Drain polled tasks on shutdown

internal/internal_task_pollers.go

@@ -310,24 +310,13 @@ func (bp *basePoller) doPoll(pollFunc func(ctx context.Context) (taskForWorker,
 	}()
 
 	if bp.workerPollCompleteOnShutdown != nil && bp.workerPollCompleteOnShutdown.Load() {
-		// Don't kill the gRPC stream. After ShutdownWorker, the server returns empty responses.
-		select {
-		case <-doneC:
-			return result, err
-		case <-bp.stopC:
-			// TEMP FIX: Give the server a reasonable window to complete the poll after
-			// ShutdownWorker. Fall back to cancelling the poll if it takes too
-			// long, e.g. when the gRPC connection was closed before Stop().
-			timer := time.NewTimer(5 * time.Second)
-			defer timer.Stop()
-			select {
-			case <-doneC:
-			case <-timer.C:
-				cancel()
-				<-doneC
-			}
-			return result, err
-		}
+		// Don't cancel the gRPC stream. After ShutdownWorker, the server
+		// completes the poll with an empty response. The poll is bounded by
+		// the gRPC timeout (pollTaskServiceTimeOut). The worker's stop
+		// timeout (WorkerStopTimeout) controls how long Stop() blocks;
+		// goroutines clean up in the background within the gRPC deadline.
+		<-doneC
+		return result, err
 	}
 
 	// Legacy: cancel in-flight polls immediately on shutdown

internal/internal_worker_base.go

@@ -216,6 +216,7 @@ type (
 		lastPollTaskErrLock    sync.Mutex
 
 		noRepoll atomic.Bool
+		pollerWG sync.WaitGroup
 	}
 
 	eagerOrPolledTask interface {
@@ -391,6 +392,7 @@ func (bw *baseWorker) Start() {
 
 		for i := 0; i < taskWorker.pollerCount; i++ {
 			bw.stopWG.Add(1)
+			bw.pollerWG.Add(1)
 			go bw.runPoller(taskWorker)
 		}
 
@@ -403,6 +405,15 @@ func (bw *baseWorker) Start() {
 		}
 	}
 
+	// When all pollers have exited, close taskQueueCh so the dispatcher
+	// knows no more polled tasks will arrive and can drain what remains.
+	bw.stopWG.Add(1)
+	go func() {
+		defer bw.stopWG.Done()
+		bw.pollerWG.Wait()
+		close(bw.taskQueueCh)
+	}()
+
 	bw.stopWG.Add(1)
 	go bw.runTaskDispatcher()
 
@@ -428,6 +439,7 @@ func (bw *baseWorker) isStop() bool {
 
 func (bw *baseWorker) runPoller(taskWorker scalableTaskPoller) {
 	defer bw.stopWG.Done()
+	defer bw.pollerWG.Done()
 	// Note: With poller autoscaling, this metric doesn't make a lot of sense since the number of pollers can go up and down.
 	bw.metricsHandler.Counter(metrics.PollerStartCounter).Inc(1)
 
@@ -561,24 +573,17 @@ func (bw *baseWorker) processTaskAsync(eagerOrPolled eagerOrPolledTask) {
 func (bw *baseWorker) runTaskDispatcher() {
 	defer bw.stopWG.Done()
 
-	for {
-		// wait for new task or worker stop
-		select {
-		case <-bw.stopCh:
-			// Currently we can drop any tasks received when closing.
-			// https://github.com/temporalio/sdk-go/issues/1197
-			return
-		case task := <-bw.taskQueueCh:
-			// for non-polled-task (local activity result as task or eager task), we don't need to rate limit
-			_, isPolledTask := task.(*polledTask)
-			if isPolledTask && bw.taskLimiter.Wait(bw.limiterContext) != nil {
-				if bw.isStop() {
-					bw.releaseSlot(task.getPermit(), SlotReleaseReasonUnused)
-					return
-				}
-			}
-			bw.processTaskAsync(task)
+	for task := range bw.taskQueueCh {
+		// For non-polled-task (local activity result as task or eager task),
+		// we don't need to rate limit. During shutdown the limiter context
+		// is cancelled, so Wait returns immediately — we still process the
+		// task rather than dropping it.
+		if _, isPolledTask := task.(*polledTask); isPolledTask {
+			// Ignore error: during shutdown the limiter context is
+			// cancelled, but we still process remaining tasks.
+			_ = bw.taskLimiter.Wait(bw.limiterContext)
 		}
+		bw.processTaskAsync(task)
 	}
 }
 
