Fix busywait on adding to full async input buffer (backport 24-3 #14522) (#15584)

Author: yumkam

ydb/library/yql/dq/actors/compute/dq_async_compute_actor.cpp

@@ -605,7 +605,20 @@ class TDqAsyncComputeActor : public TDqComputeActorBase<TDqAsyncComputeActor, TC
     }
 
     void DoExecuteImpl() override {
-        PollAsyncInput();
+        LastPollResult = PollAsyncInput();
+
+        if (LastPollResult && *LastPollResult != EResumeSource::CAPollAsyncNoSpace) {
+            // When (some) source buffers was not full, and (some) was successfully polled,
+            // initiate next DoExecute run immediately;
+            // If only reason for continuing was lack on space on all source
+            // buffers, only continue execution after run completed,
+            // (some) sources was consumed and compute waits for input
+            // (Otherwise we enter busy-poll, and there are especially bad scenario
+            // when compute is delayed by rate-limiter, we enter busy-poll here,
+            // this spends cpu, ratelimiter delays compute execution even more))
+            ContinueExecute(*std::exchange(LastPollResult, {}));
+        }
+
         if (ProcessSourcesState.Inflight == 0) {
             auto req = GetCheckpointRequest();
             CA_LOG_T("DoExecuteImpl: " << (bool) req);
@@ -1087,6 +1100,9 @@ class TDqAsyncComputeActor : public TDqComputeActorBase<TDqAsyncComputeActor, TC
             CA_LOG_T("AsyncCheckRunStatus: TakeInputChannelDataRequests: " << TakeInputChannelDataRequests.size());
             return;
         }
+        if (ProcessOutputsState.LastRunStatus == ERunStatus::PendingInput && LastPollResult) {
+            ContinueExecute(*LastPollResult);
+        }
         TBase::CheckRunStatus();
     }
 
@@ -1134,6 +1150,7 @@ class TDqAsyncComputeActor : public TDqComputeActorBase<TDqAsyncComputeActor, TC
     NMonitoring::THistogramPtr CpuTimeQuotaWaitDelay;
     NMonitoring::TDynamicCounters::TCounterPtr CpuTime;
     NDqProto::TEvComputeActorState ComputeActorState;
+    TMaybe<EResumeSource> LastPollResult;
 };
 
 

ydb/library/yql/dq/actors/compute/dq_compute_actor_impl.h

@@ -397,6 +397,14 @@ class TDqComputeActorBase : public NActors::TActorBootstrapped<TDerived>
             return;
         }
 
+        if (status == ERunStatus::PendingInput) {
+            for (auto& [id, inputTransform] : InputTransformsMap) {
+                if (!inputTransform.Buffer->Empty()) {
+                    ContinueExecute(EResumeSource::CAPendingInput);
+                }
+            }
+        }
+
         if (status != ERunStatus::Finished) {
             // If the incoming channel's buffer was full at the moment when last ChannelDataAck event had been sent,
             // there will be no attempts to send a new piece of data from the other side of this channel.
@@ -1389,26 +1397,38 @@ class TDqComputeActorBase : public NActors::TActorBootstrapped<TDerived>
         }
     }
 
-    void PollAsyncInput() {
+    [[nodiscard]]
+    TMaybe<EResumeSource> PollAsyncInput() {
+        TMaybe<EResumeSource> pollResult;
+        if (!Running) {
+            CA_LOG_T("Skip polling inputs and sources because not running");
+            return pollResult;
+        }
+
+        CA_LOG_T("Poll inputs");
+        for (auto& [inputIndex, transform] : InputTransformsMap) {
+            if (auto resume = transform.PollAsyncInput(MetricsReporter, WatermarksTracker, RuntimeSettings.AsyncInputPushLimit)) {
+                if (!pollResult || *pollResult == EResumeSource::CAPollAsyncNoSpace) {
+                    pollResult = resume;
+                }
+            }
+        }
+
         // Don't produce any input from sources if we're about to save checkpoint.
-        if (!Running || (Checkpoints && Checkpoints->HasPendingCheckpoint() && !Checkpoints->ComputeActorStateSaved())) {
+        if ((Checkpoints && Checkpoints->HasPendingCheckpoint() && !Checkpoints->ComputeActorStateSaved())) {
             CA_LOG_T("Skip polling sources because of pending checkpoint");
-            return;
+            return pollResult;
         }
 
         CA_LOG_T("Poll sources");
         for (auto& [inputIndex, source] : SourcesMap) {
             if (auto resume =  source.PollAsyncInput(MetricsReporter, WatermarksTracker, RuntimeSettings.AsyncInputPushLimit)) {
-                ContinueExecute(*resume);
-            }
-        }
-
-        CA_LOG_T("Poll inputs");
-        for (auto& [inputIndex, transform] : InputTransformsMap) {
-            if (auto resume = transform.PollAsyncInput(MetricsReporter, WatermarksTracker, RuntimeSettings.AsyncInputPushLimit)) {
-                ContinueExecute(*resume);
+                if (!pollResult || *pollResult == EResumeSource::CAPollAsyncNoSpace) {
+                    pollResult = resume;
+                }
             }
         }
+        return pollResult;
     }
 
     void OnNewAsyncInputDataArrived(const IDqComputeActorAsyncInput::TEvNewAsyncInputDataArrived::TPtr& ev) {

ydb/library/yql/dq/actors/compute/dq_sync_compute_actor_base.h

@@ -30,7 +30,7 @@ class TDqSyncComputeActorBase: public TDqComputeActorBase<TDerived, TComputeActo
     void DoExecuteImpl() override{
         auto sourcesState = static_cast<TDerived*>(this)->GetSourcesState();
 
