Fixed errors with transactions in the PQ tablet (#12884) (#12901)

Author: Alek5andr-Kotov

ydb/core/persqueue/blob.cpp

@@ -660,6 +660,7 @@ ui32 THead::GetCount() const
 
     //how much offsets before last batch and how much offsets in last batch
     Y_ABORT_UNLESS(Batches.front().GetOffset() == Offset);
+
     return Batches.back().GetOffset() - Offset + Batches.back().GetCount();
 }
 
@@ -933,16 +934,23 @@ auto TPartitionedBlob::Add(TClientBlob&& blob) -> std::optional<TFormedBlobInfo>
 
 auto TPartitionedBlob::Add(const TKey& oldKey, ui32 size) -> std::optional<TFormedBlobInfo>
 {
+    if (HeadSize + BlobsSize == 0) { //if nothing to compact at all
+        NeedCompactHead = false;
+    }
+
     std::optional<TFormedBlobInfo> res;
     if (NeedCompactHead) {
         NeedCompactHead = false;
-        GlueNewHead = false;
         res = CreateFormedBlob(0, false);
+
+        StartOffset = NewHead.Offset + NewHead.GetCount();
+        NewHead.Clear();
+        NewHead.Offset = StartOffset;
     }
 
     TKey newKey(TKeyPrefix::TypeData,
                 Partition,
-                NewHead.Offset + oldKey.GetOffset(),
+                StartOffset,
                 oldKey.GetPartNo(),
                 oldKey.GetCount(),
                 oldKey.GetInternalPartsCount(),

ydb/core/persqueue/events/internal.h

@@ -819,6 +819,7 @@ struct TEvPQ {
         ui64 TxId;
         TVector<NKikimrPQ::TPartitionOperation> Operations;
         TActorId SupportivePartitionActor;
+        bool ForceFalse = false;
     };
 
     struct TEvTxCalcPredicateResult : public TEventLocal<TEvTxCalcPredicateResult, EvTxCalcPredicateResult> {

ydb/core/persqueue/partition.cpp

@@ -251,7 +251,7 @@ ui64 TPartition::UserDataSize() const {
         return 0;
     }
 
-    // We assume that DataKyesBody contains an up-to-date set of blobs, their relevance is
+    // We assume that DataKeysBody contains an up-to-date set of blobs, their relevance is
     // maintained by the background process. However, the last block may contain several irrelevant
     // messages. Because of them, we throw out the size of the entire blob.
     auto size = Size();
@@ -938,15 +938,13 @@ void TPartition::Handle(TEvPQ::TEvProposePartitionConfig::TPtr& ev, const TActor
 
 void TPartition::HandleOnInit(TEvPQ::TEvTxCalcPredicate::TPtr& ev, const TActorContext&)
 {
+    PQ_LOG_D("HandleOnInit TEvPQ::TEvTxCalcPredicate");
+
     PendingEvents.emplace_back(ev->ReleaseBase().Release());
 }
 
 void TPartition::HandleOnInit(TEvPQ::TEvTxCommit::TPtr& ev, const TActorContext&)
 {
-    PQ_LOG_D("HandleOnInit TEvPQ::TEvTxCommit" <<
-             " Step " << ev->Get()->Step <<
-             ", TxId " << ev->Get()->TxId);
-
     PendingEvents.emplace_back(ev->ReleaseBase().Release());
 }
 
@@ -957,9 +955,32 @@ void TPartition::HandleOnInit(TEvPQ::TEvTxRollback::TPtr& ev, const TActorContex
 
 void TPartition::HandleOnInit(TEvPQ::TEvProposePartitionConfig::TPtr& ev, const TActorContext&)
 {
-    PQ_LOG_D("HandleOnInit TEvPQ::TEvProposePartitionConfig" <<
-             " Step " << ev->Get()->Step <<
-             ", TxId " << ev->Get()->TxId);
+    PendingEvents.emplace_back(ev->ReleaseBase().Release());
+}
+
+void TPartition::HandleOnInit(TEvPQ::TEvGetWriteInfoRequest::TPtr& ev, const TActorContext&)
+{
+    PQ_LOG_D("HandleOnInit TEvPQ::TEvGetWriteInfoRequest");
+
+    Y_ABORT_UNLESS(IsSupportive());
+
+    PendingEvents.emplace_back(ev->ReleaseBase().Release());
+}
+
+void TPartition::HandleOnInit(TEvPQ::TEvGetWriteInfoResponse::TPtr& ev, const TActorContext&)
+{
+    PQ_LOG_D("HandleOnInit TEvPQ::TEvGetWriteInfoResponse");
+
+    Y_ABORT_UNLESS(!IsSupportive());
+
+    PendingEvents.emplace_back(ev->ReleaseBase().Release());
+}
+
+void TPartition::HandleOnInit(TEvPQ::TEvGetWriteInfoError::TPtr& ev, const TActorContext&)
+{
+    PQ_LOG_D("HandleOnInit TEvPQ::TEvGetWriteInfoError");
+
+    Y_ABORT_UNLESS(!IsSupportive());
 
