primary_key = conversions::ConvertPropertyVector(req.split_key);
+  auto new_shard_split_data = shard_->PerformSplit(converted_primary_key, req.old_shard_version,
+                                                   req.new_lhs_shard_version, req.new_rhs_shard_version);
+
+  if (new_shard_split_data) {
+    spdlog::debug("ShardRsm performed split from version {} to versions {} and {}", req.old_shard_version,
+                  req.new_lhs_shard_version, req.new_rhs_shard_version);
+    msgs::InitializeSplitShard msg{.shard = Shard::FromSplitData(std::move(*new_shard_split_data)),
+                                   .uuid_mapping = req.uuid_mapping};
+    shard_manager_sender_.Send(std::move(msg));
+  }
+
+  return SplitResponse{};
+}
+
 msgs::ReadResponses ShardRsm::HandleRead(msgs::ScanVerticesRequest &&req) {
+  if (req.shard_map_version < shard_->Version()) {
+    spdlog::debug("ShardRsm Rejecting client request with stale ShardMap version, so that they retry");
+    return msgs::ScanVerticesResponse{
+        .error = msgs::ShardError{
+            .code = common::ErrorCode::STALE_SHARD_MAP,
+            .message = "Shard has a higher version than the requestor's ShardMap due to a Shard split or merge, so the "
+                       "requestor must refresh their ShardMap and try again",
+        }};
+  }
   auto acc = shard_->Access(req.transaction_id);
   std::optional<msgs::ShardError> shard_error;
 
@@ -409,6 +448,9 @@ msgs::ReadResponses ShardRsm::HandleRead(msgs::ScanVerticesRequest &&req) {
     resp.results = std::move(results);
   }
 
+  spdlog::trace("Shard version {} returning {} results for ScanVertices", shard_->Version().logical_id,
+                resp.results.size());
+
   return resp;
 }
 

src/storage/v3/shard_rsm.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -12,17 +12,20 @@
 #pragma once
 
 #include <memory>
-#include <variant>
+#include <optional>
 
 #include <openssl/ec.h>
+#include "io/messages.hpp"
 #include "query/v2/requests.hpp"
 #include "storage/v3/shard.hpp"
+#include "storage/v3/value_conversions.hpp"
 #include "storage/v3/vertex_accessor.hpp"
 
 namespace memgraph::storage::v3 {
 
 class ShardRsm {
-  std::unique_ptr<Shard> shard_;
+  std::shared_ptr<Shard> shard_;
+  utils::Sender<msgs::InitializeSplitShard> shard_manager_sender_;
 
   msgs::ReadResponses HandleRead(msgs::ExpandOneRequest &&req);
   msgs::ReadResponses HandleRead(msgs::GetPropertiesRequest &&req);
@@ -36,10 +39,26 @@ class ShardRsm {
   msgs::WriteResponses ApplyWrite(msgs::DeleteEdgesRequest &&req);
   msgs::WriteResponses ApplyWrite(msgs::UpdateEdgesRequest &&req);
 
+  msgs::WriteResponses ApplyWrite(msgs::SplitRequest &&req);
+
   msgs::WriteResponses ApplyWrite(msgs::CommitRequest &&req);
 
  public:
-  explicit ShardRsm(std::unique_ptr<Shard> &&shard) : shard_(std::move(shard)){};
+  ShardRsm(std::shared_ptr<Shard> &&shard, utils::Sender<msgs::InitializeSplitShard> shard_manager_sender)
+      : shard_(std::move(shard)), shard_manager_sender_{std::move(shard_manager_sender)} {};
+
+  std::optional<msgs::SuggestedSplitInfo> ShouldSplit() const noexcept {
+    auto split_info = shard_->ShouldSplit();
+    if (split_info) {
+      return msgs::SuggestedSplitInfo{
+          .shard_to_split_uuid = split_info->shard_to_split_uuid,
+          .label_id = split_info->label_id,
+          .splitting_shard_low_key = conversions::ConvertValueVector(split_info->splitting_shard_low_key),
+          .split_key = conversions::ConvertValueVector(split_info->split_key),
+          .shard_version = split_info->shard_version};
+    }
+    return std::nullopt;
+  }
 
   // NOLINTNEXTLINE(readability-convert-member-functions-to-static)
   msgs::ReadResponses Read(msgs::ReadRequests requests) {

src/storage/v3/shard_worker.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -17,9 +17,11 @@
 #include <queue>
 #include <variant>
 
+#include <boost/core/demangle.hpp>
 #include <boost/uuid/uuid.hpp>
 
 #include "coordinator/coordinator.hpp"
+#include "coordinator/hybrid_logical_clock.hpp"
 #include "coordinator/shard_map.hpp"
 #include "io/address.hpp"
 #include "io/future.hpp"
@@ -46,6 +48,7 @@ using io::RequestId;
 using io::Time;
 using io::messages::ShardMessages;
 using io::rsm::Raft;
+using msgs::InitializeSplitShardByUUID;
 using msgs::ReadRequests;
 using msgs::ReadResponses;
 using msgs::WriteRequests;
@@ -66,7 +69,7 @@ struct RouteMessage {
   Address from;
 };
 
-using Message = std::variant<ShutDown, Cron, ShardToInitialize, RouteMessage>;
+using Message = std::variant<ShutDown, Cron, ShardToInitialize, RouteMessage, InitializeSplitShardByUUID>;
 
 struct QueueInner {
   std::mutex mu{};
@@ -155,9 +158,54 @@ class ShardWorker {
   }
 
   bool Process(RouteMessage &&route_message) {
-    auto &rsm = rsm_map_.at(route_message.to.unique_id);
+    if (rsm_map_.contains(route_message.to.unique_id)) {
+      spdlog::info("ShardWorker routing message to rsm {}", route_message.to.unique_id);
+      auto &rsm = rsm_map_.at(route_message.to.unique_id);
 
-    rsm.Handle(std::move(route_message.message), route_message.request_id, route_message.from);
+      rsm.Handle(std::move(route_message.message), route_message.request_id, route_message.from);
+    } else {
+      auto type_info = std::visit([&](const auto &msg) { return io::TypeInfoFor(msg); }, route_message.message);
+
+      std::string demangled_name = boost::core::demangle(type_info.get().name());
+      spdlog::debug(
+          "ShardWorker received {} message for rsm {} which does not exist on our system (possibly due to a split not "
+          "having been applied locally yet)",
+          demangled_name, route_message.to.unique_id);
+    }
+
+    return true;
+  }
+
+  bool Process(InitializeSplitShardByUUID &&initialize_split_shard) {
+    if (rsm_map_.contains(initialize_split_shard.shard_uuid)) {
+      // it's not a bug for the coordinator to send us UUIDs that we have
+      // already created, because there may have been lag that caused
+      // the coordinator not to hear back from us.
+      return true;
+    }
+
+    auto rsm_io = io_.ForkLocal(initialize_split_shard.shard_uuid);
+
+    // TODO(tyler) get peers from Coordinator in HeartbeatResponse
+    std::vector<Address> rsm_peers = {};
+
+    Address local_shard_manager_address = io_.GetAddress().ForkLocalShardManager();
+    utils::Sender<msgs::InitializeSplitShard> local_shard_manager_sender =
+        io_.template GetSender<msgs::InitializeSplitShard>(local_shard_manager_address);
+
+    // TODO(tyler) pass this local_shard_manager_sender to the Shard so that it can communicate back to the local
+    // manager from split code
+    ShardRsm rsm_state{std::move(initialize_split_shard.shard), std::move(local_shard_manager_sender)};
+
+    ShardRaft<IoImpl> rsm{std::move(rsm_io), rsm_peers, std::move(rsm_state)};
+
+    spdlog::info("ShardWorker created a new shard with UUID {}", initialize_split_shard.shard_uuid);
+
+    // perform an initial Cron call for the new RSM
+    Time next_cron = rsm.Cron();
+    cron_schedule_.push(std::make_pair(next_cron, initialize_split_shard.shard_uuid));
+
+    rsm_map_.emplace(initialize_split_shard.shard_uuid, std::move(rsm));
 
     return true;
   }
@@ -173,6 +221,13 @@ class ShardWorker {
         auto &rsm = rsm_map_.at(uuid);
         Time next_for_uuid = rsm.Cron();
 
+        // Check if shard should split
+        if (auto split_info = rsm.ShouldSplit(); split_info) {
+          split_info->shard_to_split_uuid = uuid;
+          const auto shard_manager_addr = io_.GetAddress().ForkLocalShardManager();
+          io_.Send(shard_manager_addr, 0, *split_info);
+        }
+
         cron_schedule_.pop();
         cron_schedule_.push(std::make_pair(next_for_uuid, uuid));
       } else {
@@ -188,6 +243,7 @@ class ShardWorker {
       // it's not a bug for the coordinator to send us UUIDs that we have
       // already created, because there may have been lag that caused
       // the coordinator not to hear back from us.
+      spdlog::debug("ShardWorker told to initialize already-existing shard with UUID {} - skipping", to_init.uuid);
       return;
     }
 
@@ -196,14 +252,22 @@ class ShardWorker {
     // TODO(tyler) get peers from Coordinator in HeartbeatResponse
     std::vector<Address> rsm_peers = {};
 
-    std::unique_ptr<Shard> shard = std::make_unique<Shard>(to_init.label_id, to_init.min_key, to_init.max_key,
-                                                           to_init.schema, to_init.config, to_init.id_to_names);
+    Address local_shard_manager_address = io_.GetAddress().ForkLocalShardManager();
+    utils::Sender<msgs::InitializeSplitShard> local_shard_manager_sender =
+        io_.template GetSender<msgs::InitializeSplitShard>(local_shard_manager_address);
 
-    ShardRsm rsm_state{std::move(shard)};
+    // TODO(tyler) pass this local_shard_manager_sender to the Shard so that it can communicate back to the local
+    // manager from split code
+
+    std::unique_ptr<Shard> shard =
+        std::make_unique<Shard>(to_init.label_id, to_init.min_key, to_init.max_key, to_init.schema,
+                                to_init.new_shard_version, to_init.config, to_init.id_to_names);
+
+    ShardRsm rsm_state{std::move(shard), std::move(local_shard_manager_sender)};
 
     ShardRaft<IoImpl> rsm{std::move(rsm_io), rsm_peers, std::move(rsm_state)};
 
-    spdlog::info("SM created a new shard with UUID {}", to_init.uuid);
+    spdlog::debug("ShardWorker created a new shard with UUID {}", to_init.uuid);
 
     // perform an initial Cron call for the new RSM
     Time next_cron = rsm.Cron();