@@ -639,11 +644,10 @@ func (bw *baseWorker) pollTask(taskWorker scalableTaskPoller, slotPermit *SlotPe
 			taskWorker.pollerAutoscalerReportHandle.handleTask(task)
 		}
 
-		select {
-		case bw.taskQueueCh <- &polledTask{task: task, permit: slotPermit}:
-			didSendTask = true
-		case <-bw.stopCh:
-		}
+		// The dispatcher is guaranteed to be alive: it only exits after
+		// taskQueueCh is closed, which happens after all pollers finish.
+		bw.taskQueueCh <- &polledTask{task: task, permit: slotPermit}
+		didSendTask = true
 	}
 }
 

internal/internal_worker_base_test.go

@@ -315,6 +315,104 @@ type noopTaskProcessor struct{}
 
 func (noopTaskProcessor) ProcessTask(any) error { return nil }
 
+// TestTaskNotDroppedDuringShutdown verifies the two-stage shutdown: when a
+// poller receives a task during shutdown, the task is still dispatched and
+// processed rather than silently dropped.
+func TestTaskNotDroppedDuringShutdown(t *testing.T) {
+	taskProcessed := make(chan struct{})
+	pollStarted := make(chan struct{})
+
+	// A poller that blocks until returnTask is closed, then returns a task
+	// exactly once. Subsequent polls return nil so the poller can exit.
+	tp := &shutdownTaskPoller{
+		pollStarted: pollStarted,
+		returnTask:  make(chan struct{}),
+		task:        &testTask{},
+	}
+
+	processor := &recordingTaskProcessor{
+		processed: taskProcessed,
+	}
+
+	bw := newBaseWorker(baseWorkerOptions{
+		slotSupplier:     &testSlotSupplier{},
+		maxTaskPerSecond: 1000,
+		taskPollers: []scalableTaskPoller{
+			{taskPollerType: "test", pollerCount: 1, taskPoller: tp},
+		},
+		taskProcessor:  processor,
+		workerType:     "ShutdownTest",
+		logger:         ilog.NewNopLogger(),
+		stopTimeout:    5 * time.Second,
+		metricsHandler: metrics.NopHandler,
+	})
+
+	bw.Start()
+
+	// Wait for the poller to be actively polling.
+	<-pollStarted
+
+	// Release the poller so it returns a task, then stop the worker.
+	// The poller returns a task and then nil on subsequent polls,
+	// allowing it to exit via noRepoll/stopCh during Stop().
+	close(tp.returnTask)
+
+	// Stop exercises the real shutdown path: noRepoll, close(stopCh),
+	// limiterContextCancel, and awaitWaitGroup.
+	stopDone := make(chan struct{})
+	go func() {
+		bw.Stop()
+		close(stopDone)
+	}()
+
+	select {
+	case <-taskProcessed:
+		// Success: the task was dispatched and processed during shutdown
+	case <-time.After(5 * time.Second):
+		t.Fatal("task polled during shutdown was not processed (dropped)")
+	}
+
+	select {
+	case <-stopDone:
+		// Stop completed cleanly
+	case <-time.After(5 * time.Second):
+		t.Fatal("Stop() did not return in time")
+	}
+}
+
+// shutdownTaskPoller blocks until returnTask is closed, then returns a task
+// exactly once. Subsequent polls return nil.
+type shutdownTaskPoller struct {
+	pollStarted chan struct{}
+	returnTask  chan struct{}
+	task        taskForWorker
+	returned    atomic.Bool
+}
+
+func (p *shutdownTaskPoller) PollTask() (taskForWorker, error) {
+	select {
+	case p.pollStarted <- struct{}{}:
+	default:
+	}
+	<-p.returnTask
+	if p.returned.CompareAndSwap(false, true) {
+		return p.task, nil
+	}
+	return nil, nil
+}
+
+type recordingTaskProcessor struct {
+	processed chan struct{}
+}
+
+func (p *recordingTaskProcessor) ProcessTask(any) error {
+	select {
+	case p.processed <- struct{}{}:
+	default:
+	}
+	return nil
+}
+
 func (s *PollScalerReportHandleSuite) TestAutoscaleDownOnTimeoutWithCapability() {
 	targetSuggestion := 0
 	ps := newPollScalerReportHandle(pollScalerReportHandleOptions{