     PendingEvents.emplace_back(ev->ReleaseBase().Release());
 }
@@ -1058,7 +1079,9 @@ void TPartition::Handle(TEvPQ::TEvTxRollback::TPtr& ev, const TActorContext& ctx
 }
 
 void TPartition::Handle(TEvPQ::TEvGetWriteInfoRequest::TPtr& ev, const TActorContext& ctx) {
+    PQ_LOG_D("Handle TEvPQ::TEvGetWriteInfoRequest");
     if (ClosedInternalPartition || WaitingForPreviousBlobQuota() || (CurrentStateFunc() != &TThis::StateIdle)) {
+        PQ_LOG_D("Send TEvPQ::TEvGetWriteInfoError");
         auto* response = new TEvPQ::TEvGetWriteInfoError(Partition.InternalPartitionId,
                                                          "Write info requested while writes are not complete");
         ctx.Send(ev->Sender, response);
@@ -1084,6 +1107,7 @@ void TPartition::Handle(TEvPQ::TEvGetWriteInfoRequest::TPtr& ev, const TActorCon
     response->MessagesSizes = std::move(MessageSize.GetValues());
     response->InputLags = std::move(SupportivePartitionTimeLag);
 
+    PQ_LOG_D("Send TEvPQ::TEvGetWriteInfoResponse");
     ctx.Send(ev->Sender, response);
 }
 
@@ -1170,7 +1194,9 @@ void TPartition::Handle(TEvPQ::TEvGetWriteInfoResponse::TPtr& ev, const TActorCo
 
 
 void TPartition::Handle(TEvPQ::TEvGetWriteInfoError::TPtr& ev, const TActorContext& ctx) {
-    PQ_LOG_D("Handle TEvPQ::TEvGetWriteInfoError");
+    PQ_LOG_D("Handle TEvPQ::TEvGetWriteInfoError " <<
+             "Cookie " << ev->Get()->Cookie <<
+             ", Message " << ev->Get()->Message);
     WriteInfoResponseHandler(ev->Sender, ev->Release(), ctx);
 }
 
@@ -1856,7 +1882,6 @@ void TPartition::ProcessCommitQueue() {
         return this->ExecUserActionOrTransaction(event, request);
     };
     while (!UserActionAndTxPendingCommit.empty()) {
-        // UserActionAndTxPendingCommit.pop_front();
         auto& front = UserActionAndTxPendingCommit.front();
         auto state = ECommitState::Committed;
         if (auto* tx = get_if<TSimpleSharedPtr<TTransaction>>(&front.Event)) {
@@ -2023,8 +2048,9 @@ TPartition::EProcessResult TPartition::PreProcessUserActionOrTransaction(TSimple
         Y_ABORT_UNLESS(t->ChangeConfig);
 
         Y_ABORT_UNLESS(!ChangeConfig && !ChangingConfig);
-        if (!FirstEvent)
+        if (!FirstEvent) {
             return EProcessResult::Blocked;
+        }
         ChangingConfig = true;
         // Should remove this and add some id to TEvChangeConfig if we want to batch change of configs
         t->State = ECommitState::Committed;
@@ -2180,6 +2206,9 @@ void TPartition::CommitWriteOperations(TTransaction& t)
     PQ_LOG_D("Head=" << Head << ", NewHead=" << NewHead);
 
     if (!t.WriteInfo->BodyKeys.empty()) {
+        bool needCompactHead =
+            (Parameters->FirstCommitWriteOperations ? Head : NewHead).PackedSize != 0;
+
         PartitionedBlob = TPartitionedBlob(Partition,
                                            NewHead.Offset,
                                            "", // SourceId
@@ -2189,7 +2218,7 @@ void TPartition::CommitWriteOperations(TTransaction& t)
                                            Head,
                                            NewHead,
                                            Parameters->HeadCleared,  // headCleared
-                                           Head.PackedSize != 0,     // needCompactHead
+                                           needCompactHead,          // needCompactHead
                                            MaxBlobSize);
 
         for (auto& k : t.WriteInfo->BodyKeys) {
@@ -2200,9 +2229,9 @@ void TPartition::CommitWriteOperations(TTransaction& t)
                 CompactedKeys.emplace_back(write->Key, write->Value.size());
                 ClearOldHead(write->Key.GetOffset(), write->Key.GetPartNo(), PersistRequest.Get());
             }
+            Parameters->CurOffset += k.Key.GetCount();
         }
 
-
         PQ_LOG_D("PartitionedBlob.GetFormedBlobs().size=" << PartitionedBlob.GetFormedBlobs().size());
         if (const auto& formedBlobs = PartitionedBlob.GetFormedBlobs(); !formedBlobs.empty()) {
             ui32 curWrites = RenameTmpCmdWrites(PersistRequest.Get());
@@ -2213,17 +2242,15 @@ void TPartition::CommitWriteOperations(TTransaction& t)
                               ctx);
         }
 