src/storage/v3/splitter.cpp

@@ -0,0 +1,351 @@
+// Copyright 2023 Memgraph Ltd.
+//
+// Use of this software is governed by the Business Source License
+// included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
+// License, and you may not use this file except in compliance with the Business Source License.
+//
+// As of the Change Date specified in that file, in accordance with
+// the Business Source License, use of this software will be governed
+// by the Apache License, Version 2.0, included in the file
+// licenses/APL.txt.
+
+#include "storage/v3/splitter.hpp"
+
+#include <algorithm>
+#include <cstdint>
+#include <map>
+#include <memory>
+#include <optional>
+#include <set>
+
+#include "storage/v3/config.hpp"
+#include "storage/v3/delta.hpp"
+#include "storage/v3/id_types.hpp"
+#include "storage/v3/indices.hpp"
+#include "storage/v3/key_store.hpp"
+#include "storage/v3/name_id_mapper.hpp"
+#include "storage/v3/schemas.hpp"
+#include "storage/v3/shard.hpp"
+#include "storage/v3/transaction.hpp"
+#include "storage/v3/vertex.hpp"
+#include "utils/logging.hpp"
+
+namespace memgraph::storage::v3 {
+
+Splitter::Splitter(const LabelId primary_label, VertexContainer &vertices, EdgeContainer &edges,
+                   std::map<uint64_t, std::unique_ptr<Transaction>> &start_logical_id_to_transaction, Indices &indices,
+                   const Config &config, const std::vector<SchemaProperty> &schema, const NameIdMapper &name_id_mapper)
+    : primary_label_(primary_label),
+      vertices_(vertices),
+      edges_(edges),
+      start_logical_id_to_transaction_(start_logical_id_to_transaction),
+      indices_(indices),
+      config_(config),
+      schema_(schema),
+      name_id_mapper_(name_id_mapper) {}
+
+SplitData Splitter::SplitShard(const PrimaryKey &split_key, const std::optional<PrimaryKey> &max_primary_key,
+                               const Hlc shard_version) {
+  SplitData data{.primary_label = primary_label_,
+                 .min_primary_key = split_key,
+                 .max_primary_key = max_primary_key,
+                 .schema = schema_,
+                 .config = config_,
+                 .id_to_name = name_id_mapper_.GetIdToNameMap(),
+                 .shard_version = shard_version};
+
+  std::set<uint64_t> collected_transactions_;
+  data.vertices = CollectVertices(data, collected_transactions_, split_key);
+  data.edges = CollectEdges(collected_transactions_, data.vertices, split_key);
+  data.transactions = CollectTransactions(collected_transactions_, data.vertices, *data.edges, split_key);
+
+  return data;
+}
+
+void Splitter::ScanDeltas(std::set<uint64_t> &collected_transactions_, Delta *delta) {
+  while (delta != nullptr) {
+    collected_transactions_.insert(delta->commit_info->start_or_commit_timestamp.logical_id);
+    delta = delta->next;
+  }
+}
+
+VertexContainer Splitter::CollectVertices(SplitData &data, std::set<uint64_t> &collected_transactions_,
+                                          const PrimaryKey &split_key) {
+  data.label_indices = indices_.label_index.SplitIndexEntries(split_key);
+  data.label_property_indices = indices_.label_property_index.SplitIndexEntries(split_key);
+
+  VertexContainer splitted_data;
+  auto split_key_it = vertices_.find(split_key);
+  while (split_key_it != vertices_.end()) {
+    // Go through deltas and pick up transactions start_id/commit_id
+    ScanDeltas(collected_transactions_, split_key_it->second.delta);
+
+    auto next_it = std::next(split_key_it);
+
+    const auto &[splitted_vertex_it, inserted, node] = splitted_data.insert(vertices_.extract(split_key_it->first));
+    MG_ASSERT(inserted, "Failed to extract vertex!");
+
+    split_key_it = next_it;
+  }
+  return splitted_data;
+}
+
+std::optional<EdgeContainer> Splitter::CollectEdges(std::set<uint64_t> &collected_transactions_,
+                                                    const VertexContainer &split_vertices,
+                                                    const PrimaryKey &split_key) {
+  if (!config_.items.properties_on_edges) {
+    return std::nullopt;
+  }
+  EdgeContainer splitted_edges;
+  const auto split_vertex_edges = [&](const auto &edges_ref) {
+    // This is safe since if properties_on_edges is true, the this must be a ptr
+    for (const auto &edge_ref : edges_ref) {
+      auto *edge = std::get<2>(edge_ref).ptr;
+      const auto &other_vtx = std::get<1>(edge_ref);
+      ScanDeltas(collected_transactions_, edge->delta);
+      // Check if src and dest edge are both on splitted shard so we know if we
+      // should remove orphan edge, or make a clone
+      if (other_vtx.primary_key >= split_key) {
+        // Remove edge from shard
+        splitted_edges.insert(edges_.extract(edge->gid));
+      } else {
+        splitted_edges.insert({edge->gid, Edge{edge->gid, edge->delta}});
+      }
+    }
+  };
+
+  for (const auto &vertex : split_vertices) {
+    split_vertex_edges(vertex.second.in_edges);
+    split_vertex_edges(vertex.second.out_edges);
+  }
+  return splitted_edges;
+}
+
+std::map<uint64_t, std::unique_ptr<Transaction>> Splitter::CollectTransactions(
+    const std::set<uint64_t> &collected_transactions_, VertexContainer &cloned_vertices, EdgeContainer &cloned_edges,
+    const PrimaryKey &split_key) {
+  std::map<uint64_t, std::unique_ptr<Transaction>> transactions;
+
+  for (const auto &[commit_start, transaction] : start_logical_id_to_transaction_) {
+    // We need all transaction whose deltas need to be resolved for any of the
+    // entities
+    if (collected_transactions_.contains(transaction->commit_info->start_or_commit_timestamp.logical_id)) {
+      transactions.insert({commit_start, start_logical_id_to_transaction_[commit_start]->Clone()});
+    }
+  }
+
+  // It is necessary to clone all the transactions first so we have new addresses
+  // for deltas, before doing alignment of deltas and prev_ptr
+  AdjustClonedTransactions(transactions, cloned_vertices, cloned_edges, split_key);
+  return transactions;
+}
+
+void PruneDeltas(Transaction &cloned_transaction, std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions,
+                 const PrimaryKey &split_key) {
+  // Remove delta chains that don't point to objects on splitted shard
+  auto cloned_delta_it = cloned_transaction.deltas.begin();
+
+  while (cloned_delta_it != cloned_transaction.deltas.end()) {
+    const auto prev = cloned_delta_it->prev.Get();
+    switch (prev.type) {
+      case PreviousPtr::Type::DELTA:
+      case PreviousPtr::Type::NULLPTR:
+        ++cloned_delta_it;
+        break;
+      case PreviousPtr::Type::VERTEX: {
+        if (prev.vertex->first < split_key) {
+          // We can remove this delta chain
+          auto *current_next_delta = cloned_delta_it->next;
+          cloned_delta_it = cloned_transaction.deltas.erase(cloned_delta_it);
+
+          while (current_next_delta != nullptr) {
+            auto *next_delta = current_next_delta->next;
+            // Find next delta transaction delta list
+            auto current_transaction_it = std::ranges::find_if(
+                cloned_transactions,
+                [&start_or_commit_timestamp =
+                     current_next_delta->commit_info->start_or_commit_timestamp](const auto &transaction) {
+                  return transaction.second->start_timestamp == start_or_commit_timestamp ||
+                         transaction.second->commit_info->start_or_commit_timestamp == start_or_commit_timestamp;
+                });
+            MG_ASSERT(current_transaction_it != cloned_transactions.end(), "Error when pruning deltas!");
+            // Remove it
+            current_transaction_it->second->deltas.remove_if(
+                [&current_next_delta = *current_next_delta](const auto &delta) { return delta == current_next_delta; });
+
+            current_next_delta = next_delta;
+          }
+        } else {
+          ++cloned_delta_it;
+        }
+        break;
+      }
+      case PreviousPtr::Type::EDGE:
+        ++cloned_delta_it;
+        break;
+    }
+  }
+}
+
+void Splitter::AdjustClonedTransactions(std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions,
+                                        VertexContainer &cloned_vertices, EdgeContainer &cloned_edges,
+                                        const PrimaryKey &split_key) {
+  for (auto &[commit_start, cloned_transaction] : cloned_transactions) {
+    AdjustClonedTransaction(*cloned_transaction, *start_logical_id_to_transaction_[commit_start], cloned_transactions,
+                            cloned_vertices, cloned_edges, split_key);
+  }
+  // Prune deltas whose delta chain points to vertex/edge that should not belong on that shard
+  // Prune must be after ajdust, since next, and prev are not set and we cannot follow the chain
+  for (auto &[commit_start, cloned_transaction] : cloned_transactions) {
+    PruneDeltas(*cloned_transaction, cloned_transactions, split_key);
+  }
+}
+
+inline bool IsDeltaHeadOfChain(const PreviousPtr::Type &delta_type) {
+  return delta_type == PreviousPtr::Type::VERTEX || delta_type == PreviousPtr::Type::EDGE;
+}
+
+bool DoesPrevPtrPointsToSplittedData(const PreviousPtr::Pointer &prev_ptr, const PrimaryKey &split_key) {
+  return prev_ptr.type == PreviousPtr::Type::VERTEX && prev_ptr.vertex->first < split_key;
+}
+
+void Splitter::AdjustClonedTransaction(Transaction &cloned_transaction, const Transaction &transaction,
+                                       std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions,
+                                       VertexContainer &cloned_vertices, EdgeContainer &cloned_edges,
+                                       const PrimaryKey & /*split_key*/) {
+  auto delta_it = transaction.deltas.begin();
+  auto cloned_delta_it = cloned_transaction.deltas.begin();
+
+  while (delta_it != transaction.deltas.end()) {
+    // We can safely ignore deltas which are not head of delta chain
+    // Dont' adjust delta chain that points to irrelevant data vertices/edges
+    if (const auto delta_prev = delta_it->prev.Get(); !IsDeltaHeadOfChain(delta_prev.type)) {
+      ++delta_it;
+      ++cloned_delta_it;
+      continue;
+    }
+
+    const auto *delta = &*delta_it;
+    auto *cloned_delta = &*cloned_delta_it;
+    Delta *cloned_delta_prev_ptr = cloned_delta;
+    while (delta->next != nullptr) {
+      AdjustEdgeRef(*cloned_delta, cloned_edges);
+
+      // Align next ptr
+      AdjustDeltaNext(*delta, *cloned_delta, cloned_transactions);
+
+      // Align prev ptr
+      if (cloned_delta_prev_ptr != nullptr) {
+        AdjustDeltaPrevPtr(*delta, *cloned_delta_prev_ptr, cloned_transactions, cloned_vertices, cloned_edges);
+      }
+
+      // TODO Next delta might not belong to the cloned transaction and thats
+      // why we skip this delta of the delta chain
+      if (cloned_delta->next != nullptr) {
+        cloned_delta = cloned_delta->next;
+        cloned_delta_prev_ptr = cloned_delta;
+      } else {
+        cloned_delta_prev_ptr = nullptr;
+      }
+      delta = delta->next;
+    }
+    // Align prev ptr
+    if (cloned_delta_prev_ptr != nullptr) {
+      AdjustDeltaPrevPtr(*delta, *cloned_delta_prev_ptr, cloned_transactions, cloned_vertices, cloned_edges);
+    }
+
+    ++delta_it;
+    ++cloned_delta_it;
+  }
+  MG_ASSERT(delta_it == transaction.deltas.end() && cloned_delta_it == cloned_transaction.deltas.end(),
+            "Both iterators must be exhausted!");
+}
+
+void Splitter::AdjustEdgeRef(Delta &cloned_delta, EdgeContainer &cloned_edges) const {
+  switch (cloned_delta.action) {
+    case Delta::Action::ADD_IN_EDGE:
+    case Delta::Action::ADD_OUT_EDGE:
+    case Delta::Action::REMOVE_IN_EDGE:
+    case Delta::Action::REMOVE_OUT_EDGE: {
+      // Find edge
+      if (config_.items.properties_on_edges) {
+        // Only case when not finding is when the edge is not on splitted shard
+        // TODO Do this after prune an move condition into assert
+        if (const auto cloned_edge_it =
+                std::ranges::find_if(cloned_edges, [edge_ptr = cloned_delta.vertex_edge.edge.ptr](
+                                                       const auto &elem) { return elem.second.gid == edge_ptr->gid; });
+            cloned_edge_it != cloned_edges.end()) {
+          cloned_delta.vertex_edge.edge = EdgeRef{&cloned_edge_it->second};
+        }
+      }
+      break;
+    }
+    case Delta::Action::DELETE_OBJECT:
+    case Delta::Action::RECREATE_OBJECT:
+    case Delta::Action::SET_PROPERTY:
+    case Delta::Action::ADD_LABEL:
+    case Delta::Action::REMOVE_LABEL: {
+      // noop
+      break;
+    }
+  }
+}
+
+void Splitter::AdjustDeltaNext(const Delta &original, Delta &cloned,
+                               std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions) {
+  // Get cloned_delta->next transaction, using delta->next original transaction
+  auto cloned_transaction_it = std::ranges::find_if(cloned_transactions, [&original](const auto &elem) {
+    return elem.second->start_timestamp == original.next->commit_info->start_or_commit_timestamp ||
+           elem.second->commit_info->start_or_commit_timestamp == original.next->commit_info->start_or_commit_timestamp;
+  });
+  // TODO(jbajic) What if next in delta chain does not belong to cloned transaction?
+  // MG_ASSERT(cloned_transaction_it != cloned_transactions.end(), "Cloned transaction not found");
+  if (cloned_transaction_it == cloned_transactions.end()) return;
+  // Find cloned delta in delta list of cloned transaction
+  auto found_cloned_delta_it = std::ranges::find_if(
+      cloned_transaction_it->second->deltas, [&original](const auto &elem) { return elem.id == original.next->id; });
+  MG_ASSERT(found_cloned_delta_it != cloned_transaction_it->second->deltas.end(), "Delta with given uuid must exist!");
+  cloned.next = &*found_cloned_delta_it;
+}
+
+void Splitter::AdjustDeltaPrevPtr(const Delta &original, Delta &cloned,
+                                  std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions,
+                                  VertexContainer & /*cloned_vertices*/, EdgeContainer &cloned_edges) {
+  auto ptr = original.prev.Get();
+  switch (ptr.type) {
+    case PreviousPtr::Type::NULLPTR: {
+      // noop
+      break;
+    }
+    case PreviousPtr::Type::DELTA: {
+      // Same as for deltas except don't align next but prev
+      auto cloned_transaction_it = std::ranges::find_if(cloned_transactions, [&ptr](const auto &elem) {
+        return elem.second->start_timestamp == ptr.delta->commit_info->start_or_commit_timestamp ||
+               elem.second->commit_info->start_or_commit_timestamp == ptr.delta->commit_info->start_or_commit_timestamp;
+      });
+      MG_ASSERT(cloned_transaction_it != cloned_transactions.end(), "Cloned transaction not found");
+      // Find cloned delta in delta list of cloned transaction
+      auto found_cloned_delta_it =
+          std::ranges::find_if(cloned_transaction_it->second->deltas,
+                               [delta = ptr.delta](const auto &elem) { return elem.id == delta->id; });
+      MG_ASSERT(found_cloned_delta_it != cloned_transaction_it->second->deltas.end(),
+                "Delta with given id must exist!");
+
+      cloned.prev.Set(&*found_cloned_delta_it);
+      break;
+    }
+    case PreviousPtr::Type::VERTEX: {
+      // The vertex was extracted and it is safe to reuse address
+      cloned.prev.Set(ptr.vertex);
+      break;
+    }
+    case PreviousPtr::Type::EDGE: {
+      // We can never be here if we have properties on edge disabled
+      auto *cloned_edge = &*cloned_edges.find(ptr.edge->gid);
+      cloned.prev.Set(&cloned_edge->second);
+      break;
+    }
+  };
+}
+
+}  // namespace memgraph::storage::v3

src/storage/v3/splitter.hpp

@@ -0,0 +1,108 @@
+// Copyright 2023 Memgraph Ltd.
+//
+// Use of this software is governed by the Business Source License
+// included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
+// License, and you may not use this file except in compliance with the Business Source License.
+//
+// As of the Change Date specified in that file, in accordance with
+// the Business Source License, use of this software will be governed
+// by the Apache License, Version 2.0, included in the file
+// licenses/APL.txt.
+#pragma once
+
+#include <cstdint>
+#include <map>
+#include <memory>
+#include <optional>
+#include <set>
+
+#include "coordinator/hybrid_logical_clock.hpp"
+#include "storage/v3/config.hpp"
+#include "storage/v3/delta.hpp"
+#include "storage/v3/edge.hpp"
+#include "storage/v3/id_types.hpp"
+#include "storage/v3/indices.hpp"
+#include "storage/v3/key_store.hpp"
+#include "storage/v3/name_id_mapper.hpp"
+#include "storage/v3/schemas.hpp"
+#include "storage/v3/transaction.hpp"
+#include "storage/v3/vertex.hpp"
+#include "utils/concepts.hpp"
+
+namespace memgraph::storage::v3 {
+
+// If edge properties-on-edges is false then we don't need to send edges but
+// only vertices, since they will contain those edges
+struct SplitData {
+  LabelId primary_label;
+  PrimaryKey min_primary_key;
+  std::optional<PrimaryKey> max_primary_key;
+  std::vector<SchemaProperty> schema;
+  Config config;
+  std::unordered_map<uint64_t, std::string> id_to_name;
+  coordinator::Hlc shard_version;
+
+  VertexContainer vertices;
+  std::optional<EdgeContainer> edges;
+  std::map<uint64_t, std::unique_ptr<Transaction>> transactions;
+  std::map<LabelId, LabelIndex::IndexContainer> label_indices;
+  std::map<std::pair<LabelId, PropertyId>, LabelPropertyIndex::IndexContainer> label_property_indices;
+};
+
+class Splitter final {
+ public:
+  Splitter(LabelId primary_label, VertexContainer &vertices, EdgeContainer &edges,
+           std::map<uint64_t, std::unique_ptr<Transaction>> &start_logical_id_to_transaction, Indices &indices,
+           const Config &config, const std::vector<SchemaProperty> &schema, const NameIdMapper &name_id_mapper_);
+
+  Splitter(const Splitter &) = delete;
+  Splitter(Splitter &&) noexcept = delete;
+  Splitter &operator=(const Splitter &) = delete;
+  Splitter operator=(Splitter &&) noexcept = delete;
+  ~Splitter() = default;
+
+  SplitData SplitShard(const PrimaryKey &split_key, const std::optional<PrimaryKey> &max_primary_key,
+                       coordinator::Hlc shard_version);
+
+ private:
+  VertexContainer CollectVertices(SplitData &data, std::set<uint64_t> &collected_transactions_start_id,
+                                  const PrimaryKey &split_key);
+
+  std::optional<EdgeContainer> CollectEdges(std::set<uint64_t> &collected_transactions_start_id,
+                                            const VertexContainer &split_vertices, const PrimaryKey &split_key);
+
+  std::map<uint64_t, std::unique_ptr<Transaction>> CollectTransactions(
+      const std::set<uint64_t> &collected_transactions_start_id, VertexContainer &cloned_vertices,
+      EdgeContainer &cloned_edges, const PrimaryKey &split_key);
+
+  static void ScanDeltas(std::set<uint64_t> &collected_transactions_start_id, Delta *delta);
+
+  void AdjustClonedTransaction(Transaction &cloned_transaction, const Transaction &transaction,
+                               std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions,
+                               VertexContainer &cloned_vertices, EdgeContainer &cloned_edges,
+                               const PrimaryKey &split_key);
+
+  void AdjustClonedTransactions(std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions,
+                                VertexContainer &cloned_vertices, EdgeContainer &cloned_edges,
+                                const PrimaryKey &split_key);
+
+  void AdjustEdgeRef(Delta &cloned_delta, EdgeContainer &cloned_edges) const;
+
+  static void AdjustDeltaNext(const Delta &original, Delta &cloned,
+                              std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions);
+
+  static void AdjustDeltaPrevPtr(const Delta &original, Delta &cloned,
+                                 std::map<uint64_t, std::unique_ptr<Transaction>> &cloned_transactions,
+                                 VertexContainer &cloned_vertices, EdgeContainer &cloned_edges);
+
+  const LabelId primary_label_;
+  VertexContainer &vertices_;
+  EdgeContainer &edges_;
+  std::map<uint64_t, std::unique_ptr<Transaction>> &start_logical_id_to_transaction_;
+  Indices &indices_;
+  const Config &config_;
+  const std::vector<SchemaProperty> schema_;
+  const NameIdMapper &name_id_mapper_;
+};
+
+}  // namespace memgraph::storage::v3

src/storage/v3/transaction.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -31,6 +31,16 @@ struct CommitInfo {
 };
 
 struct Transaction {
+  Transaction(coordinator::Hlc start_timestamp, CommitInfo new_commit_info, std::list<Delta> deltas,
+              uint64_t command_id, bool must_abort, bool is_aborted, IsolationLevel isolation_level)
+      : start_timestamp{start_timestamp},
+        commit_info{std::make_unique<CommitInfo>(new_commit_info)},
+        command_id(command_id),
+        deltas(std::move(deltas)),
+        must_abort(must_abort),
+        is_aborted(is_aborted),
+        isolation_level(isolation_level){};
+
   Transaction(coordinator::Hlc start_timestamp, IsolationLevel isolation_level)
       : start_timestamp(start_timestamp),
         commit_info(std::make_unique<CommitInfo>(CommitInfo{false, {start_timestamp}})),
@@ -54,6 +64,55 @@ struct Transaction {
 
