Fix race between table merge and borrowed gc compaction. Fixes #3154. (#3164)

Author: snaury

ydb/core/tx/datashard/datashard_split_dst.cpp

@@ -120,13 +120,11 @@ class TDataShard::TTxInitSplitMergeDestination : public NTabletFlatExecutor::TTr
 class TDataShard::TTxSplitTransferSnapshot : public NTabletFlatExecutor::TTransactionBase<TDataShard> {
 private:
     TEvDataShard::TEvSplitTransferSnapshot::TPtr Ev;
-    bool LastSnapshotReceived;
 
 public:
     TTxSplitTransferSnapshot(TDataShard* ds, TEvDataShard::TEvSplitTransferSnapshot::TPtr& ev)
         : NTabletFlatExecutor::TTransactionBase<TDataShard>(ds)
         , Ev(ev)
-        , LastSnapshotReceived(false)
     {}
 
     TTxType GetTxType() const override { return TXTYPE_SPLIT_TRANSFER_SNAPSHOT; }
@@ -257,8 +255,6 @@ class TDataShard::TTxSplitTransferSnapshot : public NTabletFlatExecutor::TTransa
         }
 
         if (Self->ReceiveSnapshotsFrom.empty()) {
-            LastSnapshotReceived = true;
-
             const auto minVersion = mvcc ? Self->GetSnapshotManager().GetLowWatermark()
                                          : Self->GetSnapshotManager().GetMinWriteVersion();
 
@@ -295,6 +291,12 @@ class TDataShard::TTxSplitTransferSnapshot : public NTabletFlatExecutor::TTransa
             // Note: we persist Ready, but keep current state in memory until Complete
             Self->SetPersistState(TShardState::Ready, txc);
             Self->State = TShardState::SplitDstReceivingSnapshot;
+
+            // Schedule a new transaction that will move shard to the Ready state
+            // and finish initialization. This new transaction is guaranteed to
+            // wait until async LoanTable above is complete and new parts are
+            // fully merged into the table.
+            Self->Execute(new TTxLastSnapshotReceived(Self));
         }
 
         return true;
@@ -307,39 +309,52 @@ class TDataShard::TTxSplitTransferSnapshot : public NTabletFlatExecutor::TTransa
         LOG_DEBUG_S(ctx, NKikimrServices::TX_DATASHARD, Self->TabletID() << " ack snapshot OpId " << opId);
 
         ctx.Send(ackTo, new TEvDataShard::TEvSplitTransferSnapshotAck(opId, Self->TabletID()));
+    }
 
-        // Note: we skip init in an unlikely event of state resetting between Execute and Complete
-        if (LastSnapshotReceived && Self->State == TShardState::SplitDstReceivingSnapshot) {
-            // We have received all the data, finish shard initialization
-            // Note: previously we used TxInit, however received system tables
-            // have been empty for years now, and since pipes are still open we
-            // may receive requests between TxInit loading the Ready state and
-            // its Complete method initializing everything properly. Instead
-            // necessary steps are repeated here.
-            Self->State = TShardState::Ready;
-
-            // We are already in StateWork, but we need to repeat many steps now that we are Ready
-            Self->SwitchToWork(ctx);
-
-            // We can send the registration request now that we are ready
-            Self->SendRegistrationRequestTimeCast(ctx);
-
-            // Initialize snapshot expiration queue with current context time
-            Self->GetSnapshotManager().InitExpireQueue(ctx.Now());
-            if (Self->GetSnapshotManager().HasExpiringSnapshots()) {
-                Self->PlanCleanup(ctx);
-            }
+    class TTxLastSnapshotReceived : public NTabletFlatExecutor::TTransactionBase<TDataShard> {
+    public:
+        TTxLastSnapshotReceived(TDataShard* self)
+            : TTransactionBase(self)
+        {}
 
-            // Initialize change senders
-            Self->KillChangeSender(ctx);
-            Self->CreateChangeSender(ctx);
-            Self->MaybeActivateChangeSender(ctx);
-            Self->EmitHeartbeats();
+        bool Execute(TTransactionContext&, const TActorContext&) override {
+            return true;
+        }
 
-            // Switch mvcc state if needed
-            Self->CheckMvccStateChangeCanStart(ctx);
+        void Complete(const TActorContext& ctx) override {
+            // Note: we skip init in an unlikely event of state resetting before reaching Complete
+            if (Self->State == TShardState::SplitDstReceivingSnapshot) {
+                // We have received all the data, finish shard initialization
+                // Note: previously we used TxInit, however received system tables
+                // have been empty for years now, and since pipes are still open we
+                // may receive requests between TxInit loading the Ready state and
+                // its Complete method initializing everything properly. Instead
+                // necessary steps are repeated here.
+                Self->State = TShardState::Ready;
+
+                // We are already in StateWork, but we need to repeat many steps now that we are Ready
+                Self->SwitchToWork(ctx);
+
+                // We can send the registration request now that we are ready
+                Self->SendRegistrationRequestTimeCast(ctx);
+
+                // Initialize snapshot expiration queue with current context time
+                Self->GetSnapshotManager().InitExpireQueue(ctx.Now());
+                if (Self->GetSnapshotManager().HasExpiringSnapshots()) {
+                    Self->PlanCleanup(ctx);
+                }
+
+                // Initialize change senders
+                Self->KillChangeSender(ctx);
+                Self->CreateChangeSender(ctx);
+                Self->MaybeActivateChangeSender(ctx);
+                Self->EmitHeartbeats();
+
+                // Switch mvcc state if needed
+                Self->CheckMvccStateChangeCanStart(ctx);
+            }
         }
-    }
+    };
 };
 