-        const auto& last = t.WriteInfo->BodyKeys.back();
-
-        NewHead.Offset += (last.Key.GetOffset() + last.Key.GetCount());
+        NewHead.Clear();
+        NewHead.Offset = Parameters->CurOffset;
     }
 
     if (!t.WriteInfo->BlobsFromHead.empty()) {
         auto& first = t.WriteInfo->BlobsFromHead.front();
         NewHead.PartNo = first.GetPartNo();
 
-        Parameters->CurOffset = NewHead.Offset;
-        Parameters->HeadCleared = !t.WriteInfo->BodyKeys.empty();
+        Parameters->HeadCleared = Parameters->HeadCleared || !t.WriteInfo->BodyKeys.empty();
 
         PartitionedBlob = TPartitionedBlob(Partition,
                                            NewHead.Offset,
@@ -2268,6 +2295,8 @@ void TPartition::CommitWriteOperations(TTransaction& t)
         }
     }
 
+    Parameters->FirstCommitWriteOperations = false;
+
     WriteInfosApplied.emplace_back(std::move(t.WriteInfo));
 }
 
@@ -2544,6 +2573,8 @@ void TPartition::ChangePlanStepAndTxId(ui64 step, ui64 txId)
 
 void TPartition::ResendPendingEvents(const TActorContext& ctx)
 {
+    PQ_LOG_D("Resend pending events. Count " << PendingEvents.size());
+
     while (!PendingEvents.empty()) {
         ctx.Schedule(TDuration::Zero(), PendingEvents.front().release());
         PendingEvents.pop_front();
@@ -3416,8 +3447,6 @@ void TPartition::Handle(TEvPQ::TEvCheckPartitionStatusRequest::TPtr& ev, const T
 
 void TPartition::HandleOnInit(TEvPQ::TEvDeletePartition::TPtr& ev, const TActorContext&)
 {
-    PQ_LOG_D("HandleOnInit TEvPQ::TEvDeletePartition");
-
     Y_ABORT_UNLESS(IsSupportive());
 
     PendingEvents.emplace_back(ev->ReleaseBase().Release());

ydb/core/persqueue/partition.h

@@ -416,6 +416,9 @@ class TPartition : public TActorBootstrapped<TPartition> {
     void HandleOnInit(TEvPQ::TEvTxCommit::TPtr& ev, const TActorContext& ctx);
     void HandleOnInit(TEvPQ::TEvTxRollback::TPtr& ev, const TActorContext& ctx);
     void HandleOnInit(TEvPQ::TEvProposePartitionConfig::TPtr& ev, const TActorContext& ctx);
+    void HandleOnInit(TEvPQ::TEvGetWriteInfoRequest::TPtr& ev, const TActorContext& ctx);
+    void HandleOnInit(TEvPQ::TEvGetWriteInfoResponse::TPtr& ev, const TActorContext& ctx);
+    void HandleOnInit(TEvPQ::TEvGetWriteInfoError::TPtr& ev, const TActorContext& ctx);
 
     void ChangePlanStepAndTxId(ui64 step, ui64 txId);
 
@@ -523,10 +526,10 @@ class TPartition : public TActorBootstrapped<TPartition> {
             HFuncTraced(NReadQuoterEvents::TEvQuotaUpdated, Handle);
             HFuncTraced(NReadQuoterEvents::TEvAccountQuotaCountersUpdated, Handle);
             HFuncTraced(NReadQuoterEvents::TEvQuotaCountersUpdated, Handle);
-            HFuncTraced(TEvPQ::TEvGetWriteInfoRequest, Handle);
+            HFuncTraced(TEvPQ::TEvGetWriteInfoRequest, HandleOnInit);
 
-            HFuncTraced(TEvPQ::TEvGetWriteInfoResponse, Handle);
-            HFuncTraced(TEvPQ::TEvGetWriteInfoError, Handle);
+            HFuncTraced(TEvPQ::TEvGetWriteInfoResponse, HandleOnInit);
+            HFuncTraced(TEvPQ::TEvGetWriteInfoError, HandleOnInit);
             HFuncTraced(TEvPQ::TEvDeletePartition, HandleOnInit);
             IgnoreFunc(TEvPQ::TEvTxBatchComplete);
         default:
@@ -618,6 +621,7 @@ class TPartition : public TActorBootstrapped<TPartition> {
         ui64 CurOffset;
         bool OldPartsCleared;
         bool HeadCleared;
+        bool FirstCommitWriteOperations = true;
     };
 
     static void RemoveMessages(TMessageQueue& src, TMessageQueue& dst);

ydb/core/persqueue/partition_write.cpp

@@ -460,7 +460,7 @@ void TPartition::OnHandleWriteResponse(const TActorContext& ctx)
 
 void TPartition::Handle(TEvPQ::TEvHandleWriteResponse::TPtr&, const TActorContext& ctx)
 {
-    PQ_LOG_T("TPartition::HandleOnWrite TEvHandleWriteResponse.");
+    PQ_LOG_T("TPartition::Handle TEvHandleWriteResponse.");
     OnHandleWriteResponse(ctx);
 }