   ~Transaction() {}
 
+  std::list<Delta> CopyDeltas(CommitInfo *commit_info) const {
+    std::list<Delta> copied_deltas;
+    for (const auto &delta : deltas) {
+      switch (delta.action) {
+        case Delta::Action::DELETE_OBJECT:
+          copied_deltas.emplace_back(Delta::DeleteObjectTag{}, commit_info, delta.id, command_id,
+                                     delta.idempotency_token);
+          break;
+        case Delta::Action::RECREATE_OBJECT:
+          copied_deltas.emplace_back(Delta::RecreateObjectTag{}, commit_info, delta.id, command_id);
+          break;
+        case Delta::Action::ADD_LABEL:
+          copied_deltas.emplace_back(Delta::AddLabelTag{}, delta.label, commit_info, delta.id, command_id);
+          break;
+        case Delta::Action::REMOVE_LABEL:
+          copied_deltas.emplace_back(Delta::RemoveLabelTag{}, delta.label, commit_info, delta.id, command_id);
+          break;
+        case Delta::Action::ADD_IN_EDGE:
+          copied_deltas.emplace_back(Delta::AddInEdgeTag{}, delta.vertex_edge.edge_type, delta.vertex_edge.vertex_id,
+                                     delta.vertex_edge.edge, commit_info, delta.id, command_id);
+          break;
+        case Delta::Action::ADD_OUT_EDGE:
+          copied_deltas.emplace_back(Delta::AddOutEdgeTag{}, delta.vertex_edge.edge_type, delta.vertex_edge.vertex_id,
+                                     delta.vertex_edge.edge, commit_info, delta.id, command_id);
+          break;
+        case Delta::Action::REMOVE_IN_EDGE:
+          copied_deltas.emplace_back(Delta::RemoveInEdgeTag{}, delta.vertex_edge.edge_type, delta.vertex_edge.vertex_id,
+                                     delta.vertex_edge.edge, commit_info, delta.id, command_id);
+          break;
+        case Delta::Action::REMOVE_OUT_EDGE:
+          copied_deltas.emplace_back(Delta::RemoveOutEdgeTag{}, delta.vertex_edge.edge_type,
+                                     delta.vertex_edge.vertex_id, delta.vertex_edge.edge, commit_info, delta.id,
+                                     command_id);
+          break;
+        case Delta::Action::SET_PROPERTY:
+          copied_deltas.emplace_back(Delta::SetPropertyTag{}, delta.property.key, delta.property.value, commit_info,
+                                     delta.id, command_id);
+          break;
+      }
+    }
+    return copied_deltas;
+  }
+
+  // This does not solve the whole problem of copying deltas
+  std::unique_ptr<Transaction> Clone() const {
+    return std::make_unique<Transaction>(start_timestamp, *commit_info, CopyDeltas(commit_info.get()), command_id,
+                                         must_abort, is_aborted, isolation_level);
+  }
+
   coordinator::Hlc start_timestamp;
   std::unique_ptr<CommitInfo> commit_info;
   uint64_t command_id;

src/utils/concrete_msg_sender.hpp

@@ -0,0 +1,42 @@
+// Copyright 2023 Memgraph Ltd.
+//
+// Use of this software is governed by the Business Source License
+// included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
+// License, and you may not use this file except in compliance with the Business Source License.
+//
+// As of the Change Date specified in that file, in accordance with
+// the Business Source License, use of this software will be governed
+// by the Apache License, Version 2.0, included in the file
+// licenses/APL.txt.
+
+#pragma once
+
+#include <concepts>
+#include <functional>
+#include <type_traits>
+
+namespace memgraph::utils {
+
+// TODO(tyler) ensure that Message continues to represent
+// reasonable constraints around message types over time,
+// as we adapt things to use Thrift-generated message types.
+template <typename T>
+concept Message = std::movable<T> && std::copyable<T>;
+
+/// This is a concrete type that allows one message type to be
+/// sent to a single address. Initially intended to be used by the
+/// Shard to send messages to the local ShardManager.
+template <Message M>
+class Sender {
+ public:
+  Sender() = default;
+
+  explicit Sender(std::function<void(M)> sender) : sender_(sender) {}
+
+  void Send(M &&message) { sender_(std::forward<M>(message)); }
+
+ private:
+  std::function<void(M)> sender_;
+};
+
+}  // namespace memgraph::utils

tests/benchmark/CMakeLists.txt

@@ -79,3 +79,6 @@ target_link_libraries(${test_prefix}data_structures_contains mg-utils mg-storage
 
 add_benchmark(data_structures_remove.cpp)
 target_link_libraries(${test_prefix}data_structures_remove mg-utils mg-storage-v3)
+
+add_benchmark(storage_v3_split.cpp)
+target_link_libraries(${test_prefix}storage_v3_split mg-storage-v3 mg-query-v2)

tests/benchmark/storage_v3_split.cpp

@@ -0,0 +1,194 @@
+// Copyright 2023 Memgraph Ltd.
+//
+// Use of this software is governed by the Business Source License
+// included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
+// License, and you may not use this file except in compliance with the Business Source License.
+//
+// As of the Change Date specified in that file, in accordance with
+// the Business Source License, use of this software will be governed
+// by the Apache License, Version 2.0, included in the file
+// licenses/APL.txt.
+
+#include <cstdint>
+#include <optional>
+#include <vector>
+
+#include <benchmark/benchmark.h>
+#include <gflags/gflags.h>
+
+#include "storage/v3/id_types.hpp"
+#include "storage/v3/key_store.hpp"
+#include "storage/v3/property_value.hpp"
+#include "storage/v3/shard.hpp"
+#include "storage/v3/vertex.hpp"
+#include "storage/v3/vertex_id.hpp"
+
+namespace memgraph::benchmark {
+
+class ShardSplitBenchmark : public ::benchmark::Fixture {
+ protected:
+  using PrimaryKey = storage::v3::PrimaryKey;
+  using PropertyId = storage::v3::PropertyId;
+  using PropertyValue = storage::v3::PropertyValue;
+  using LabelId = storage::v3::LabelId;
+  using EdgeTypeId = storage::v3::EdgeTypeId;
+  using Shard = storage::v3::Shard;
+  using VertexId = storage::v3::VertexId;
+  using Gid = storage::v3::Gid;
+
+  void SetUp(const ::benchmark::State &state) override {
+    storage.emplace(primary_label, min_pk, std::nullopt, schema_property_vector, last_hlc);
+    storage->StoreMapping(
+        {{1, "label"}, {2, "property"}, {3, "edge_property"}, {4, "secondary_label"}, {5, "secondary_prop"}});
+  }
+
+  void TearDown(const ::benchmark::State &) override { storage = std::nullopt; }
+
+  const PropertyId primary_property{PropertyId::FromUint(2)};
+  const PropertyId secondary_property{PropertyId::FromUint(5)};
+  std::vector<storage::v3::SchemaProperty> schema_property_vector = {
+      storage::v3::SchemaProperty{primary_property, common::SchemaType::INT}};
+  const std::vector<PropertyValue> min_pk{PropertyValue{0}};
+  const LabelId primary_label{LabelId::FromUint(1)};
+  const LabelId secondary_label{LabelId::FromUint(4)};
+  const EdgeTypeId edge_type_id{EdgeTypeId::FromUint(3)};
+  std::optional<Shard> storage;
+
+  coordinator::Hlc last_hlc{0, io::Time{}};
+
+  coordinator::Hlc GetNextHlc() {
+    ++last_hlc.logical_id;
+    last_hlc.coordinator_wall_clock += std::chrono::seconds(1);
+    return last_hlc;
+  }
+};
+
+BENCHMARK_DEFINE_F(ShardSplitBenchmark, BigDataSplit)(::benchmark::State &state) {
+  const auto number_of_vertices{state.range(0)};
+  std::random_device r;
+  std::default_random_engine e1(r());
+  std::uniform_int_distribution<int> uniform_dist(0, number_of_vertices);
+
+  for (int64_t i{0}; i < number_of_vertices; ++i) {
+    auto acc = storage->Access(GetNextHlc());
+    MG_ASSERT(acc.CreateVertexAndValidate({secondary_label}, PrimaryKey{PropertyValue(i)},
+                                          {{secondary_property, PropertyValue(i)}})
+                  .HasValue(),
+              "Failed creating with pk {}", i);
+    if (i > 1) {
+      const auto vtx1 = uniform_dist(e1) % i;
+      const auto vtx2 = uniform_dist(e1) % i;
+
+      MG_ASSERT(acc.CreateEdge(VertexId{primary_label, {PropertyValue(vtx1)}},
+                               VertexId{primary_label, {PropertyValue(vtx2)}}, edge_type_id, Gid::FromUint(i))
+                    .HasValue(),
+                "Failed on {} and {}", vtx1, vtx2);
+    }
+    acc.Commit(GetNextHlc());
+  }
+  for (auto _ : state) {
+    auto data = storage->PerformSplit(PrimaryKey{PropertyValue{number_of_vertices / 2}}, last_hlc, GetNextHlc());
+  }
+}
+
+BENCHMARK_DEFINE_F(ShardSplitBenchmark, BigDataSplitWithGc)(::benchmark::State &state) {
+  const auto number_of_vertices{state.range(0)};
+  std::random_device r;
+  std::default_random_engine e1(r());
+  std::uniform_int_distribution<int> uniform_dist(0, number_of_vertices);
+
+  for (int64_t i{0}; i < number_of_vertices; ++i) {
+    auto acc = storage->Access(GetNextHlc());
+    MG_ASSERT(acc.CreateVertexAndValidate({secondary_label}, PrimaryKey{PropertyValue(i)},
+                                          {{secondary_property, PropertyValue(i)}})
+                  .HasValue(),
+              "Failed creating with pk {}", i);
+    if (i > 1) {
+      const auto vtx1 = uniform_dist(e1) % i;
+      const auto vtx2 = uniform_dist(e1) % i;
+
+      MG_ASSERT(acc.CreateEdge(VertexId{primary_label, {PropertyValue(vtx1)}},
+                               VertexId{primary_label, {PropertyValue(vtx2)}}, edge_type_id, Gid::FromUint(i))
+                    .HasValue(),
+                "Failed on {} and {}", vtx1, vtx2);
+    }
+    acc.Commit(GetNextHlc());
+  }
+  storage->CollectGarbage(GetNextHlc().coordinator_wall_clock);
+  for (auto _ : state) {
+    auto data = storage->PerformSplit(PrimaryKey{PropertyValue{number_of_vertices / 2}}, last_hlc, GetNextHlc());
+  }
+}
+
+BENCHMARK_DEFINE_F(ShardSplitBenchmark, BigDataSplitWithFewTransactions)(::benchmark::State &state) {
+  const auto number_of_vertices = state.range(0);
+  const auto number_of_edges = state.range(1);
+  const auto number_of_transactions = state.range(2);
+  std::random_device r;
+  std::default_random_engine e1(r());
+  std::uniform_int_distribution<int> uniform_dist(0, number_of_vertices);
+
+  const auto max_transactions_needed = std::max(number_of_vertices, number_of_edges);
+  for (int64_t vertex_counter{number_of_vertices}, edge_counter{number_of_edges}, i{0};
+       vertex_counter > 0 || edge_counter > 0; --vertex_counter, --edge_counter, ++i) {
+    auto acc = storage->Access(GetNextHlc());
+    if (vertex_counter > 0) {
+      MG_ASSERT(acc.CreateVertexAndValidate({secondary_label}, PrimaryKey{PropertyValue(i)},
+                                            {{secondary_property, PropertyValue(i)}})
+                    .HasValue(),
+                "Failed creating with pk {}", i);
+    }
+    if (edge_counter > 0 && i > 1) {
+      const auto vtx1 = uniform_dist(e1) % std::min(i, number_of_vertices);
+      const auto vtx2 = uniform_dist(e1) % std::min(i, number_of_vertices);
+
+      MG_ASSERT(acc.CreateEdge(VertexId{primary_label, {PropertyValue(vtx1)}},
+                               VertexId{primary_label, {PropertyValue(vtx2)}}, edge_type_id, Gid::FromUint(i))
+                    .HasValue(),
+                "Failed on {} and {}", vtx1, vtx2);
+    }
+
+    acc.Commit(GetNextHlc());
+    if (i == max_transactions_needed - number_of_transactions) {
+      storage->CollectGarbage(GetNextHlc().coordinator_wall_clock);
+    }
+  }
+
+  for (auto _ : state) {
+    // Don't create shard since shard deallocation can take some time as well
+    auto data = storage->PerformSplit(PrimaryKey{PropertyValue{number_of_vertices / 2}}, last_hlc, GetNextHlc());
+  }
+}
+
+// Range:
+// Number of vertices
+// This run is pessimistic, number of vertices corresponds with number if transactions
+BENCHMARK_REGISTER_F(ShardSplitBenchmark, BigDataSplit)
+    ->RangeMultiplier(10)
+    ->Range(100'000, 100'000)
+    ->Unit(::benchmark::kMillisecond);
+
+// Range:
+// Number of vertices
+// This run is optimistic, in this run there are no transactions
+BENCHMARK_REGISTER_F(ShardSplitBenchmark, BigDataSplitWithGc)
+    ->RangeMultiplier(10)
+    ->Range(100'000, 1'000'000)
+    ->Unit(::benchmark::kMillisecond);
+
+// Args:
+// Number of vertices
+// Number of edges
+// Number of transaction
+BENCHMARK_REGISTER_F(ShardSplitBenchmark, BigDataSplitWithFewTransactions)
+    ->Args({100'000, 100'000, 1'000})
+    ->Args({100'000, 100'000, 10'000})
+    ->Args({1'000'000, 100'000, 1'000})
+    ->Args({1'000'000, 100'000, 10'000})
+    ->Args({100'000, 1'000'000, 1'000})
+    ->Args({1'000'000, 1'00'000, 10'000})
+    ->Unit(::benchmark::kMillisecond);
+
+}  // namespace memgraph::benchmark
+
+BENCHMARK_MAIN();

tests/simulation/CMakeLists.txt

@@ -31,6 +31,7 @@ add_simulation_test(raft.cpp)
 add_simulation_test(trial_query_storage/query_storage_test.cpp)
 add_simulation_test(sharded_map.cpp)
 add_simulation_test(shard_rsm.cpp)
-add_simulation_test(cluster_property_test.cpp)
-add_simulation_test(cluster_property_test_cypher_queries.cpp)
+add_simulation_test(bulk_load.cpp)
+add_simulation_test(random_workload.cpp)
 add_simulation_test(request_router.cpp)
+add_simulation_test(property_test_demo.cpp)

tests/simulation/bulk_load.cpp

@@ -33,24 +33,41 @@ using io::Time;
 using io::simulator::SimulatorConfig;
 using storage::v3::kMaximumCronInterval;
 