 class TDataShard::TTxSplitReplicationSourceOffsets : public NTabletFlatExecutor::TTransactionBase<TDataShard> {

ydb/core/tx/datashard/datashard_ut_common_kqp.h

@@ -11,6 +11,9 @@ namespace NKqpHelpers {
     using TEvExecuteDataQueryRequest = NKikimr::NGRpcService::TGrpcRequestOperationCall<Ydb::Table::ExecuteDataQueryRequest,
         Ydb::Table::ExecuteDataQueryResponse>;
 
+    using TEvExecuteSchemeQueryRequest = NKikimr::NGRpcService::TGrpcRequestOperationCall<Ydb::Table::ExecuteSchemeQueryRequest,
+        Ydb::Table::ExecuteSchemeQueryResponse>;
+
     using TEvCreateSessionRequest = NKikimr::NGRpcService::TGrpcRequestOperationCall<Ydb::Table::CreateSessionRequest,
         Ydb::Table::CreateSessionResponse>;
 
@@ -224,6 +227,31 @@ namespace NKqpHelpers {
         return FormatResult(result);
     }
 
+    inline Ydb::Table::ExecuteSchemeQueryRequest MakeSchemeRequestRPC(
+        const TString& sql, const TString& sessionId)
+    {
+        Ydb::Table::ExecuteSchemeQueryRequest request;
+        request.set_session_id(sessionId);
+        request.set_yql_text(sql);
+        return request;
+    }
+
+    inline NThreading::TFuture<Ydb::Table::ExecuteSchemeQueryResponse> SendRequest(
+        TTestActorRuntime& runtime, Ydb::Table::ExecuteSchemeQueryRequest&& request, const TString& database = {})
+    {
+        return NRpcService::DoLocalRpc<TEvExecuteSchemeQueryRequest>(
+            std::move(request), database, /* token */ "", runtime.GetActorSystem(0));
+    }
+
+    inline TString KqpSchemeExec(TTestActorRuntime& runtime, const TString& query) {
+        TString sessionId = CreateSessionRPC(runtime);
+        auto response = AwaitResponse(runtime, SendRequest(runtime, MakeSchemeRequestRPC(query, sessionId)));
+        if (response.operation().status() != Ydb::StatusIds::SUCCESS) {
+            return TStringBuilder() << "ERROR: " << response.operation().status();
+        }
+        return "SUCCESS";
+    }
+
 } // namespace NKqpHelpers
 } // namespace NDataShard
 } // namespace NKikimr

ydb/core/tx/datashard/datashard_ut_snapshot.cpp

@@ -4540,6 +4540,130 @@ Y_UNIT_TEST_SUITE(DataShardSnapshots) {
         }
     }
 
