feat(cluster): add migration finalization (#2507)

* feat(cluster): add migration finalization

src/server/cluster/cluster_config.h

@@ -23,7 +23,13 @@ using SlotId = uint16_t;
 using SlotSet = absl::flat_hash_set<SlotId>;
 
 // MigrationState constants are ordered in state changing order
-enum class MigrationState : uint8_t { C_NO_STATE, C_CONNECTING, C_FULL_SYNC, C_STABLE_SYNC };
+enum class MigrationState : uint8_t {
+  C_NO_STATE,
+  C_CONNECTING,
+  C_FULL_SYNC,
+  C_STABLE_SYNC,
+  C_FINISHED
+};
 
 class ClusterConfig {
  public:

src/server/cluster/cluster_family.cc

@@ -403,7 +403,9 @@ void ClusterFamily::DflyCluster(CmdArgList args, ConnectionContext* cntx) {
   } else if (sub_cmd == "START-SLOT-MIGRATION") {
     return DflyClusterStartSlotMigration(args, cntx);
   } else if (sub_cmd == "SLOT-MIGRATION-STATUS") {
-    return DflySlotMigrationStatus(args, cntx);
+    return DflyClusterSlotMigrationStatus(args, cntx);
+  } else if (sub_cmd == "SLOT-MIGRATION-FINALIZE") {
+    return DflyClusterMigrationFinalize(args, cntx);
   }
 
   return cntx->SendError(UnknownSubCmd(sub_cmd, "DFLYCLUSTER"), kSyntaxErrType);
@@ -531,6 +533,9 @@ void ClusterFamily::DflyClusterConfig(CmdArgList args, ConnectionContext* cntx)
     tracker.TrackOnThread();
   };
 
+  // TODO think about another place for it
+  RemoveFinishedIncomingMigrations();
+
   server_family_->service().proactor_pool().AwaitFiberOnAll(std::move(cb));
   DCHECK(tl_cluster_config != nullptr);
 
@@ -629,7 +634,7 @@ void ClusterFamily::DflyClusterStartSlotMigration(CmdArgList args, ConnectionCon
   }
   node->Start(cntx);
 
-  return cntx->SendOk();
+  return cntx->SendLong(node->GetSyncId());
 }
 
 static std::string_view state_to_str(MigrationState state) {
@@ -642,12 +647,14 @@ static std::string_view state_to_str(MigrationState state) {
       return "FULL_SYNC"sv;
     case MigrationState::C_STABLE_SYNC:
       return "STABLE_SYNC"sv;
+    case MigrationState::C_FINISHED:
+      return "FINISHED"sv;
   }
   DCHECK(false) << "Unknown State value " << static_cast<underlying_type_t<MigrationState>>(state);
   return "UNDEFINED_STATE"sv;
 }
 
-void ClusterFamily::DflySlotMigrationStatus(CmdArgList args, ConnectionContext* cntx) {
+void ClusterFamily::DflyClusterSlotMigrationStatus(CmdArgList args, ConnectionContext* cntx) {
   CmdArgParser parser(args);
   auto* rb = static_cast<RedisReplyBuilder*>(cntx->reply_builder());
 
@@ -689,6 +696,40 @@ void ClusterFamily::DflySlotMigrationStatus(CmdArgList args, ConnectionContext*
   return rb->SendSimpleString(state_to_str(MigrationState::C_NO_STATE));
 }
 
+void ClusterFamily::DflyClusterMigrationFinalize(CmdArgList args, ConnectionContext* cntx) {
+  CmdArgParser parser{args};
+  auto sync_id = parser.Next<uint32_t>();
+
+  if (auto err = parser.Error(); err) {
+    return cntx->SendError(err->MakeReply());
+  }
+
+  auto migration = GetOutgoingMigration(sync_id);
+  if (!migration)
+    return cntx->SendError(kIdNotFound);
+
+  if (migration->GetState() != MigrationState::C_STABLE_SYNC) {
+    return cntx->SendError("Migration process is not in STABLE_SYNC state");
+  }
+
+  shard_set->pool()->AwaitFiberOnAll([migration](auto*) {
+    if (const auto* shard = EngineShard::tlocal(); shard)
+      migration->Finalize(shard->shard_id());
+  });
+
+  // TODO do next after ACK
+  util::ThisFiber::SleepFor(500ms);
+
+  shard_set->pool()->AwaitFiberOnAll([migration](auto*) {
+    if (const auto* shard = EngineShard::tlocal(); shard)
+      migration->Cancel(shard->shard_id());
+  });
+
+  RemoveOutgoingMigration(sync_id);
+
+  return cntx->SendOk();
+}
+
 void ClusterFamily::DflyMigrate(CmdArgList args, ConnectionContext* cntx) {
   ToUpper(&args[0]);
   string_view sub_cmd = ArgS(args, 0);
@@ -718,6 +759,19 @@ ClusterSlotMigration* ClusterFamily::AddMigration(std::string host_ip, uint16_t
       .get();
 }
 