-RC_GTEST_PROP(RandomClusterConfig, HappyPath, (ClusterConfig cluster_config, NonEmptyOpVec ops, uint64_t rng_seed)) {
+RC_GTEST_PROP(RandomClusterConfig, BulkLoadAndSplit,
+              (ClusterConfig cluster_config, uint8_t inserts, uint64_t rng_seed)) {
   spdlog::cfg::load_env_levels();
 
+  // This is a static workload that just inserts vertices and reads them back, which implicitly triggers concurrent
+  // splits
+  std::vector<Op> ops{};
+
+  for (int key = 0; key < inserts; key++) {
+    Op op1 = {.inner = AssertShardsSplit{}};
+    ops.emplace_back(std::move(op1));
+
+    Op op2 = {.inner = CreateVertex{.first = 0, .second = key}};
+    ops.emplace_back(std::move(op2));
+  }
+
+  Op op1 = {.inner = AssertShardsSplit{}};
+  ops.emplace_back(std::move(op1));
+
+  ops.emplace_back(Op{.inner = ScanAll{}});
+
   SimulatorConfig sim_config{
       .drop_percent = 0,
-      .perform_timeouts = false,
+      .perform_timeouts = true,
       .scramble_messages = true,
       .rng_seed = rng_seed,
       .start_time = Time::min(),
+      // TODO(tyler) set abort_time to something more restrictive than Time::max()
       .abort_time = Time::max(),
   };
 
-  std::vector<std::string> queries = {"CREATE (n:test_label{property_1: 0, property_2: 0});", "MATCH (n) RETURN n;"};
+  auto [sim_stats_1, latency_stats_1] = RunClusterSimulation(sim_config, cluster_config, ops);
+  auto [sim_stats_2, latency_stats_2] = RunClusterSimulation(sim_config, cluster_config, ops);
 
-  auto [sim_stats_1, latency_stats_1] = RunClusterSimulationWithQueries(sim_config, cluster_config, queries);
-  auto [sim_stats_2, latency_stats_2] = RunClusterSimulationWithQueries(sim_config, cluster_config, queries);
-
-  if (latency_stats_1 != latency_stats_2) {
+  if (latency_stats_1 != latency_stats_2 || sim_stats_1 != sim_stats_2) {
     spdlog::error("simulator stats diverged across runs");
     spdlog::error("run 1 simulator stats: {}", sim_stats_1);
     spdlog::error("run 2 simulator stats: {}", sim_stats_2);
@@ -59,6 +76,8 @@ RC_GTEST_PROP(RandomClusterConfig, HappyPath, (ClusterConfig cluster_config, Non
     RC_ASSERT(latency_stats_1 == latency_stats_2);
     RC_ASSERT(sim_stats_1 == sim_stats_2);
   }
+
+  spdlog::trace("passed stats comparison - all good!");
 }
 
 }  // namespace memgraph::tests::simulation

tests/simulation/cluster_config.hpp

@@ -20,11 +20,11 @@ namespace memgraph::tests::simulation {
 struct ClusterConfig {
   int servers;
   int replication_factor;
-  int shards;
+  int split_threshold;
 
   friend std::ostream &operator<<(std::ostream &in, const ClusterConfig &cluster) {
     in << "ClusterConfig { servers: " << cluster.servers << ", replication_factor: " << cluster.replication_factor
-       << ", shards: " << cluster.shards << " }";
+       << " }";
     return in;
   }
 };
@@ -44,7 +44,7 @@ struct Arbitrary<ClusterConfig> {
         gen::set(&ClusterConfig::servers, gen::inRange(kMinimumServers, kMaximumServers)),
         gen::set(&ClusterConfig::replication_factor,
                  gen::inRange(kMinimumReplicationFactor, kMaximumReplicationFactor)),
-        gen::set(&ClusterConfig::shards, gen::inRange(kMinimumShards, kMaximumShards)));
+        gen::set(&ClusterConfig::split_threshold, gen::inRange(kMinimumSplitThreshold, kMaximumSplitThreshold)));
   }
 };
 

tests/simulation/common.hpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -130,6 +130,7 @@ class MockedShardRsm {
   msgs::DeleteEdgesResponse ApplyImpl(msgs::DeleteEdgesRequest rqst) { return {}; }
   msgs::UpdateEdgesResponse ApplyImpl(msgs::UpdateEdgesRequest rqst) { return {}; }
   msgs::CommitResponse ApplyImpl(msgs::CommitRequest rqst) { return {}; }
+  msgs::CommitResponse ApplyImpl(msgs::SplitRequest rqst) { return {}; }
 
   WriteResponses Apply(WriteRequests write_requests) {
     return {std::visit([this]<typename T>(T &&request) { return WriteResponses{ApplyImpl(std::forward<T>(request))}; },

tests/simulation/generated_operations.hpp

@@ -35,13 +35,20 @@ struct CreateVertex {
 };
 
 struct ScanAll {
-  friend std::ostream &operator<<(std::ostream &in, const ScanAll &get) {
+  friend std::ostream &operator<<(std::ostream &in, const ScanAll & /* unused */) {
     in << "ScanAll {}";
     return in;
   }
 };
 
-using OpVariant = std::variant<CreateVertex, ScanAll>;
+struct AssertShardsSplit {
+  friend std::ostream &operator<<(std::ostream &in, const AssertShardsSplit & /* unused */) {
+    in << "AssertShardsSplit {}";
+    return in;
+  }
+};
+
+using OpVariant = std::variant<CreateVertex, ScanAll, AssertShardsSplit>;
 
 struct Op {
   OpVariant inner;
@@ -81,8 +88,8 @@ using namespace memgraph::tests::simulation;
 template <>
 struct Arbitrary<CreateVertex> {
   static Gen<CreateVertex> arbitrary() {
-    return gen::build<CreateVertex>(gen::set(&CreateVertex::first, gen::inRange(0, kMaximumShards + 1)),
-                                    gen::set(&CreateVertex::second, gen::inRange(0, kMaximumShards + 1)));
+    return gen::build<CreateVertex>(gen::set(&CreateVertex::first, gen::inRange(0, kMaximumKey)),
+                                    gen::set(&CreateVertex::second, gen::inRange(0, kMaximumKey)));
   }
 };
 
@@ -91,15 +98,23 @@ struct Arbitrary<ScanAll> {
   static Gen<ScanAll> arbitrary() { return gen::just(ScanAll{}); }
 };
 
+template <>
+struct Arbitrary<AssertShardsSplit> {
+  static Gen<AssertShardsSplit> arbitrary() { return gen::just(AssertShardsSplit{}); }
+};
+
 OpVariant opHoist(ScanAll op) { return op; }
 OpVariant opHoist(CreateVertex op) { return op; }
+OpVariant opHoist(AssertShardsSplit op) { return op; }
 
 template <>
 struct ::rc::Arbitrary<Op> {
   static Gen<Op> arbitrary() {
     return gen::build<Op>(gen::set(
-        &Op::inner, gen::oneOf(gen::map(gen::arbitrary<CreateVertex>(), [](CreateVertex op) { return opHoist(op); }),
-                               gen::map(gen::arbitrary<ScanAll>(), [](ScanAll op) { return opHoist(op); }))));
+        &Op::inner,
+        gen::oneOf(gen::map(gen::arbitrary<CreateVertex>(), [](CreateVertex op) { return opHoist(op); }),
+                   gen::map(gen::arbitrary<ScanAll>(), [](ScanAll op) { return opHoist(op); }),
+                   gen::map(gen::arbitrary<AssertShardsSplit>(), [](AssertShardsSplit op) { return opHoist(op); }))));
   }
 };
 

tests/simulation/property_test_demo.cpp

@@ -0,0 +1,37 @@
+// Copyright 2023 Memgraph Ltd.
+//
+// Use of this software is governed by the Business Source License
+// included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
+// License, and you may not use this file except in compliance with the Business Source License.
+//
+// As of the Change Date specified in that file, in accordance with
+// the Business Source License, use of this software will be governed
+// by the Apache License, Version 2.0, included in the file
+// licenses/APL.txt.
+
+// This test serves as an example of a property-based model test.
+// It generates a cluster configuration and a set of operations to
+// apply against both the real system and a greatly simplified model.
+
+#include <chrono>
+
+#include <gtest/gtest.h>
+#include <rapidcheck.h>
+#include <rapidcheck/gtest.h>
+#include <spdlog/cfg/env.h>
+
+#include "generated_operations.hpp"
+
+namespace memgraph::tests::simulation {
+
+void my_code(NonEmptyOpVec input) { RC_ASSERT(input.ops.size() < 3); }
+
+RC_GTEST_PROP(PropertyTestDemo, Demo1, (NonEmptyOpVec ops, uint64_t rng_seed)) {
+  spdlog::cfg::load_env_levels();
+
+  my_code(ops);
+
+  spdlog::trace("passed stats comparison - all good!");
+}
+
+}  // namespace memgraph::tests::simulation

tests/simulation/random_workload.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -33,12 +33,13 @@ using io::Time;
 using io::simulator::SimulatorConfig;
 using storage::v3::kMaximumCronInterval;
 
-RC_GTEST_PROP(RandomClusterConfig, HappyPath, (ClusterConfig cluster_config, NonEmptyOpVec ops, uint64_t rng_seed)) {
+RC_GTEST_PROP(RandomClusterConfig, RandomWorkload,
+              (ClusterConfig cluster_config, NonEmptyOpVec ops, uint64_t rng_seed)) {
   spdlog::cfg::load_env_levels();
 
   SimulatorConfig sim_config{
       .drop_percent = 0,
-      .perform_timeouts = false,
+      .perform_timeouts = true,
       .scramble_messages = true,
       .rng_seed = rng_seed,
       .start_time = Time::min(),
@@ -49,7 +50,7 @@ RC_GTEST_PROP(RandomClusterConfig, HappyPath, (ClusterConfig cluster_config, Non
   auto [sim_stats_1, latency_stats_1] = RunClusterSimulation(sim_config, cluster_config, ops.ops);
   auto [sim_stats_2, latency_stats_2] = RunClusterSimulation(sim_config, cluster_config, ops.ops);
 
-  if (latency_stats_1 != latency_stats_2) {
+  if (latency_stats_1 != latency_stats_2 || sim_stats_1 != sim_stats_2) {
     spdlog::error("simulator stats diverged across runs");
     spdlog::error("run 1 simulator stats: {}", sim_stats_1);
     spdlog::error("run 2 simulator stats: {}", sim_stats_2);
@@ -58,6 +59,8 @@ RC_GTEST_PROP(RandomClusterConfig, HappyPath, (ClusterConfig cluster_config, Non
     RC_ASSERT(latency_stats_1 == latency_stats_2);
     RC_ASSERT(sim_stats_1 == sim_stats_2);
   }
+
+  spdlog::trace("passed stats comparison - all good!");
 }
 
 }  // namespace memgraph::tests::simulation

tests/simulation/request_router.cpp

@@ -24,6 +24,7 @@
 #include "common/types.hpp"
 #include "coordinator/coordinator_client.hpp"
 #include "coordinator/coordinator_rsm.hpp"
+#include "coordinator/shard_map.hpp"
 #include "io/address.hpp"
 #include "io/errors.hpp"
 #include "io/rsm/raft.hpp"
@@ -39,7 +40,7 @@
 #include "utils/result.hpp"
 
 namespace memgraph::query::v2::tests {
-using coordinator::AddressAndStatus;
+using coordinator::PeerMetadata;
 using CompoundKey = coordinator::PrimaryKey;
 using coordinator::Coordinator;
 using coordinator::CoordinatorClient;
@@ -100,7 +101,8 @@ ShardMap CreateDummyShardmap(coordinator::Address a_io_1, coordinator::Address a
       SchemaProperty{.property_id = property_id_2, .type = type_2},
   };
 
-  auto label_success = sm.InitializeNewLabel(label_name, schema, 1, sm.shard_map_version);
+  const auto split_threshold = 4;
+  auto label_success = sm.InitializeNewLabel(label_name, schema, 1, split_threshold, sm.shard_map_version);
   MG_ASSERT(label_success);
 
   const LabelId label_id = sm.labels.at(label_name);
@@ -109,11 +111,11 @@ ShardMap CreateDummyShardmap(coordinator::Address a_io_1, coordinator::Address a
   shards_for_label.clear();
 
   // add first shard at [0, 0]
-  AddressAndStatus aas1_1{.address = a_io_1, .status = Status::CONSENSUS_PARTICIPANT};
-  AddressAndStatus aas1_2{.address = a_io_2, .status = Status::CONSENSUS_PARTICIPANT};
-  AddressAndStatus aas1_3{.address = a_io_3, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas1_1{.address = a_io_1, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas1_2{.address = a_io_2, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas1_3{.address = a_io_3, .status = Status::CONSENSUS_PARTICIPANT};
 
-  ShardMetadata shard1 = ShardMetadata{.peers = {aas1_1, aas1_2, aas1_3}, .version = 1};
+  ShardMetadata shard1 = ShardMetadata{.peers = {aas1_1, aas1_2, aas1_3}, .version = {1}};
 
   auto key1 = storage::v3::PropertyValue(0);
   auto key2 = storage::v3::PropertyValue(0);
@@ -121,11 +123,11 @@ ShardMap CreateDummyShardmap(coordinator::Address a_io_1, coordinator::Address a
   shards_for_label[compound_key_1] = shard1;
 
   // add second shard at [12, 13]
-  AddressAndStatus aas2_1{.address = b_io_1, .status = Status::CONSENSUS_PARTICIPANT};
-  AddressAndStatus aas2_2{.address = b_io_2, .status = Status::CONSENSUS_PARTICIPANT};
-  AddressAndStatus aas2_3{.address = b_io_3, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas2_1{.address = b_io_1, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas2_2{.address = b_io_2, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas2_3{.address = b_io_3, .status = Status::CONSENSUS_PARTICIPANT};
 
