feat: support durable task cancellation without foreign keys

sdk/go/worker_runtime.go

@@ -14,6 +14,7 @@ type runtimeAttempt struct {
 	startID, renewID, finishID string
 	guard                      *executionGuard
 	invoked, returned, lost    bool
+	cancelRequested            bool
 	result                     WorkerFinish
 	nextTry, nextRenew         time.Time
 	retryDelay                 time.Duration
@@ -152,9 +153,11 @@ func (w *Worker) Run(parent context.Context) error {
 			}
 		}
 		var assignments <-chan WorkerAssignment
+		var cancellations <-chan WorkerCancellation
 		var closed <-chan struct{}
 		if r.conn != nil {
 			assignments = r.conn.assignments
+			cancellations = r.conn.cancellations
 			closed = r.conn.Done()
 		}
 		select {
@@ -175,6 +178,10 @@ func (w *Worker) Run(parent context.Context) error {
 				j.result = done.result
 				j.nextTry = time.Time{}
 			}
+		case c := <-cancellations:
+			if j := r.jobs[c.AttemptID]; j != nil && j.assignment.Task.ID == c.TaskID && j.assignment.Epoch == c.Epoch {
+				r.cancelAttempt(j)
+			}
 		case reply := <-r.replies:
 			if err := r.receive(reply); err != nil {
 				return r.failure(err)
@@ -232,19 +239,30 @@ func (r *workerRuntime) remove(j *runtimeAttempt) {
 }
 func (r *workerRuntime) observeExpired() {
 	for _, j := range r.jobs {
-		if j.guard != nil && j.guard.ctx.Err() != nil && !j.lost {
+		if j.guard != nil && j.guard.ctx.Err() != nil && !j.lost && !j.cancelRequested {
 			j.lost = true
 			r.event(j, "lease_lost", "LEASE_LOST")
 		}
 	}
 }
+
+func (r *workerRuntime) cancelAttempt(j *runtimeAttempt) {
+	if j.cancelRequested {
+		return
+	}
+	j.cancelRequested = true
+	if j.guard != nil {
+		j.guard.stop(ErrTaskCancellationRequested)
+	}
+	r.event(j, "cancel_requested", "CANCEL_REQUESTED")
+}
 func (r *workerRuntime) desiredCapacity() uint32 {
 	if r.draining {
 		return 0
 	}
 	ghosts := uint32(0)
 	for _, j := range r.jobs {
-		if j.returned || j.lost {
+		if j.returned || j.lost || j.cancelRequested {
 			ghosts++
 		}
 	}
@@ -320,7 +338,10 @@ func (r *workerRuntime) establish(c runtimeConnected) error {
 		if j == nil {
 			continue
 		}
-		if j.guard != nil {
+		if lease.CancelRequestedAt != nil {
+			r.cancelAttempt(j)
+		}
+		if j.guard != nil && !j.cancelRequested {
 			if !j.guard.update(lease) {
 				j.lost = true
 				r.event(j, "lease_lost", "LEASE_LOST")
@@ -398,7 +419,7 @@ func (r *workerRuntime) dispatch() error {
 				}
 				id = &j.finishID
 			case "renew":
-				if !j.invoked || j.lost || j.returned || now.Before(j.nextRenew) {
+				if !j.invoked || j.lost || j.cancelRequested || j.returned || now.Before(j.nextRenew) {
 					continue
 				}
 				id = &j.renewID
@@ -509,7 +530,17 @@ func (r *workerRuntime) receive(reply runtimeReply) error {
 		r.remove(j)
 		return nil
 	}
+	if reply.result.Lease.CancelRequestedAt != nil {
+		r.cancelAttempt(j)
+	}
 	if command.kind == "start" {
+		if j.cancelRequested {
+			// 开始确认已收到，但本进程未调用处理器；可以明确证明本次没有副作用。
+			j.invoked, j.returned = true, true
+			j.result = cancelledHandlerResult(j.assignment.Attempt(), nil)
+			j.nextTry = time.Time{}
+			return nil
+		}
 		guard, ok := newExecutionGuard(r.ctx, reply.result.Lease)
 		if !ok {
 			r.event(j, "lease_lost", "START_NOT_AUTHORIZED")
@@ -529,14 +560,14 @@ func (r *workerRuntime) receive(reply runtimeReply) error {
 	} else {
 		r.worker.renewals.Add(1)
 		j.renewID = ""
-		if !j.guard.update(reply.result.Lease) {
+		if !j.cancelRequested && !j.guard.update(reply.result.Lease) {
 			j.lost = true
 			r.event(j, "lease_lost", "LEASE_LOST")
 		}
 	}
 	j.nextTry = time.Time{}
 	j.retryDelay = 0
-	if !j.lost {
+	if !j.lost && !j.cancelRequested {
 		j.nextRenew = time.Now().Add(max(20*time.Millisecond, time.Until(j.guard.until())/3))
 	}
 	return nil