+void ClusterFamily::RemoveFinishedIncomingMigrations() {
+  lock_guard lk(migration_mu_);
+  auto removed_items_it =
+      std::remove_if(incoming_migrations_jobs_.begin(), incoming_migrations_jobs_.end(),
+                     [](const auto& m) { return m->GetState() == MigrationState::C_FINISHED; });
+  incoming_migrations_jobs_.erase(removed_items_it, incoming_migrations_jobs_.end());
+}
+
+void ClusterFamily::RemoveOutgoingMigration(uint32_t sync_id) {
+  lock_guard lk(migration_mu_);
+  outgoing_migration_jobs_.erase(sync_id);
+}
+
 void ClusterFamily::MigrationConf(CmdArgList args, ConnectionContext* cntx) {
   VLOG(1) << "Create slot migration config";
   CmdArgParser parser{args};
@@ -793,6 +847,7 @@ void ClusterFamily::DflyMigrateFlow(CmdArgList args, ConnectionContext* cntx) {
     return cntx->SendError(kIdNotFound);
 
   cntx->conn()->Migrate(shard_set->pool()->at(shard_id));
+  server_family_->journal()->StartInThread();
 
   cntx->SendOk();
 
@@ -825,7 +880,6 @@ void ClusterFamily::DflyMigrateFullSyncCut(CmdArgList args, ConnectionContext* c
 
   (*migration_it)->SetStableSyncForFlow(shard_id);
   if ((*migration_it)->GetState() == MigrationState::C_STABLE_SYNC) {
-    (*migration_it)->Stop();
     LOG(INFO) << "STABLE-SYNC state is set for sync_id " << sync_id;
   }
 

src/server/cluster/cluster_family.h

@@ -51,7 +51,8 @@ class ClusterFamily {
 
  private:  // Slots migration section
   void DflyClusterStartSlotMigration(CmdArgList args, ConnectionContext* cntx);
-  void DflySlotMigrationStatus(CmdArgList args, ConnectionContext* cntx);
+  void DflyClusterSlotMigrationStatus(CmdArgList args, ConnectionContext* cntx);
+  void DflyClusterMigrationFinalize(CmdArgList args, ConnectionContext* cntx);
 
   // DFLYMIGRATE is internal command defines several steps in slots migrations process
   void DflyMigrate(CmdArgList args, ConnectionContext* cntx);
@@ -74,6 +75,9 @@ class ClusterFamily {
   ClusterSlotMigration* AddMigration(std::string host_ip, uint16_t port,
                                      std::vector<ClusterConfig::SlotRange> slots);
 
+  void RemoveFinishedIncomingMigrations();
+  void RemoveOutgoingMigration(uint32_t sync_id);
+
   // store info about migration and create unique session id
   uint32_t CreateOutgoingMigration(ConnectionContext* cntx, uint16_t port,
                                    std::vector<ClusterConfig::SlotRange> slots);

src/server/cluster/cluster_shard_migration.cc

@@ -74,10 +74,14 @@ void ClusterShardMigration::FullSyncShardFb(Context* cntx) {
 
     TouchIoTime();
 
-    if (!tx_data->is_ping) {
+    if (tx_data->opcode == journal::Op::FIN) {
-      ExecuteTxWithNoShardSync(std::move(*tx_data), cntx);
+      VLOG(2) << "Flow " << source_shard_id_ << " is finalized";
-    } else {
+      is_finalized_ = true;
+      break;
+    } else if (tx_data->opcode == journal::Op::PING) {
       // TODO check about ping logic
+    } else {
+      ExecuteTxWithNoShardSync(std::move(*tx_data), cntx);
     }
   }
 }

src/server/cluster/cluster_shard_migration.h

@@ -31,9 +31,14 @@ class ClusterShardMigration : public ProtocolClient {
     return is_stable_sync_.load();
   }
 