-  ShardMetadata shard2 = ShardMetadata{.peers = {aas2_1, aas2_2, aas2_3}, .version = 1};
+  ShardMetadata shard2 = ShardMetadata{.peers = {aas2_1, aas2_2, aas2_3}, .version = {1}};
 
   auto key3 = storage::v3::PropertyValue(12);
   auto key4 = storage::v3::PropertyValue(13);

tests/simulation/shard_rsm.cpp

@@ -1496,9 +1496,13 @@ int TestMessages() {
 
   std::vector<SchemaProperty> schema_prop = {get_schema_property()};
 
-  auto shard_ptr1 = std::make_unique<Shard>(get_primary_label(), min_prim_key, max_prim_key, schema_prop);
-  auto shard_ptr2 = std::make_unique<Shard>(get_primary_label(), min_prim_key, max_prim_key, schema_prop);
-  auto shard_ptr3 = std::make_unique<Shard>(get_primary_label(), min_prim_key, max_prim_key, schema_prop);
+  const coordinator::Hlc shard_version{0};
+  auto shard_ptr1 =
+      std::make_unique<Shard>(get_primary_label(), min_prim_key, max_prim_key, schema_prop, shard_version);
+  auto shard_ptr2 =
+      std::make_unique<Shard>(get_primary_label(), min_prim_key, max_prim_key, schema_prop, shard_version);
+  auto shard_ptr3 =
+      std::make_unique<Shard>(get_primary_label(), min_prim_key, max_prim_key, schema_prop, shard_version);
 
   shard_ptr1->StoreMapping(
       {{1, "label"}, {2, "prop1"}, {3, "label1"}, {4, "prop2"}, {5, "prop3"}, {6, "prop4"}, {7, "e_prop"}});
@@ -1511,9 +1515,13 @@ int TestMessages() {
   std::vector<Address> address_for_2{shard_server_1_address, shard_server_3_address};
   std::vector<Address> address_for_3{shard_server_1_address, shard_server_2_address};
 
-  ConcreteShardRsm shard_server1(std::move(shard_server_io_1), address_for_1, ShardRsm(std::move(shard_ptr1)));
-  ConcreteShardRsm shard_server2(std::move(shard_server_io_2), address_for_2, ShardRsm(std::move(shard_ptr2)));
-  ConcreteShardRsm shard_server3(std::move(shard_server_io_3), address_for_3, ShardRsm(std::move(shard_ptr3)));
+  utils::Sender<msgs::InitializeSplitShard> local_shard_manager_sender{[](const auto &elem) {}};
+  ConcreteShardRsm shard_server1(std::move(shard_server_io_1), address_for_1,
+                                 ShardRsm(std::move(shard_ptr1), local_shard_manager_sender));
+  ConcreteShardRsm shard_server2(std::move(shard_server_io_2), address_for_2,
+                                 ShardRsm(std::move(shard_ptr2), local_shard_manager_sender));
+  ConcreteShardRsm shard_server3(std::move(shard_server_io_3), address_for_3,
+                                 ShardRsm(std::move(shard_ptr3), local_shard_manager_sender));
 
   auto server_thread1 = std::jthread([&shard_server1]() { shard_server1.Run(); });
   simulator.IncrementServerCountAndWaitForQuiescentState(shard_server_1_address);

tests/simulation/sharded_map.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -33,12 +33,12 @@
 #include "utils/result.hpp"
 
 using memgraph::common::SchemaType;
-using memgraph::coordinator::AddressAndStatus;
 using memgraph::coordinator::Coordinator;
 using memgraph::coordinator::CoordinatorClient;
 using memgraph::coordinator::CoordinatorRsm;
 using memgraph::coordinator::HlcRequest;
 using memgraph::coordinator::HlcResponse;
+using memgraph::coordinator::PeerMetadata;
 using memgraph::coordinator::PrimaryKey;
 using memgraph::coordinator::ShardMap;
 using memgraph::coordinator::ShardMetadata;
@@ -94,9 +94,10 @@ ShardMap CreateDummyShardmap(Address a_io_1, Address a_io_2, Address a_io_3, Add
       SchemaProperty{.property_id = property_id_1, .type = type_1},
       SchemaProperty{.property_id = property_id_2, .type = type_2},
   };
-  size_t replication_factor = 3;
+  const size_t replication_factor = 3;
+  const auto split_threshold = 999;
   std::optional<LabelId> label_id_opt =
-      sm.InitializeNewLabel(label_name, schema, replication_factor, sm.shard_map_version);
+      sm.InitializeNewLabel(label_name, schema, replication_factor, split_threshold, sm.shard_map_version);
   MG_ASSERT(label_id_opt.has_value());
 
   const LabelId label_id = label_id_opt.value();
@@ -105,9 +106,9 @@ ShardMap CreateDummyShardmap(Address a_io_1, Address a_io_2, Address a_io_3, Add
   shards_for_label.clear();
 
   // add first shard at [0, 0]
-  AddressAndStatus aas1_1{.address = a_io_1, .status = Status::CONSENSUS_PARTICIPANT};
-  AddressAndStatus aas1_2{.address = a_io_2, .status = Status::CONSENSUS_PARTICIPANT};
-  AddressAndStatus aas1_3{.address = a_io_3, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas1_1{.address = a_io_1, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas1_2{.address = a_io_2, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas1_3{.address = a_io_3, .status = Status::CONSENSUS_PARTICIPANT};
 
   ShardMetadata shard1 = ShardMetadata{.peers = {aas1_1, aas1_2, aas1_3}, .version = 1};
 
@@ -117,9 +118,9 @@ ShardMap CreateDummyShardmap(Address a_io_1, Address a_io_2, Address a_io_3, Add
   shards_for_label.emplace(compound_key_1, shard1);
 
   // add second shard at [12, 13]
-  AddressAndStatus aas2_1{.address = b_io_1, .status = Status::CONSENSUS_PARTICIPANT};
-  AddressAndStatus aas2_2{.address = b_io_2, .status = Status::CONSENSUS_PARTICIPANT};
-  AddressAndStatus aas2_3{.address = b_io_3, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas2_1{.address = b_io_1, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas2_2{.address = b_io_2, .status = Status::CONSENSUS_PARTICIPANT};
+  PeerMetadata aas2_3{.address = b_io_3, .status = Status::CONSENSUS_PARTICIPANT};
 
   ShardMetadata shard2 = ShardMetadata{.peers = {aas2_1, aas2_2, aas2_3}, .version = 1};
 

tests/simulation/simulation_interpreter.hpp

@@ -27,9 +27,18 @@
 
 namespace memgraph::io::simulator {
 
-class SimulatedInterpreter {
-  using ResultStream = query::v2::DiscardValueResultStream;
+class FlatStream {
+  std::vector<query::v2::TypedValue> all_results_;
 
+ public:
+  void Result(const std::vector<query::v2::TypedValue> &new_values) {
+    all_results_.insert(all_results_.end(), new_values.begin(), new_values.end());
+  }
+
+  std::vector<query::v2::TypedValue> &&TakeResults() { return std::move(all_results_); }
+};
+
+class SimulatedInterpreter {
  public:
   explicit SimulatedInterpreter(std::unique_ptr<query::v2::InterpreterContext> interpreter_context)
       : interpreter_context_(std::move(interpreter_context)) {
@@ -42,23 +51,8 @@ class SimulatedInterpreter {
   SimulatedInterpreter &operator=(SimulatedInterpreter &&) = delete;
   ~SimulatedInterpreter() = default;
 
-  void InstallSimulatorTicker(Simulator &simulator) {
-    interpreter_->InstallSimulatorTicker(simulator.GetSimulatorTickClosure());
-  }
-
-  std::vector<ResultStream> RunQueries(const std::vector<std::string> &queries) {
-    std::vector<ResultStream> results;
-    results.reserve(queries.size());
-
-    for (const auto &query : queries) {
-      results.emplace_back(RunQuery(query));
-    }
-    return results;
-  }
-
- private:
-  ResultStream RunQuery(const std::string &query) {
-    ResultStream stream;
+  std::vector<query::v2::TypedValue> RunQuery(const std::string &query) {
+    FlatStream stream;
 
     std::map<std::string, memgraph::storage::v3::PropertyValue> params;
     const std::string *username = nullptr;
@@ -66,9 +60,10 @@ class SimulatedInterpreter {
     interpreter_->Prepare(query, params, username);
     interpreter_->PullAll(&stream);
 
-    return stream;
+    return stream.TakeResults();
   }
 
+ private:
   std::unique_ptr<query::v2::InterpreterContext> interpreter_context_;
   std::unique_ptr<query::v2::Interpreter> interpreter_;
 };

tests/simulation/test_cluster.hpp

@@ -36,6 +36,7 @@
 #include "utils/print_helpers.hpp"
 #include "utils/variant_helpers.hpp"
 
+#include "generated_operations.hpp"
 #include "simulation_interpreter.hpp"
 
 namespace memgraph::tests::simulation {
@@ -71,6 +72,13 @@ using storage::v3::SchemaProperty;
 using CompoundKey = std::pair<int, int>;
 using ShardClient = RsmClient<SimulatorTransport, WriteRequests, WriteResponses, ReadRequests, ReadResponses>;
 
+struct SimClientContext {
+  CoordinatorClient<SimulatorTransport> coordinator_client;
+  const ClusterConfig &cluster_config;
+  io::simulator::SimulatedInterpreter interpreter;
+  std::set<CompoundKey> correctness_model;
+};
+
 MachineManager<SimulatorTransport> MkMm(Simulator &simulator, std::vector<Address> coordinator_addresses, Address addr,
                                         ShardMap shard_map) {
   MachineConfig config{
@@ -116,9 +124,8 @@ void WaitForShardsToInitialize(CoordinatorClient<SimulatorTransport> &coordinato
   }
 }
 
-ShardMap TestShardMap(int n_splits, int replication_factor) {
+ShardMap TestShardMap(const ClusterConfig &cluster_config) {
   ShardMap sm{};
-
   const std::string label_name = std::string("test_label");
 
   // register new properties
@@ -135,65 +142,100 @@ ShardMap TestShardMap(int n_splits, int replication_factor) {
       SchemaProperty{.property_id = property_id_2, .type = type_2},
   };
 
-  std::optional<LabelId> label_id = sm.InitializeNewLabel(label_name, schema, replication_factor, sm.shard_map_version);
+  std::optional<LabelId> label_id = sm.InitializeNewLabel(label_name, schema, cluster_config.replication_factor,
+                                                          cluster_config.split_threshold, sm.shard_map_version);
   RC_ASSERT(label_id.has_value());
 
-  // split the shard at N split points
-  for (int64_t i = 1; i < n_splits; ++i) {
-    const auto key1 = memgraph::storage::v3::PropertyValue(i);
-    const auto key2 = memgraph::storage::v3::PropertyValue(0);
-
-    const auto split_point = {key1, key2};
-
-    const bool split_success = sm.SplitShard(sm.shard_map_version, label_id.value(), split_point);
-
-    RC_ASSERT(split_success);
-  }
-
   return sm;
 }
 
-void ExecuteOp(query::v2::RequestRouter<SimulatorTransport> &request_router, std::set<CompoundKey> &correctness_model,
-               CreateVertex create_vertex) {
+void ExecuteOp(SimClientContext &context, CreateVertex create_vertex) {
   const auto key1 = memgraph::storage::v3::PropertyValue(create_vertex.first);
   const auto key2 = memgraph::storage::v3::PropertyValue(create_vertex.second);
 
   std::vector<msgs::Value> primary_key = {msgs::Value(int64_t(create_vertex.first)),
                                           msgs::Value(int64_t(create_vertex.second))};
 
-  if (correctness_model.contains(std::make_pair(create_vertex.first, create_vertex.second))) {
+  if (context.correctness_model.contains(std::make_pair(create_vertex.first, create_vertex.second))) {
     // TODO(tyler) remove this early-return when we have properly handled setting non-unique vertexes
     return;
   }
 
-  auto label_id = request_router.NameToLabel("test_label");
+  std::string query = fmt::format("CREATE (n:test_label{{property_1: {}, property_2: {}}});", create_vertex.first,
+                                  create_vertex.second);
 
-  msgs::NewVertex nv{.primary_key = primary_key};
-  nv.label_ids.push_back({label_id});
+  auto result_stream = context.interpreter.RunQuery(query);
 
-  std::vector<msgs::NewVertex> new_vertices;
-  new_vertices.push_back(std::move(nv));
+  // TODO(tyler) is there a better way to assert the actual success of the operation?
+  RC_ASSERT(result_stream.size() == 0);
 
-  auto result = request_router.CreateVertices(std::move(new_vertices));
+  // RC_ASSERT(!result_stream[0].error.has_value());
 
-  RC_ASSERT(result.size() == 1);
-  RC_ASSERT(!result[0].error.has_value());
-
-  correctness_model.emplace(std::make_pair(create_vertex.first, create_vertex.second));
+  context.correctness_model.emplace(std::make_pair(create_vertex.first, create_vertex.second));
 }
 
-void ExecuteOp(query::v2::RequestRouter<SimulatorTransport> &request_router, std::set<CompoundKey> &correctness_model,
-               ScanAll scan_all) {
-  auto results = request_router.ScanVertices("test_label");
+void ExecuteOp(SimClientContext &context, ScanAll scan_all) {
+  auto results = context.interpreter.RunQuery("MATCH (n) RETURN n;");
 
-  RC_ASSERT(results.size() == correctness_model.size());
+  auto result_set = std::set<CompoundKey>{};
 
-  for (const auto &vertex_accessor : results) {
-    const auto properties = vertex_accessor.Properties();
-    const auto primary_key = vertex_accessor.Id().second;
-    const CompoundKey model_key = std::make_pair(primary_key[0].int_v, primary_key[1].int_v);
-    RC_ASSERT(correctness_model.contains(model_key));
+  for (const auto &typed_value : results) {
+    auto &vertex = typed_value.ValueVertex();
+    auto &pk = vertex.Properties();
+    auto &[l1, v1] = pk[0];
+    auto &[l2, v2] = pk[1];
+    auto compound_key = std::make_pair(v1.int_v, v2.int_v);
+    result_set.emplace(compound_key);
   }
+
+  // TODO(tyler) we are getting more results back than expected, due to
+  // race conditions during shard splits causing multiple shards to be
+  // queried by the client. The std::set above papers-over this, but
+  // it's an issue that needs to be addressed eventually.
+  RC_ASSERT(result_set == context.correctness_model);
+}
+
+void ExecuteOp(SimClientContext &context, AssertShardsSplit assert_shards_split) {
+  // 1 -> 1
+  // 2 -> 1
+  // 3 -> 1
+  // 4 -> 2
+  // 5 -> 2
+  // 6 -> 2
+  // 7 -> 3
+  const int max_shard_size = context.cluster_config.split_threshold - 1;
+  const int min_shards = (std::max((size_t)1, context.correctness_model.size()) - 1) / max_shard_size;
+  const int minimum_expected_shards = min_shards + 1;
+  // TODO(tyler) make this a higher number of retries
+  const int maximum_attempts = 100;
+  size_t initialized_shards;
+
+  for (int i = 0; i < maximum_attempts; i++) {
+    GetShardMapRequest req{};
+    CoordinatorReadRequests read_req = req;
+    auto read_res = context.coordinator_client.SendReadRequest(read_req);
+    if (read_res.HasError()) {
+      // timed out
+      continue;
+    }
+    auto response_result = read_res.GetValue();
+    auto response = std::get<GetShardMapResponse>(response_result);
+    auto shard_map = response.shard_map;
+
+    initialized_shards = shard_map.InitializedShards();
+
+    if (initialized_shards >= minimum_expected_shards) {
+      spdlog::info(
+          "AssertShardsSplit returning after we see {} initialized shards ({} minimum expected with model size {} and "
+          "split threshold {})",
+          initialized_shards, minimum_expected_shards, context.correctness_model.size(),
+          context.cluster_config.split_threshold);
+      return;
+    }
+  }
+
+  spdlog::error("expected {} shards, but we only have {}", minimum_expected_shards, initialized_shards);
+  MG_ASSERT(false, "exceeded maximum attempts for waiting for the expected number of shard splits to occur");
 }
 