-        TBase::PollAsyncInput();
+        auto lastPollResult = TBase::PollAsyncInput();
         ERunStatus status = TaskRunner->Run();
 
         CA_LOG_T("Resume execution, run status: " << status);
@@ -44,6 +44,13 @@ class TDqSyncComputeActorBase: public TDqComputeActorBase<TDerived, TComputeActo
         }
 
         TBase::ProcessOutputsImpl(status);
+
+        if (lastPollResult && (*lastPollResult != EResumeSource::CAPollAsyncNoSpace || status == ERunStatus::PendingInput)) {
+            // If only reason for continuing was lack on space on all sources,
+            // only continue execution when input was consumed;
+            // otherwise this may result in busy-poll
+            TBase::ContinueExecute(*lastPollResult);
+        }
     }
 
     void DoTerminateImpl() override {

ydb/library/yql/dq/actors/input_transforms/dq_input_transform_lookup.cpp

@@ -185,7 +185,7 @@ class TInputTransformStreamLookupBase
             }
         }
         finished = IsFinished();
-        return 0;
+        return AwaitingQueue.RowCount();
     }
 
     TMaybe<google::protobuf::Any> ExtraData() override {

ydb/library/yql/dq/actors/task_runner/task_runner_actor_local.cpp

@@ -144,6 +144,18 @@ class TLocalTaskRunnerActor
                 return false;
             }
         }
+        for (const auto transformId: InputTransforms) {
+            const auto t = TaskRunner->GetInputTransform(transformId);
+            if (t) {
+                auto [_, transform] = *t;
+                if (!transform->Empty()) {
+                    return false;
+                }
+                if (transform->IsPending()) {
+                    return false;
+                }
+            }
+        }
         return true;
     }
 
@@ -443,6 +455,7 @@ class TLocalTaskRunnerActor
         for (auto i = 0; i != inputs.size(); ++i) {
             if (auto t = TaskRunner->GetInputTransform(i)) {
                 inputTransforms[i] = *t;
+                InputTransforms.emplace(i);
             }
         }
 
@@ -490,6 +503,7 @@ class TLocalTaskRunnerActor
     const TTxId TxId;
     const ui64 TaskId;
     THashSet<ui32> Inputs;
+    THashSet<ui32> InputTransforms;
     THashSet<ui32> Sources;
     TIntrusivePtr<NDq::IDqTaskRunner> TaskRunner;
     THashSet<ui32> InputChannelsWithDisabledCheckpoints;

ydb/library/yql/dq/runtime/dq_async_input.cpp

@@ -6,6 +6,7 @@ namespace NYql::NDq {
 class TDqAsyncInputBuffer : public TDqInputImpl<TDqAsyncInputBuffer, IDqAsyncInputBuffer> {
     using TBaseImpl = TDqInputImpl<TDqAsyncInputBuffer, IDqAsyncInputBuffer>;
     friend TBaseImpl;
+    bool Pending = false;
 public:
     TDqAsyncInputBufferStats PushStats;
     TDqInputStats PopStats;
@@ -32,7 +33,7 @@ class TDqAsyncInputBuffer : public TDqInputImpl<TDqAsyncInputBuffer, IDqAsyncInp
     }
 
     void Push(NKikimr::NMiniKQL::TUnboxedValueBatch&& batch, i64 space) override {
-        Y_ABORT_UNLESS(!batch.empty() || !space);
+        Pending = space != 0;
         if (!batch.empty()) {
             AddBatch(std::move(batch), space);
         }
@@ -41,6 +42,10 @@ class TDqAsyncInputBuffer : public TDqInputImpl<TDqAsyncInputBuffer, IDqAsyncInp
     virtual void Push(TDqSerializedBatch&&, i64) override {
         YQL_ENSURE(!"Unimplemented");
     }
+
+    bool IsPending() const override {
+        return Pending;
+    }
 };
 
 IDqAsyncInputBuffer::TPtr CreateDqAsyncInputBuffer(

ydb/library/yql/dq/runtime/dq_async_input.h

@@ -23,6 +23,8 @@ class IDqAsyncInputBuffer : public IDqInput {
     virtual void Push(TDqSerializedBatch&& batch, i64 space) = 0;
 
     virtual void Finish() = 0;
+
+    virtual bool IsPending() const { return false; };
 };
 
 IDqAsyncInputBuffer::TPtr CreateDqAsyncInputBuffer(ui64 inputIndex, const TString& type, NKikimr::NMiniKQL::TType* inputType, ui64 maxBufferBytes,