+  bool IsFinalized() {
+    return is_finalized_;
+  }
+
+  void JoinFlow();
+
  private:
   void FullSyncShardFb(Context* cntx);
-  void JoinFlow();
 
   void ExecuteTx(TransactionData&& tx_data, bool inserted_by_me, Context* cntx);
   void ExecuteTxWithNoShardSync(TransactionData&& tx_data, Context* cntx);
@@ -45,6 +50,7 @@ class ClusterShardMigration : public ProtocolClient {
   std::unique_ptr<JournalExecutor> executor_;
   Fiber sync_fb_;
   std::atomic_bool is_stable_sync_ = false;
+  bool is_finalized_ = false;
 };
 
 }  // namespace dfly

src/server/cluster/cluster_slot_migration.cc

@@ -113,6 +113,11 @@ void ClusterSlotMigration::SetStableSyncForFlow(uint32_t flow) {
   }
 }
 
+bool ClusterSlotMigration::IsFinalized() const {
+  return std::all_of(shard_flows_.begin(), shard_flows_.end(),
+                     [](const auto& el) { return el->IsFinalized(); });
+}
+
 void ClusterSlotMigration::Stop() {
   for (auto& flow : shard_flows_) {
     flow->Cancel();
@@ -129,6 +134,10 @@ void ClusterSlotMigration::MainMigrationFb() {
     LOG(WARNING) << "Error syncing with " << server().Description() << " " << ec << " "
                  << ec.message();
   }
+
+  if (IsFinalized()) {
+    state_ = MigrationState::C_FINISHED;
+  }
 }
 
 std::error_code ClusterSlotMigration::InitiateSlotsMigration() {
@@ -137,6 +146,13 @@ std::error_code ClusterSlotMigration::InitiateSlotsMigration() {
     shard_flows_[i].reset(new ClusterShardMigration(server(), i, sync_id_, &service_));
   }
 
+  absl::Cleanup cleanup = [this]() {
+    // We do the following operations regardless of outcome.
+    for (auto& flow : shard_flows_) {
+      flow->JoinFlow();
+    }
+  };
+
   // Switch to new error handler that closes flow sockets.
   auto err_handler = [this](const auto& ge) mutable {
     // Make sure the flows are not in a state transition

src/server/cluster/cluster_slot_migration.h

@@ -36,6 +36,7 @@ class ClusterSlotMigration : ProtocolClient {
   }
 
   void SetStableSyncForFlow(uint32_t flow);
+
   void Stop();
 
  private:
@@ -45,6 +46,9 @@ class ClusterSlotMigration : ProtocolClient {
   // Creates flows, one per shard on the source node and manage migration process
   std::error_code InitiateSlotsMigration();
 
+  // may be called after we finish all flows
+  bool IsFinalized() const;
+
  private:
   Service& service_;
   Mutex flows_op_mu_;

src/server/cluster/outgoing_slot_migration.cc

@@ -18,10 +18,14 @@ class OutgoingMigration::SliceSlotMigration {
     state_ = MigrationState::C_FULL_SYNC;
   }
 
-  ~SliceSlotMigration() {
+  void Cancel() {
     streamer_.Cancel();
   }
 
+  void Finalize() {
+    streamer_.SendFinalize();
+  }
+
   MigrationState GetState() const {
     return state_ == MigrationState::C_FULL_SYNC && streamer_.IsSnapshotFinished()
                ? MigrationState::C_STABLE_SYNC
@@ -56,8 +60,20 @@ void OutgoingMigration::StartFlow(DbSlice* slice, uint32_t sync_id, journal::Jou
       std::make_unique<SliceSlotMigration>(slice, std::move(sset), sync_id, journal, &cntx_, dest);
 }
 
+void OutgoingMigration::Finalize(uint32_t shard_id) {
+  slot_migrations_[shard_id]->Finalize();
+}
+
+void OutgoingMigration::Cancel(uint32_t shard_id) {
+  slot_migrations_[shard_id]->Cancel();
+}
+
 MigrationState OutgoingMigration::GetState() const {
   std::lock_guard lck(flows_mu_);
+  return GetStateImpl();
+}
+
+MigrationState OutgoingMigration::GetStateImpl() const {
   MigrationState min_state = MigrationState::C_STABLE_SYNC;
   for (const auto& slot_migration : slot_migrations_) {
     if (slot_migration)

src/server/cluster/outgoing_slot_migration.h

@@ -15,7 +15,7 @@ class Journal;
 
 class DbSlice;
 