 /// This struct exists as a way of detaching
@@ -222,15 +264,17 @@ std::pair<SimulatorStats, LatencyHistogramSummaries> RunClusterSimulation(const
   auto machine_1_addr = Address::TestAddress(1);
   auto cli_addr = Address::TestAddress(2);
   auto cli_addr_2 = Address::TestAddress(3);
+  auto cli_addr_3 = Address::TestAddress(4);
 
   Io<SimulatorTransport> cli_io = simulator.Register(cli_addr);
   Io<SimulatorTransport> cli_io_2 = simulator.Register(cli_addr_2);
+  Io<SimulatorTransport> cli_io_3 = simulator.Register(cli_addr_3);
 
   auto coordinator_addresses = std::vector{
       machine_1_addr,
   };
 
-  ShardMap initialization_sm = TestShardMap(cluster_config.shards - 1, cluster_config.replication_factor);
+  ShardMap initialization_sm = TestShardMap(cluster_config);
 
   auto mm_1 = MkMm(simulator, coordinator_addresses, machine_1_addr, initialization_sm);
   Address coordinator_address = mm_1.CoordinatorAddress();
@@ -242,19 +286,19 @@ std::pair<SimulatorStats, LatencyHistogramSummaries> RunClusterSimulation(const
 
   // TODO(tyler) clarify addresses of coordinator etc... as it's a mess
 
-  CoordinatorClient<SimulatorTransport> coordinator_client(cli_io, coordinator_address, {coordinator_address});
-  WaitForShardsToInitialize(coordinator_client);
-
-  query::v2::RequestRouter<SimulatorTransport> request_router(std::move(coordinator_client), std::move(cli_io));
-  std::function<bool()> tick_simulator = simulator.GetSimulatorTickClosure();
-  request_router.InstallSimulatorTicker(tick_simulator);
-
-  request_router.StartTransaction();
-
   auto correctness_model = std::set<CompoundKey>{};
 
+  SimClientContext context{
+      .coordinator_client = CoordinatorClient<SimulatorTransport>(cli_io, coordinator_address, {coordinator_address}),
+      .cluster_config = cluster_config,
+      .interpreter = io::simulator::SetUpInterpreter(coordinator_address, simulator),
+      .correctness_model = correctness_model,
+  };
+
+  WaitForShardsToInitialize(context.coordinator_client);
+
   for (const Op &op : ops) {
-    std::visit([&](auto &o) { ExecuteOp(request_router, correctness_model, o); }, op.inner);
+    std::visit([&](auto &o) { ExecuteOp(context, o); }, op.inner);
   }
 
   // We have now completed our workload without failing any assertions, so we can
@@ -281,65 +325,4 @@ std::pair<SimulatorStats, LatencyHistogramSummaries> RunClusterSimulation(const
   return std::make_pair(stats, histo);
 }
 
-std::pair<SimulatorStats, LatencyHistogramSummaries> RunClusterSimulationWithQueries(
-    const SimulatorConfig &sim_config, const ClusterConfig &cluster_config, const std::vector<std::string> &queries) {
-  spdlog::info("========================== NEW SIMULATION ==========================");
-
-  auto simulator = Simulator(sim_config);
-
-  auto machine_1_addr = Address::TestAddress(1);
-  auto cli_addr = Address::TestAddress(2);
-  auto cli_addr_2 = Address::TestAddress(3);
-
-  Io<SimulatorTransport> cli_io = simulator.Register(cli_addr);
-  Io<SimulatorTransport> cli_io_2 = simulator.Register(cli_addr_2);
-
-  auto coordinator_addresses = std::vector{
-      machine_1_addr,
-  };
-
-  ShardMap initialization_sm = TestShardMap(cluster_config.shards - 1, cluster_config.replication_factor);
-
-  auto mm_1 = MkMm(simulator, coordinator_addresses, machine_1_addr, initialization_sm);
-  Address coordinator_address = mm_1.CoordinatorAddress();
-
-  auto mm_thread_1 = std::jthread(RunMachine, std::move(mm_1));
-  simulator.IncrementServerCountAndWaitForQuiescentState(machine_1_addr);
-
-  auto detach_on_error = DetachIfDropped{.handle = mm_thread_1};
-
-  // TODO(tyler) clarify addresses of coordinator etc... as it's a mess
-
-  CoordinatorClient<SimulatorTransport> coordinator_client(cli_io, coordinator_address, {coordinator_address});
-  WaitForShardsToInitialize(coordinator_client);
-
-  auto simulated_interpreter = io::simulator::SetUpInterpreter(coordinator_address, simulator);
-  simulated_interpreter.InstallSimulatorTicker(simulator);
-
-  auto query_results = simulated_interpreter.RunQueries(queries);
-
-  // We have now completed our workload without failing any assertions, so we can
-  // disable detaching the worker thread, which will cause the mm_thread_1 jthread
-  // to be joined when this function returns.
-  detach_on_error.detach = false;
-
-  simulator.ShutDown();
-
-  mm_thread_1.join();
-
-  SimulatorStats stats = simulator.Stats();
-
-  spdlog::info("total messages:     {}", stats.total_messages);
-  spdlog::info("dropped messages:   {}", stats.dropped_messages);
-  spdlog::info("timed out requests: {}", stats.timed_out_requests);
-  spdlog::info("total requests:     {}", stats.total_requests);
-  spdlog::info("total responses:    {}", stats.total_responses);
-  spdlog::info("simulator ticks:    {}", stats.simulator_ticks);
-
-  auto histo = cli_io_2.ResponseLatencies();
-
-  spdlog::info("========================== SUCCESS :) ==========================");
-  return std::make_pair(stats, histo);
-}
-
 }  // namespace memgraph::tests::simulation

tests/simulation/testing_constants.hpp

@@ -13,9 +13,7 @@
 
 namespace memgraph::tests::simulation {
 
-// TODO(tyler) increase this when we start standing up multiple machines in cluster tests
-static constexpr auto kMinimumShards = 1;
-static constexpr auto kMaximumShards = kMinimumShards + 10;
+static constexpr auto kMaximumKey = 5;
 
 // TODO(tyler) increase this when we start standing up multiple machines in cluster tests
 static constexpr auto kMinimumServers = 1;
@@ -25,4 +23,7 @@ static constexpr auto kMaximumServers = kMinimumServers + 1;
 static constexpr auto kMinimumReplicationFactor = 1;
 static constexpr auto kMaximumReplicationFactor = kMinimumReplicationFactor + 1;
 
+static constexpr auto kMinimumSplitThreshold = 4;
+static constexpr auto kMaximumSplitThreshold = 5;
+
 }  // namespace memgraph::tests::simulation

tests/unit/CMakeLists.txt

@@ -294,6 +294,9 @@ target_link_libraries(${test_prefix}storage_v3_expr mg-storage-v3 mg-expr)
 add_unit_test(storage_v3_schema.cpp)
 target_link_libraries(${test_prefix}storage_v3_schema mg-storage-v3)
 
+add_unit_test(storage_v3_shard_split.cpp)
+target_link_libraries(${test_prefix}storage_v3_shard_split mg-storage-v3  mg-query-v2)
+
 # Test mg-query-v2
 # These are commented out because of the new TypedValue in the query engine
 # add_unit_test(query_v2_interpreter.cpp ${CMAKE_SOURCE_DIR}/src/glue/v2/communication.cpp)

tests/unit/high_density_shard_create_scan.cpp

@@ -124,7 +124,7 @@ void WaitForShardsToInitialize(CoordinatorClient<LocalTransport> &coordinator_cl
   }
 }
 
-ShardMap TestShardMap(int shards, int replication_factor, int gap_between_shards) {
+ShardMap TestShardMap(int shards, int replication_factor, int split_threshold) {
   ShardMap sm{};
 
   const auto label_name = std::string("test_label");
@@ -143,21 +143,10 @@ ShardMap TestShardMap(int shards, int replication_factor, int gap_between_shards
       SchemaProperty{.property_id = property_id_2, .type = type_2},
   };
 
-  std::optional<LabelId> label_id = sm.InitializeNewLabel(label_name, schema, replication_factor, sm.shard_map_version);
+  std::optional<LabelId> label_id =
+      sm.InitializeNewLabel(label_name, schema, replication_factor, split_threshold, sm.shard_map_version);
   MG_ASSERT(label_id.has_value());
 
-  // split the shard at N split points
-  for (int64_t i = 1; i < shards; ++i) {
-    const auto key1 = memgraph::storage::v3::PropertyValue(i * gap_between_shards);
-    const auto key2 = memgraph::storage::v3::PropertyValue(0);
-
-    const auto split_point = {key1, key2};
-
-    const bool split_success = sm.SplitShard(sm.shard_map_version, label_id.value(), split_point);
-
-    MG_ASSERT(split_success);
-  }
-
   return sm;
 }
 
@@ -205,15 +194,15 @@ void ExecuteOp(query::v2::RequestRouter<LocalTransport> &request_router, std::se
   }
 }
 
-void RunWorkload(int shards, int replication_factor, int create_ops, int scan_ops, int shard_worker_threads,
-                 int gap_between_shards) {
+void RunWorkload(int shards, int replication_factor, int split_threshold, int create_ops, int scan_ops,
+                 int shard_worker_threads, int gap_between_shards) {
   spdlog::info("======================== NEW TEST ========================");
   spdlog::info("shards:               ", shards);
   spdlog::info("replication factor:   ", replication_factor);
   spdlog::info("create ops:           ", create_ops);
   spdlog::info("scan all ops:         ", scan_ops);
   spdlog::info("shard worker threads: ", shard_worker_threads);
-  spdlog::info("gap between shards:   ", gap_between_shards);
+  spdlog::info("split threshold:      ", split_threshold);
 
   LocalSystem local_system;
 
@@ -228,7 +217,7 @@ void RunWorkload(int shards, int replication_factor, int create_ops, int scan_op
   };
 
   auto time_before_shard_map_creation = cli_io_2.Now();
-  ShardMap initialization_sm = TestShardMap(shards, replication_factor, gap_between_shards);
+  ShardMap initialization_sm = TestShardMap(shards, replication_factor, split_threshold);
   auto time_after_shard_map_creation = cli_io_2.Now();
 
   auto mm_1 = MkMm(local_system, coordinator_addresses, machine_1_addr, initialization_sm, shard_worker_threads);
@@ -281,19 +270,20 @@ void RunWorkload(int shards, int replication_factor, int create_ops, int scan_op
 }
 
 TEST(MachineManager, ManyShards) {
-  auto shards_attempts = {1, 64};
-  auto shard_worker_thread_attempts = {1, 32};
-  auto replication_factor = 1;
-  auto create_ops = 128;
-  auto scan_ops = 1;
+  const auto shards_attempts = {1, 64};
+  const auto shard_worker_thread_attempts = {1, 32};
+  const auto replication_factor = 1;
+  const auto create_ops = 128;
+  const auto scan_ops = 1;
+  const auto shards = 1;
+  const auto split_threshold = create_ops / shards;
 
   std::cout << "splits threads scan_all_microseconds\n";
 
   for (const auto shards : shards_attempts) {
-    auto gap_between_shards = create_ops / shards;
-
     for (const auto shard_worker_threads : shard_worker_thread_attempts) {
-      RunWorkload(shards, replication_factor, create_ops, scan_ops, shard_worker_threads, gap_between_shards);
+      RunWorkload(shards, replication_factor, split_threshold, create_ops, scan_ops, shard_worker_threads,
+                  split_threshold);
     }
   }
 }

tests/unit/machine_manager.cpp

@@ -82,29 +82,13 @@ ShardMap TestShardMap() {
   };
 
   const size_t replication_factor = 1;
+  const size_t split_threshold = 4;
 
-  const auto label_id = sm.InitializeNewLabel(kLabelName, schema, replication_factor, sm.shard_map_version);
+  const auto label_id =
+      sm.InitializeNewLabel(kLabelName, schema, replication_factor, split_threshold, sm.shard_map_version);
   EXPECT_TRUE(label_id.has_value());
 
   sm.AllocateEdgeTypeIds(std::vector<std::string>{"edge_type"});
-  // split the shard at N split points
-  // NB: this is the logic that should be provided by the "split file"
-  // TODO(tyler) split points should account for signedness
-  const size_t n_splits = 16;
-  const auto split_interval = std::numeric_limits<int64_t>::max() / n_splits;
-
-  for (int64_t i = 0; i < n_splits; ++i) {
-    const int64_t value = i * split_interval;
-
-    const auto key1 = memgraph::storage::v3::PropertyValue(value);
-    const auto key2 = memgraph::storage::v3::PropertyValue(0);
-
-    const CompoundKey split_point = {key1, key2};
-
-    const auto split_success = sm.SplitShard(sm.shard_map_version, label_id.value(), split_point);
-
-    EXPECT_TRUE(split_success);
-  }
 
   return sm;
 }

tests/unit/query_v2_expression_evaluator.cpp

@@ -188,7 +188,8 @@ ShardMap CreateDummyShardmap() {
       SchemaProperty{.property_id = property_id_7, .type = type_1},
   };
 