+    void CompactBorrowed(TTestActorRuntime& runtime, ui64 shardId, const TTableId& tableId) {
+        auto msg = MakeHolder<TEvDataShard::TEvCompactBorrowed>(tableId.PathId);
+        auto sender = runtime.AllocateEdgeActor();
+        runtime.SendToPipe(shardId, sender, msg.Release(), 0, GetPipeConfigWithRetries());
+        runtime.GrabEdgeEventRethrow<TEvDataShard::TEvCompactBorrowedResult>(sender);
+    }
+
+    Y_UNIT_TEST(PostMergeNotCompactedTooEarly) {
+        TPortManager pm;
+        TServerSettings serverSettings(pm.GetPort(2134));
+        serverSettings.SetDomainName("Root")
+            .SetUseRealThreads(false)
+            .SetDomainPlanResolution(100);
+
+        Tests::TServer::TPtr server = new TServer(serverSettings);
+        auto &runtime = *server->GetRuntime();
+        auto sender = runtime.AllocateEdgeActor();
+
+        runtime.SetLogPriority(NKikimrServices::TX_DATASHARD, NLog::PRI_TRACE);
+
+        InitRoot(server, sender);
+
+        TDisableDataShardLogBatching disableDataShardLogBatching;
+
+        KqpSchemeExec(runtime, R"(
+            CREATE TABLE `/Root/table` (key int, value bytes, PRIMARY KEY (key))
+            WITH (AUTO_PARTITIONING_MIN_PARTITIONS_COUNT = 1,
+                  PARTITION_AT_KEYS = (5));
+        )");
+
+        const auto shards = GetTableShards(server, sender, "/Root/table");
+        UNIT_ASSERT_VALUES_EQUAL(shards.size(), 2u);
+        const auto tableId = ResolveTableId(server, sender, "/Root/table");
+
+        for (int i = 0; i < 20; ++i) {
+            Cerr << "... upserting key " << i << Endl;
+            auto query = Sprintf(R"(
+                UPSERT INTO `/Root/table` (key, value) VALUES (%d, '%s');
+            )", i, TString(128 * 1024, 'x').c_str());
+            ExecSQL(server, sender, query);
+            if (i >= 5) {
+                Cerr << "... compacting shard " << shards.at(1) << Endl;
+                CompactTable(runtime, shards.at(1), tableId, false);
+            } else if (i == 4) {
+                Cerr << "... compacting shard " << shards.at(0) << Endl;
+                CompactTable(runtime, shards.at(0), tableId, false);
+            }
+        }
+
+        // Read (and snapshot) current data, so it doesn't go away on compaction
+        UNIT_ASSERT_VALUES_EQUAL(
+            KqpSimpleExec(runtime, "SELECT COUNT(*) FROM `/Root/table`;"),
+            "{ items { uint64_value: 20 } }");
+
+        // Delete all the data in shard 0, this is small and will stay in memtable
+        // But when borrowed dst compaction will have pressure to compact it all
+        ExecSQL(server, sender, "DELETE FROM `/Root/table` WHERE key < 5");
+
+        std::vector<TEvDataShard::TEvSplitTransferSnapshot::TPtr> snapshots;
+        auto captureSnapshots = runtime.AddObserver<TEvDataShard::TEvSplitTransferSnapshot>(
+            [&](TEvDataShard::TEvSplitTransferSnapshot::TPtr& ev) {
+                auto* msg = ev->Get();
+                Cerr << "... captured snapshot from " << msg->Record.GetSrcTabletId() << Endl;
+                snapshots.emplace_back(ev.Release());
+            });
+
+        Cerr << "... merging table" << Endl;
+        SetSplitMergePartCountLimit(server->GetRuntime(), -1);
+        ui64 txId = AsyncMergeTable(server, sender, "/Root/table", shards);
+        Cerr << "... started merge " << txId << Endl;
+        WaitFor(runtime, [&]{ return snapshots.size() >= 2; }, "both src tablet snapshots");
+
+        std::vector<TEvBlobStorage::TEvGet::TPtr> gets;
+        auto captureGets = runtime.AddObserver<TEvBlobStorage::TEvGet>(
+            [&](TEvBlobStorage::TEvGet::TPtr& ev) {
+                auto* msg = ev->Get();
+                if (msg->Queries[0].Id.TabletID() == shards.at(1)) {
+                    Cerr << "... blocking blob get of " << msg->Queries[0].Id << Endl;
+                    gets.emplace_back(ev.Release());
+                }
+            });
+
+        // Release snapshot for shard 0 then shard 1
+        captureSnapshots.Remove();
+        Cerr << "... unlocking snapshots from tablet " << shards.at(0) << Endl;
+        for (auto& ev : snapshots) {
+            if (ev && ev->Get()->Record.GetSrcTabletId() == shards.at(0)) {
+                runtime.Send(ev.Release(), 0, true);
+            }
+        }
+        Cerr << "... unblocking snapshots from tablet " << shards.at(1) << Endl;
+        for (auto& ev : snapshots) {
+            if (ev && ev->Get()->Record.GetSrcTabletId() == shards.at(1)) {
+                runtime.Send(ev.Release(), 0, true);
+            }
+        }
+
+        // Let it commit above snapshots and incorrectly compact after the first one is loaded and merged
+        runtime.SimulateSleep(TDuration::Seconds(1));
+        UNIT_ASSERT(gets.size() > 0);
+
+        Cerr << "... unblocking blob gets" << Endl;
+        captureGets.Remove();
+        for (auto& ev : gets) {
+            runtime.Send(ev.Release(), 0, true);
+        }
+
+        // Let it finish loading the second snapshot
+        runtime.SimulateSleep(TDuration::Seconds(1));
+
+        // Wait for merge to complete and start a borrowed compaction
+        // When bug is present it will cause newly compacted to part to have epoch larger than previously compacted
+        WaitTxNotification(server, sender, txId);
+        const auto merged = GetTableShards(server, sender, "/Root/table");
+        UNIT_ASSERT_VALUES_EQUAL(merged.size(), 1u);
+        Cerr << "... compacting borrowed parts in shard " << merged.at(0) << Endl;
+        CompactBorrowed(runtime, merged.at(0), tableId);
+
+        // Validate we have an expected number of rows
+        UNIT_ASSERT_VALUES_EQUAL(
+            KqpSimpleExec(runtime, "SELECT COUNT(*) FROM `/Root/table`;"),
+            "{ items { uint64_value: 15 } }");
+    }
+
 }
 
 } // namespace NKikimr