-// Whole slots migration process information
+// Whole outgoing slots migration manager
 class OutgoingMigration {
  public:
   OutgoingMigration() = default;
@@ -25,6 +25,9 @@ class OutgoingMigration {
 
   void StartFlow(DbSlice* slice, uint32_t sync_id, journal::Journal* journal, io::Sink* dest);
 
+  void Finalize(uint32_t shard_id);
+  void Cancel(uint32_t shard_id);
+
   MigrationState GetState() const;
 
   const std::string& GetHostIp() const {
@@ -35,6 +38,7 @@ class OutgoingMigration {
   };
 
  private:
+  MigrationState GetStateImpl() const;
   // SliceSlotMigration manages state and data transfering for the corresponding shard
   class SliceSlotMigration;
 

src/server/journal/serializer.cc

@@ -195,7 +195,7 @@ io::Result<journal::ParsedEntry> JournalReader::ReadEntry() {
   entry.dbid = dbid_;
   entry.opcode = opcode;
 
-  if (opcode == journal::Op::PING) {
+  if (opcode == journal::Op::PING || opcode == journal::Op::FIN) {
     return entry;
   }
 

src/server/journal/streamer.cc

@@ -92,6 +92,15 @@ void RestoreStreamer::Start(io::Sink* dest) {
   });
 }
 
+void RestoreStreamer::SendFinalize() {
+  VLOG(2) << "DFLYMIGRATE FINALIZE for " << sync_id_ << " : " << db_slice_->shard_id();
+  journal::Entry entry(journal::Op::FIN, 0 /*db_id*/, 0 /*slot_id*/);
+
+  JournalWriter writer{this};
+  writer.Write(entry);
+  NotifyWritten(true);
+}
+
 void RestoreStreamer::Cancel() {
   fiber_cancellation_.Cancel();
   snapshot_fb_.JoinIfNeeded();

src/server/journal/streamer.h

@@ -60,6 +60,8 @@ class RestoreStreamer : public JournalStreamer {
   // Cancel() must be called if Start() is called
   void Cancel() override;
 
+  void SendFinalize();
+
   bool IsSnapshotFinished() const {
     return snapshot_finished_;
   }

src/server/journal/tx_executor.cc

@@ -49,10 +49,12 @@ void MultiShardExecution::CancelAllBlockingEntities() {
 
 bool TransactionData::AddEntry(journal::ParsedEntry&& entry) {
   ++journal_rec_count;
+  opcode = entry.opcode;
 
   switch (entry.opcode) {
     case journal::Op::PING:
-      is_ping = true;
+      return true;
+    case journal::Op::FIN:
       return true;
     case journal::Op::EXPIRED:
     case journal::Op::COMMAND:
@@ -112,7 +114,7 @@ std::optional<TransactionData> TransactionReader::NextTxData(JournalReader* read
     // Check if journal command can be executed right away.
     // Expiration checks lock on master, so it never conflicts with running multi transactions.
     if (res->opcode == journal::Op::EXPIRED || res->opcode == journal::Op::COMMAND ||
-        res->opcode == journal::Op::PING)
+        res->opcode == journal::Op::PING || res->opcode == journal::Op::FIN)
       return TransactionData::FromSingle(std::move(res.value()));
 
     // Otherwise, continue building multi command.

src/server/journal/tx_executor.h

@@ -51,7 +51,7 @@ struct TransactionData {
   uint32_t shard_cnt{0};
   absl::InlinedVector<journal::ParsedEntry::CmdData, 1> commands{0};
   uint32_t journal_rec_count{0};  // Count number of source entries to check offset.
-  bool is_ping = false;           // For Op::PING entries.
+  journal::Op opcode = journal::Op::NOOP;
 };
 
 // Utility for reading TransactionData from a journal reader.