-  auto label_success = sm.InitializeNewLabel(label_name, schema, 1, sm.shard_map_version);
+  const auto split_threshold = 4;
+  auto label_success = sm.InitializeNewLabel(label_name, schema, 1, split_threshold, sm.shard_map_version);
   MG_ASSERT(label_success);
 
   const LabelId label_id = sm.labels.at(label_name);

tests/unit/storage_v3.cpp

@@ -78,9 +78,9 @@ class StorageV3 : public ::testing::TestWithParam<bool> {
   const PropertyId primary_property{PropertyId::FromUint(2)};
   std::vector<storage::v3::SchemaProperty> schema_property_vector = {
       storage::v3::SchemaProperty{primary_property, common::SchemaType::INT}};
-  Shard store{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector,
-              Config{.gc = {.reclamation_interval = reclamation_interval}}};
   coordinator::Hlc last_hlc{0, io::Time{}};
+  Shard store{primary_label,          min_pk,   std::nullopt /*max_primary_key*/,
+              schema_property_vector, last_hlc, Config{.gc = {.reclamation_interval = reclamation_interval}}};
 };
 INSTANTIATE_TEST_SUITE_P(WithGc, StorageV3, ::testing::Values(true));
 INSTANTIATE_TEST_SUITE_P(WithoutGc, StorageV3, ::testing::Values(false));
@@ -2650,7 +2650,7 @@ TEST_P(StorageV3, TestCreateVertexAndValidate) {
               (std::map<PropertyId, PropertyValue>{{prop1, PropertyValue(111)}}));
   }
   {
-    Shard store(primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector);
+    Shard store(primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector, last_hlc);
     auto acc = store.Access(GetNextHlc());
     auto vertex1 = acc.CreateVertexAndValidate({}, {PropertyValue{0}}, {});
     auto vertex2 = acc.CreateVertexAndValidate({}, {PropertyValue{0}}, {});

tests/unit/storage_v3_edge.cpp

@@ -54,10 +54,9 @@ class StorageEdgeTest : public ::testing::TestWithParam<bool> {
   const PropertyId primary_property{PropertyId::FromUint(2)};
   std::vector<storage::v3::SchemaProperty> schema_property_vector = {
       storage::v3::SchemaProperty{primary_property, common::SchemaType::INT}};
-  Shard store{primary_label, min_pk, max_pk, schema_property_vector,
-              Config{.items = {.properties_on_edges = GetParam()}}};
-
   coordinator::Hlc last_hlc{0, io::Time{}};
+  Shard store{primary_label,          min_pk,   max_pk,
+              schema_property_vector, last_hlc, Config{.items = {.properties_on_edges = GetParam()}}};
 };
 
 INSTANTIATE_TEST_SUITE_P(EdgesWithProperties, StorageEdgeTest, ::testing::Values(true));

tests/unit/storage_v3_expr.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -136,7 +136,8 @@ class ExpressionEvaluatorTest : public ::testing::Test {
 
   std::vector<storage::v3::SchemaProperty> schema_property_vector = {
       storage::v3::SchemaProperty{primary_property, common::SchemaType::INT}};
-  Shard db{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector};
+  coordinator::Hlc last_hlc{0, io::Time{}};
+  Shard db{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector, last_hlc};
 
   Shard::Accessor storage_dba{db.Access(GetNextHlc())};
   DbAccessor dba{&storage_dba};
@@ -149,8 +150,6 @@ class ExpressionEvaluatorTest : public ::testing::Test {
   Frame frame{128};
   ExpressionEvaluator eval{&frame, symbol_table, ctx, &dba, View::OLD};
 
-  coordinator::Hlc last_hlc{0, io::Time{}};
-
   void SetUp() override { db.StoreMapping({{1, "label"}, {2, "property"}}); }
 
   std::vector<PropertyId> NamesToProperties(const std::vector<std::string> &property_names) {

tests/unit/storage_v3_indices.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -44,7 +44,8 @@ class IndexTest : public testing::Test {
       storage::v3::SchemaProperty{primary_property, common::SchemaType::INT}};
   const std::vector<PropertyValue> min_pk{PropertyValue{0}};
   const LabelId primary_label{LabelId::FromUint(1)};
-  Shard storage{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector};
+  coordinator::Hlc last_hlc{0, io::Time{}};
+  Shard storage{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector, last_hlc};
 
   const PropertyId prop_id{PropertyId::FromUint(5)};
   const PropertyId prop_val{PropertyId::FromUint(6)};
@@ -55,7 +56,6 @@ class IndexTest : public testing::Test {
   static constexpr io::Duration one_time_unit{1};
   int primary_key_id{0};
   int vertex_id{0};
-  coordinator::Hlc last_hlc{0, io::Time{}};
 
   LabelId NameToLabelId(std::string_view label_name) { return storage.NameToLabel(label_name); }
 

tests/unit/storage_v3_isolation_level.cpp

@@ -82,8 +82,9 @@ TEST_P(StorageIsolationLevelTest, Visibility) {
 
   for (auto override_isolation_level_index{0U}; override_isolation_level_index < isolation_levels.size();
        ++override_isolation_level_index) {
-    Shard store{primary_label, min_pk, max_pk, schema_property_vector,
-                Config{.transaction = {.isolation_level = default_isolation_level}}};
+    coordinator::Hlc last_hlc{0, io::Time{}};
+    Shard store{primary_label,          min_pk,   max_pk,
+                schema_property_vector, last_hlc, Config{.transaction = {.isolation_level = default_isolation_level}}};
     const auto override_isolation_level = isolation_levels[override_isolation_level_index];
     auto creator = store.Access(GetNextHlc());
     auto default_isolation_level_reader = store.Access(GetNextHlc());

tests/unit/storage_v3_shard_rsm.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -18,7 +18,6 @@
 #include <memory>
 #include <optional>
 #include <string>
-#include <variant>
 #include <vector>
 
 #include "common/types.hpp"
@@ -58,15 +57,20 @@ class ShardRSMTest : public testing::Test {
     PropertyValue max_pk(static_cast<int64_t>(10000000));
     std::vector<PropertyValue> max_prim_key = {max_pk};
 
-    auto shard_ptr1 = std::make_unique<Shard>(primary_label, min_prim_key, max_prim_key, std::vector{schema_prop});
+    coordinator::Hlc shard_version{GetTransactionId()};
+    auto shard_ptr1 =
+        std::make_unique<Shard>(primary_label, min_prim_key, max_prim_key, std::vector{schema_prop}, shard_version);
     shard_ptr1->StoreMapping({{1, "primary_label"},
                               {2, "primary_label2"},
                               {3, "label"},
                               {4, "primary_prop1"},
                               {5, "primary_prop2"},
                               {6, "prop"}});
+    coordinator::Address local_shard_manager_address = coordinator::Address();
+    utils::Sender<msgs::InitializeSplitShard> local_shard_manager_sender{[](const auto &elem) {}};
+
     shard_ptr1->CreateSchema(primary_label2, {{primary_property2, SchemaType::INT}});
-    shard_rsm = std::make_unique<ShardRsm>(std::move(shard_ptr1));
+    shard_rsm = std::make_unique<ShardRsm>(std::move(shard_ptr1), std::move(local_shard_manager_sender));
   }
 
   LabelId NameToLabel(const std::string &name) { return LabelId::FromUint(id_mapper_.NameToId(name)); }

tests/unit/storage_v3_shard_split.cpp

@@ -0,0 +1,500 @@
+// Copyright 2023 Memgraph Ltd.
+//
+// Use of this software is governed by the Business Source License
+// included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
+// License, and you may not use this file except in compliance with the Business Source License.
+//
+// As of the Change Date specified in that file, in accordance with
+// the Business Source License, use of this software will be governed
+// by the Apache License, Version 2.0, included in the file
+// licenses/APL.txt.
+
+#include <cstdint>
+#include <memory>
+
+#include <gmock/gmock-matchers.h>
+#include <gmock/gmock.h>
+#include <gtest/gtest.h>
+
+#include "coordinator/hybrid_logical_clock.hpp"
+#include "query/v2/requests.hpp"
+#include "storage/v3/delta.hpp"
+#include "storage/v3/id_types.hpp"
+#include "storage/v3/key_store.hpp"
+#include "storage/v3/mvcc.hpp"
+#include "storage/v3/property_value.hpp"
+#include "storage/v3/shard.hpp"
+#include "storage/v3/vertex.hpp"
+#include "storage/v3/vertex_id.hpp"
+
+using testing::Pair;
+using testing::UnorderedElementsAre;
+
+namespace memgraph::storage::v3::tests {
+
+class ShardSplitTest : public testing::Test {
+ protected:
+  void SetUp() override {
+    storage.StoreMapping(
+        {{1, "label"}, {2, "property"}, {3, "edge_property"}, {4, "secondary_label"}, {5, "secondary_prop"}});
+  }
+
+  const PropertyId primary_property{PropertyId::FromUint(2)};
+  const PropertyId secondary_property{PropertyId::FromUint(5)};
+  std::vector<storage::v3::SchemaProperty> schema_property_vector = {
+      storage::v3::SchemaProperty{primary_property, common::SchemaType::INT}};
+  const std::vector<PropertyValue> min_pk{PropertyValue{0}};
+  const LabelId primary_label{LabelId::FromUint(1)};
+  const LabelId secondary_label{LabelId::FromUint(4)};
+  const EdgeTypeId edge_type_id{EdgeTypeId::FromUint(3)};
+  coordinator::Hlc last_hlc{0, io::Time{}};
+  Shard storage{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector, last_hlc};
+
+  coordinator::Hlc GetNextHlc() {
+    ++last_hlc.logical_id;
+    last_hlc.coordinator_wall_clock += std::chrono::seconds(1);
+    return last_hlc;
+  }
+
+  void AssertShardState(auto &shard, const int split_min, const int split_max) {
+    auto acc = shard.Access(GetNextHlc());
+    for (int i{0}; i < split_min; ++i) {
+      EXPECT_FALSE(acc.FindVertex(PrimaryKey{{PropertyValue(i)}}, View::OLD).has_value());
+    }
+    for (int i{split_min}; i < split_max; ++i) {
+      const auto vtx = acc.FindVertex(PrimaryKey{{PropertyValue(i)}}, View::OLD);
+      ASSERT_TRUE(vtx.has_value());
+      EXPECT_TRUE(vtx->InEdges(View::OLD)->size() == 1 || vtx->OutEdges(View::OLD)->size() == 1);
+    }
+  }
+};
+
+void AssertEqVertexContainer(const VertexContainer &actual, const VertexContainer &expected) {
+  ASSERT_EQ(actual.size(), expected.size());
+
+  auto expected_it = expected.begin();
+  auto actual_it = actual.begin();
+  while (expected_it != expected.end()) {
+    EXPECT_EQ(actual_it->first, expected_it->first);
+    EXPECT_EQ(actual_it->second.deleted, expected_it->second.deleted);
+    EXPECT_EQ(actual_it->second.labels, expected_it->second.labels);
+
+    auto *expected_delta = expected_it->second.delta;
+    auto *actual_delta = actual_it->second.delta;
+    // This asserts delta chain
+    while (expected_delta != nullptr) {
+      EXPECT_EQ(actual_delta->action, expected_delta->action);
+      EXPECT_EQ(actual_delta->id, expected_delta->id);
+
+      switch (expected_delta->action) {
+        case Delta::Action::ADD_LABEL:
+        case Delta::Action::REMOVE_LABEL: {
+          EXPECT_EQ(actual_delta->label, expected_delta->label);
+          break;
+        }
+        case Delta::Action::SET_PROPERTY: {
+          EXPECT_EQ(actual_delta->property.key, expected_delta->property.key);
+          EXPECT_EQ(actual_delta->property.value, expected_delta->property.value);
+          break;
+        }
+        case Delta::Action::ADD_IN_EDGE:
+        case Delta::Action::ADD_OUT_EDGE:
+        case Delta::Action::REMOVE_IN_EDGE:
+        case Delta::Action::RECREATE_OBJECT:
+        case Delta::Action::DELETE_OBJECT:
+        case Delta::Action::REMOVE_OUT_EDGE: {
+          break;
+        }
+      }
+
+      const auto expected_prev = expected_delta->prev.Get();
+      const auto actual_prev = actual_delta->prev.Get();
+      switch (expected_prev.type) {
+        case PreviousPtr::Type::NULLPTR: {
+          ASSERT_EQ(actual_prev.type, PreviousPtr::Type::NULLPTR) << "Expected type is nullptr!";
+          break;
+        }
+        case PreviousPtr::Type::DELTA: {
+          ASSERT_EQ(actual_prev.type, PreviousPtr::Type::DELTA) << "Expected type is delta!";
+          EXPECT_EQ(actual_prev.delta->action, expected_prev.delta->action);
+          EXPECT_EQ(actual_prev.delta->id, expected_prev.delta->id);
+          break;
+        }
+        case v3::PreviousPtr::Type::EDGE: {
+          ASSERT_EQ(actual_prev.type, PreviousPtr::Type::EDGE) << "Expected type is edge!";
+          EXPECT_EQ(actual_prev.edge->gid, expected_prev.edge->gid);
+          break;
+        }
+        case v3::PreviousPtr::Type::VERTEX: {
+          ASSERT_EQ(actual_prev.type, PreviousPtr::Type::VERTEX) << "Expected type is vertex!";
+          EXPECT_EQ(actual_prev.vertex->first, expected_prev.vertex->first);
+          break;
+        }
+      }
+
+      expected_delta = expected_delta->next;
+      actual_delta = actual_delta->next;
+    }
+    EXPECT_EQ(expected_delta, nullptr);
+    EXPECT_EQ(actual_delta, nullptr);
+    ++expected_it;
+    ++actual_it;
+  }
+}
+
+void AssertEqDeltaLists(const std::list<Delta> &actual, const std::list<Delta> &expected) {
+  EXPECT_EQ(actual.size(), expected.size());
+  auto actual_it = actual.begin();
+  auto expected_it = expected.begin();
+  while (actual_it != actual.end()) {
+    EXPECT_EQ(actual_it->id, expected_it->id);
+    EXPECT_EQ(actual_it->action, expected_it->action);
+    ++actual_it;
+    ++expected_it;
+  }
+}
+
+void AddDeltaToDeltaChain(Vertex *object, Delta *new_delta) {
+  auto *delta_holder = GetDeltaHolder(object);
+
+  new_delta->next = delta_holder->delta;
+  new_delta->prev.Set(object);
+  if (delta_holder->delta) {
+    delta_holder->delta->prev.Set(new_delta);
+  }
+  delta_holder->delta = new_delta;
+}
+
+TEST_F(ShardSplitTest, TestBasicSplitWithVertices) {
+  auto acc = storage.Access(GetNextHlc());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(1)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(2)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(3)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(4)}, {}).HasError());
+  EXPECT_FALSE(
+      acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(5)}, {{secondary_property, PropertyValue(121)}})
+          .HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(6)}, {}).HasError());
+  auto current_hlc = GetNextHlc();
+  acc.Commit(current_hlc);
+
+  auto splitted_data = *storage.PerformSplit({PropertyValue(4)}, last_hlc, GetNextHlc());
+  EXPECT_EQ(splitted_data.vertices.size(), 3);
+  EXPECT_EQ(splitted_data.edges->size(), 0);
+  EXPECT_EQ(splitted_data.transactions.size(), 1);
+  EXPECT_EQ(splitted_data.label_indices.size(), 0);
+  EXPECT_EQ(splitted_data.label_property_indices.size(), 0);
+
+  CommitInfo commit_info{.start_or_commit_timestamp = current_hlc};
+  Delta delta_delete1{Delta::DeleteObjectTag{}, &commit_info, 4, 1};
+  Delta delta_delete2{Delta::DeleteObjectTag{}, &commit_info, 5, 2};
+  Delta delta_remove_label{Delta::RemoveLabelTag{}, secondary_label, &commit_info, 7, 4};
+  Delta delta_set_property{Delta::SetPropertyTag{}, secondary_property, PropertyValue(), &commit_info, 6, 4};
+  Delta delta_delete3{Delta::DeleteObjectTag{}, &commit_info, 8, 3};
+
+  VertexContainer expected_vertices;
+  auto [it_4, inserted1] = expected_vertices.emplace(PrimaryKey{PropertyValue{4}}, VertexData(&delta_delete1));
+  delta_delete1.prev.Set(&*it_4);
+  auto [it_5, inserted2] = expected_vertices.emplace(PrimaryKey{PropertyValue{5}}, VertexData(&delta_delete2));
+  delta_delete2.prev.Set(&*it_5);
+  auto [it_6, inserted3] = expected_vertices.emplace(PrimaryKey{PropertyValue{6}}, VertexData(&delta_delete3));
+  delta_delete3.prev.Set(&*it_6);
+  it_5->second.labels.push_back(secondary_label);
+  AddDeltaToDeltaChain(&*it_5, &delta_set_property);
+  AddDeltaToDeltaChain(&*it_5, &delta_remove_label);
+
+  AssertEqVertexContainer(splitted_data.vertices, expected_vertices);
+
+  // This is to ensure that the transaction that we have don't point to invalid
+  // object on the other shard
+  std::list<Delta> expected_deltas;
+  expected_deltas.emplace_back(Delta::DeleteObjectTag{}, &commit_info, 4, 1);
+  expected_deltas.emplace_back(Delta::DeleteObjectTag{}, &commit_info, 5, 2);
+  expected_deltas.emplace_back(Delta::SetPropertyTag{}, secondary_property, PropertyValue(), &commit_info, 6, 4);
+  expected_deltas.emplace_back(Delta::RemoveLabelTag{}, secondary_label, &commit_info, 7, 4);
+  expected_deltas.emplace_back(Delta::DeleteObjectTag{}, &commit_info, 8, 3);
+  AssertEqDeltaLists(splitted_data.transactions.begin()->second->deltas, expected_deltas);
+}
+
+TEST_F(ShardSplitTest, TestBasicSplitVerticesAndEdges) {
+  auto acc = storage.Access(GetNextHlc());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(1)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(2)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(3)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(4)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(5)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(6)}, {}).HasError());
+
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(1)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(5)}}, edge_type_id, Gid::FromUint(1))
+                   .HasError());
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(4)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(6)}}, edge_type_id, Gid::FromUint(2))
+                   .HasError());
+
+  auto current_hlc = GetNextHlc();
+  acc.Commit(current_hlc);
+
+  auto splitted_data = *storage.PerformSplit({PropertyValue(4)}, last_hlc, GetNextHlc());
+  EXPECT_EQ(splitted_data.vertices.size(), 3);
+  EXPECT_EQ(splitted_data.edges->size(), 2);
+  EXPECT_EQ(splitted_data.transactions.size(), 1);
+  EXPECT_EQ(splitted_data.label_indices.size(), 0);
+  EXPECT_EQ(splitted_data.label_property_indices.size(), 0);
+
+  CommitInfo commit_info{.start_or_commit_timestamp = current_hlc};
+  Delta delta_delete1{Delta::DeleteObjectTag{}, &commit_info, 12, 1};
+  Delta delta_delete2{Delta::DeleteObjectTag{}, &commit_info, 13, 1};
+  Delta delta_delete3{Delta::DeleteObjectTag{}, &commit_info, 14, 1};
+  Delta delta_add_in_edge1{Delta::RemoveInEdgeTag{},
+                           edge_type_id,
+                           VertexId{primary_label, {PropertyValue(1)}},
+                           EdgeRef{Gid::FromUint(1)},
+                           &commit_info,
+                           17,
+                           1};
+  Delta delta_add_out_edge2{Delta::RemoveOutEdgeTag{},
+                            edge_type_id,
+                            VertexId{primary_label, {PropertyValue(6)}},
+                            EdgeRef{Gid::FromUint(2)},
+                            &commit_info,
+                            19,
+                            1};
+  Delta delta_add_in_edge2{Delta::RemoveInEdgeTag{},
+                           edge_type_id,
+                           VertexId{primary_label, {PropertyValue(4)}},
+                           EdgeRef{Gid::FromUint(2)},
+                           &commit_info,
+                           20,
+                           1};
+  VertexContainer expected_vertices;
+  auto [vtx4, inserted4] = expected_vertices.emplace(PrimaryKey{PropertyValue{4}}, VertexData(&delta_delete1));
+  auto [vtx5, inserted5] = expected_vertices.emplace(PrimaryKey{PropertyValue{5}}, VertexData(&delta_delete2));
+  auto [vtx6, inserted6] = expected_vertices.emplace(PrimaryKey{PropertyValue{6}}, VertexData(&delta_delete3));
+  AddDeltaToDeltaChain(&*vtx4, &delta_add_out_edge2);
+  AddDeltaToDeltaChain(&*vtx5, &delta_add_in_edge1);
+  AddDeltaToDeltaChain(&*vtx6, &delta_add_in_edge2);
+
+  AssertEqVertexContainer(splitted_data.vertices, expected_vertices);
+}
+
+TEST_F(ShardSplitTest, TestBasicSplitBeforeCommit) {
+  auto acc = storage.Access(GetNextHlc());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(1)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(2)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(3)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(4)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(5)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(6)}, {}).HasError());
+
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(1)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(2)}}, edge_type_id, Gid::FromUint(0))
+                   .HasError());
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(1)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(5)}}, edge_type_id, Gid::FromUint(1))
+                   .HasError());
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(4)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(6)}}, edge_type_id, Gid::FromUint(2))
+                   .HasError());
+
+  auto splitted_data = *storage.PerformSplit({PropertyValue(4)}, last_hlc, GetNextHlc());
+  EXPECT_EQ(splitted_data.vertices.size(), 3);
+  EXPECT_EQ(splitted_data.edges->size(), 2);
+  EXPECT_EQ(splitted_data.transactions.size(), 1);
+  EXPECT_EQ(splitted_data.label_indices.size(), 0);
+  EXPECT_EQ(splitted_data.label_property_indices.size(), 0);
+}
+
+TEST_F(ShardSplitTest, TestBasicSplitWithCommitedAndOngoingTransactions) {
+  {
+    auto acc = storage.Access(GetNextHlc());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(1)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(2)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(3)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(4)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(5)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(6)}, {}).HasError());
+
+    acc.Commit(GetNextHlc());
+  }
+  auto acc = storage.Access(GetNextHlc());
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(1)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(2)}}, edge_type_id, Gid::FromUint(0))
+                   .HasError());
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(3)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(5)}}, edge_type_id, Gid::FromUint(1))
+                   .HasError());
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(4)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(6)}}, edge_type_id, Gid::FromUint(2))
+                   .HasError());
+
+  auto splitted_data = *storage.PerformSplit({PropertyValue(4)}, last_hlc, GetNextHlc());
+  EXPECT_EQ(splitted_data.vertices.size(), 3);
+  EXPECT_EQ(splitted_data.edges->size(), 2);
+  EXPECT_EQ(splitted_data.transactions.size(), 2);
+  EXPECT_EQ(splitted_data.label_indices.size(), 0);
+  EXPECT_EQ(splitted_data.label_property_indices.size(), 0);
+}
+
+TEST_F(ShardSplitTest, TestBasicSplitWithLabelIndex) {
+  auto acc = storage.Access(GetNextHlc());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(1)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(2)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(3)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(4)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(5)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(6)}, {}).HasError());
+  acc.Commit(GetNextHlc());
+  storage.CreateIndex(secondary_label);
+
+  auto splitted_data = *storage.PerformSplit({PropertyValue(4)}, last_hlc, GetNextHlc());
+
+  EXPECT_EQ(splitted_data.vertices.size(), 3);
+  EXPECT_EQ(splitted_data.edges->size(), 0);
+  EXPECT_EQ(splitted_data.transactions.size(), 1);
+  EXPECT_EQ(splitted_data.label_indices.size(), 1);
+  EXPECT_EQ(splitted_data.label_property_indices.size(), 0);
+}
+
+TEST_F(ShardSplitTest, TestBasicSplitWithLabelPropertyIndex) {
+  auto acc = storage.Access(GetNextHlc());
+  EXPECT_FALSE(
+      acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(1)}, {{secondary_property, PropertyValue(1)}})
+          .HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(2)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(3)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(4)}, {}).HasError());
+  EXPECT_FALSE(
+      acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(5)}, {{secondary_property, PropertyValue(21)}})
+          .HasError());
+  EXPECT_FALSE(
+      acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(6)}, {{secondary_property, PropertyValue(22)}})
+          .HasError());
+  acc.Commit(GetNextHlc());
+  storage.CreateIndex(secondary_label, secondary_property);
+
+  auto splitted_data = *storage.PerformSplit({PropertyValue(4)}, last_hlc, GetNextHlc());
+
+  EXPECT_EQ(splitted_data.vertices.size(), 3);
+  EXPECT_EQ(splitted_data.edges->size(), 0);
+  EXPECT_EQ(splitted_data.transactions.size(), 1);
+  EXPECT_EQ(splitted_data.label_indices.size(), 0);
+  EXPECT_EQ(splitted_data.label_property_indices.size(), 1);
+}
+
+TEST_F(ShardSplitTest, TestSplittingShardsWithGcDestroyOriginalShard) {
+  const auto split_value{4};
+  PrimaryKey splitted_value{{PropertyValue(4)}};
+  std::unique_ptr<Shard> splitted_shard;
+
+  {
+    Shard storage2{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector, GetNextHlc()};
+    auto acc = storage2.Access(GetNextHlc());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(1)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(2)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(3)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(4)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(5)}, {}).HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(6)}, {}).HasError());
+
+    EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(1)}},
+                                VertexId{primary_label, PrimaryKey{PropertyValue(2)}}, edge_type_id, Gid::FromUint(0))
+                     .HasError());
+    EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(3)}},
+                                VertexId{primary_label, PrimaryKey{PropertyValue(5)}}, edge_type_id, Gid::FromUint(1))
+                     .HasError());
+    EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(4)}},
+                                VertexId{primary_label, PrimaryKey{PropertyValue(6)}}, edge_type_id, Gid::FromUint(2))
+                     .HasError());
+    acc.Commit(GetNextHlc());
+
+    auto splitted_data = storage2.PerformSplit({PropertyValue(split_value)}, last_hlc, GetNextHlc());
+    EXPECT_EQ(splitted_data->vertices.size(), 3);
+    EXPECT_EQ(splitted_data->edges->size(), 2);
+    EXPECT_EQ(splitted_data->transactions.size(), 1);
+    EXPECT_EQ(splitted_data->label_indices.size(), 0);
+    EXPECT_EQ(splitted_data->label_property_indices.size(), 0);
+
+    // Create a new shard
+    splitted_shard = Shard::FromSplitData(std::move(*splitted_data));
+    // Call gc on old shard
+    storage2.CollectGarbage(GetNextHlc().coordinator_wall_clock);
+    // Destroy original
+  }
+
+  splitted_shard->CollectGarbage(GetNextHlc().coordinator_wall_clock);
+  AssertShardState(*splitted_shard, 4, 6);
+}
+
+TEST_F(ShardSplitTest, TestSplittingShardsWithGcDestroySplittedShard) {
+  PrimaryKey splitted_value{{PropertyValue(4)}};
+
+  auto acc = storage.Access(GetNextHlc());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(1)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(2)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(3)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(4)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(5)}, {}).HasError());
+  EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(6)}, {}).HasError());
+
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(1)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(2)}}, edge_type_id, Gid::FromUint(0))
+                   .HasError());
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(3)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(5)}}, edge_type_id, Gid::FromUint(1))
+                   .HasError());
+  EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(4)}},
+                              VertexId{primary_label, PrimaryKey{PropertyValue(6)}}, edge_type_id, Gid::FromUint(2))
+                   .HasError());
+  acc.Commit(GetNextHlc());
+
+  auto splitted_data = storage.PerformSplit({PropertyValue(4)}, last_hlc, GetNextHlc());
+  EXPECT_EQ(splitted_data->vertices.size(), 3);
+  EXPECT_EQ(splitted_data->edges->size(), 2);
+  EXPECT_EQ(splitted_data->transactions.size(), 1);
+  EXPECT_EQ(splitted_data->label_indices.size(), 0);
+  EXPECT_EQ(splitted_data->label_property_indices.size(), 0);
+
+  {
+    // Create a new shard
+    auto splitted_shard = Shard::FromSplitData(std::move(*splitted_data));
+    // Call gc on new shard
+    splitted_shard->CollectGarbage(GetNextHlc().coordinator_wall_clock);
+    // Destroy splitted shard
+  }
+
+  storage.CollectGarbage(GetNextHlc().coordinator_wall_clock);
+  AssertShardState(storage, 1, 3);
+}
+
+TEST_F(ShardSplitTest, TestBigSplit) {
+  int pk{0};
+  for (int64_t i{0}; i < 10'000; ++i) {
+    auto acc = storage.Access(GetNextHlc());
+    EXPECT_FALSE(
+        acc.CreateVertexAndValidate({secondary_label}, {PropertyValue(pk++)}, {{secondary_property, PropertyValue(i)}})
+            .HasError());
+    EXPECT_FALSE(acc.CreateVertexAndValidate({}, {PropertyValue(pk++)}, {}).HasError());
+
+    EXPECT_FALSE(acc.CreateEdge(VertexId{primary_label, PrimaryKey{PropertyValue(pk - 2)}},
+                                VertexId{primary_label, PrimaryKey{PropertyValue(pk - 1)}}, edge_type_id,
+                                Gid::FromUint(pk))
+                     .HasError());
+    acc.Commit(GetNextHlc());
+  }
+  storage.CreateIndex(secondary_label, secondary_property);
+
+  const auto split_value = pk / 2;
+  auto splitted_data = *storage.PerformSplit({PropertyValue(split_value)}, last_hlc, GetNextHlc());
+
+  EXPECT_EQ(splitted_data.vertices.size(), 10000);
+  EXPECT_EQ(splitted_data.edges->size(), 5000);
+  EXPECT_EQ(splitted_data.transactions.size(), 5000);
+  EXPECT_EQ(splitted_data.label_indices.size(), 0);
+  EXPECT_EQ(splitted_data.label_property_indices.size(), 1);
+
+  auto shard = Shard::FromSplitData(std::move(splitted_data));
+  AssertShardState(*shard, split_value, split_value * 2);
+}
+
+}  // namespace memgraph::storage::v3::tests

tests/unit/storage_v3_vertex_accessors.cpp

@@ -1,4 +1,4 @@
-// Copyright 2022 Memgraph Ltd.
+// Copyright 2023 Memgraph Ltd.
 //
 // Use of this software is governed by the Business Source License
 // included in the file licenses/BSL.txt; by using this file, you agree to be bound by the terms of the Business Source
@@ -57,9 +57,8 @@ class StorageV3Accessor : public ::testing::Test {
   const PropertyId primary_property{PropertyId::FromUint(2)};
   std::vector<storage::v3::SchemaProperty> schema_property_vector = {
       storage::v3::SchemaProperty{primary_property, common::SchemaType::INT}};
-  Shard storage{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector};
-
   coordinator::Hlc last_hlc{0, io::Time{}};
+  Shard storage{primary_label, min_pk, std::nullopt /*max_primary_key*/, schema_property_vector, last_hlc};
 };
 
 TEST_F(StorageV3Accessor, TestPrimaryLabel) {