src/server/journal/types.h

@@ -22,6 +22,7 @@ enum class Op : uint8_t {
   MULTI_COMMAND = 11,
   EXEC = 12,
   PING = 13,
+  FIN = 14
 };
 
 struct EntryBase {

src/server/replica.cc

@@ -803,7 +803,7 @@ void DflyShardReplica::StableSyncDflyReadFb(Context* cntx) {
 
     last_io_time_ = Proactor()->GetMonotonicTimeNs();
 
-    if (!tx_data->is_ping) {
+    if (tx_data->opcode != journal::Op::PING) {
       if (use_multi_shard_exe_sync_) {
         InsertTxDataToShardResource(std::move(*tx_data));
       } else {

src/server/transaction.cc

@@ -307,8 +307,10 @@ void Transaction::InitByKeys(const KeyIndex& key_index) {
     unique_shard_cnt_ = 1;
     if (is_stub)  // stub transactions don't migrate
       DCHECK_EQ(unique_shard_id_, Shard(kv_args_.front(), shard_set->size()));
-    else
+    else {
+      unique_slot_checker_.Add(kv_args_.front());
       unique_shard_id_ = Shard(kv_args_.front(), shard_set->size());
+    }
 
     // Multi transactions that execute commands on their own (not stubs) can't shrink the backing
     // array, as it still might be read by leftover callbacks.

tests/dragonfly/cluster_test.py

@@ -787,7 +787,7 @@ async def test_cluster_slot_migration(df_local_factory: DflyInstanceFactory):
     res = await c_nodes_admin[1].execute_command(
         "DFLYCLUSTER", "START-SLOT-MIGRATION", "127.0.0.1", str(nodes[0].admin_port), "5200", "5259"
     )
-    assert "OK" == res
+    assert 1 == res
 
     while (
         await c_nodes_admin[1].execute_command(
@@ -862,12 +862,9 @@ async def test_cluster_data_migration(df_local_factory: DflyInstanceFactory):
         c_nodes_admin,
     )
 
-    assert await c_nodes[0].set("KEY0", "value")
-    assert await c_nodes[0].set("KEY1", "value")
     assert await c_nodes[1].set("KEY2", "value")
     assert await c_nodes[1].set("KEY3", "value")
-    assert await c_nodes[0].set("KEY4", "value")
+
-    assert await c_nodes[0].set("KEY5", "value")
     assert await c_nodes[1].set("KEY6", "value")
     assert await c_nodes[1].set("KEY7", "value")
     assert await c_nodes[0].set("KEY8", "value")
@@ -882,12 +879,14 @@ async def test_cluster_data_migration(df_local_factory: DflyInstanceFactory):
     assert await c_nodes[0].set("KEY17", "value")
     assert await c_nodes[1].set("KEY18", "value")
     assert await c_nodes[1].set("KEY19", "value")
-    assert await c_nodes[0].execute_command("DBSIZE") == 10
 
     res = await c_nodes_admin[1].execute_command(
         "DFLYCLUSTER", "START-SLOT-MIGRATION", "127.0.0.1", str(nodes[0].admin_port), "3000", "9000"
     )
-    assert "OK" == res
+    assert 1 == res
+
+    assert await c_nodes[0].set("KEY0", "value")
+    assert await c_nodes[0].set("KEY1", "value")
 
     while (
         await c_nodes_admin[1].execute_command(
@@ -897,6 +896,26 @@ async def test_cluster_data_migration(df_local_factory: DflyInstanceFactory):
     ):
         await asyncio.sleep(0.05)
 
+    assert await c_nodes[0].set("KEY4", "value")
+    assert await c_nodes[0].set("KEY5", "value")
+    assert await c_nodes[0].execute_command("DBSIZE") == 10
+
+    # TODO remove when we add slot blocking
+    await asyncio.sleep(0.5)
+
+    res = await c_nodes_admin[0].execute_command("DFLYCLUSTER", "SLOT-MIGRATION-FINALIZE", "1")
+    assert "OK" == res
+
+    await asyncio.sleep(0.5)
+
+    while (
+        await c_nodes_admin[1].execute_command(
+            "DFLYCLUSTER", "SLOT-MIGRATION-STATUS", "127.0.0.1", str(nodes[0].admin_port)
+        )
+        != "FINISHED"
+    ):
+        await asyncio.sleep(0.05)
+
     await push_config(
         config.replace("LAST_SLOT_CUTOFF", "2999").replace("NEXT_SLOT_CUTOFF", "3000"),
         c_nodes_admin,