fix(supervisor): bound the eager restart with a deadline

maybeRestartAfterTask ran its restart on a bare background context while
the drain loop in doRestartLocked selects only on ctx.Done(). A task that
never completed blocked the drain forever, pinning isRestarting. Since a
planned restart now reports healthy, that left the node claiming health
for good. LibreOffice escaped it because its concurrency of 1 drains no
slots, but Chromium defaults to 6.

Give that restart its own deadline. On expiry it aborts and the next task
retries it.

pkg/gotenberg/supervisor.go

@@ -115,6 +115,17 @@ const (
 	restartReasonMaxRequests = "max_requests"
 )
 
+// defaultEagerRestartTimeout bounds the restart triggered after the maximum
+// request limit. That restart runs on a background context, unlike the one from
+// ensureHealthy which inherits the request deadline, so without a deadline of
+// its own the drain loop in [processSupervisor.doRestartLocked] would wait
+// forever on a task that never completes. That would pin isRestarting and,
+// with it, the health reported by [processSupervisor.Healthy]. Sized well above
+// --api-timeout (30s by default) plus the engine start timeouts (20s by
+// default) so it never fires while tasks are merely slow. The eager restart is
+// opportunistic: on expiry it aborts, and the next task retries it.
+const defaultEagerRestartTimeout = 2 * time.Minute
+
 type processSupervisor struct {
 	logger         *slog.Logger
 	engine         string
@@ -152,6 +163,9 @@ type processSupervisor struct {
 	consecutiveHealthFailures atomic.Int64  // reset to 0 on every successful probe
 	idleMu                    sync.Mutex    // protects idleStopChan
 	idleStopChan              chan struct{} // signal to stop the idle ticker goroutine
+	// eagerRestartTimeout bounds the restart from maybeRestartAfterTask.
+	// Defaults to [defaultEagerRestartTimeout]; only tests shorten it.
+	eagerRestartTimeout time.Duration
 }
 
 // NewProcessSupervisor initializes a new [ProcessSupervisor]. engine names the
@@ -175,6 +189,7 @@ func NewProcessSupervisor(logger *slog.Logger, engine string, process Process, m
 		maxQueueSize:        maxQueueSize,
 		maxConcurrency:      maxConcurrency,
 		idleShutdownTimeout: idleShutdownTimeout,
+		eagerRestartTimeout: defaultEagerRestartTimeout,
 	}
 	b.reqCounter.Store(0)
 	b.reqQueueSize.Store(0)
@@ -558,9 +573,10 @@ func (s *processSupervisor) ensureHealthy(ctx context.Context) error {
 }
 
 // maybeRestartAfterTask checks if the maximum request limit has been reached
-// and, if so, triggers an asynchronous restart. If a restart is initiated, it
-// takes ownership of the caller's semaphore slot (the caller must not release
-// it). Returns true if ownership was taken.
+// and, if so, triggers an asynchronous restart bounded by
+// [eagerRestartTimeout]. If a restart is initiated, it takes ownership of the
+// caller's semaphore slot (the caller must not release it). Returns true if
+// ownership was taken.
 func (s *processSupervisor) maybeRestartAfterTask(logger *slog.Logger) bool {
 	if s.maxReqLimit <= 0 || s.reqCounter.Load() < s.maxReqLimit {
 		return false
@@ -573,7 +589,10 @@ func (s *processSupervisor) maybeRestartAfterTask(logger *slog.Logger) bool {
 	s.logger.DebugContext(context.Background(), "max request limit reached, restarting eagerly...")
 
 	go func() {
-		restartErr := s.doRestartLocked(context.Background(), restartReasonMaxRequests)
+		ctx, cancel := context.WithTimeout(context.Background(), s.eagerRestartTimeout)
+		defer cancel()
+
+		restartErr := s.doRestartLocked(ctx, restartReasonMaxRequests)
 		s.restartMutex.Unlock()
 		if restartErr != nil {
 			s.logger.ErrorContext(context.Background(), fmt.Sprintf("process restart after task: %v", restartErr))

pkg/gotenberg/supervisor_test.go

@@ -381,6 +381,91 @@ func TestProcessSupervisor_Healthy_UnplannedRestart(t *testing.T) {
 	close(release)
 }
 
+// TestProcessSupervisor_doRestartLocked_DrainDeadline verifies that a drain
+// unable to acquire every slot gives up on the context deadline and clears
+// isRestarting. Without a deadline on the eager restart, a task that never
+// completes would pin the flag and, since a planned restart reports healthy,
+// leave the supervisor claiming health forever.
+func TestProcessSupervisor_doRestartLocked_DrainDeadline(t *testing.T) {
+	logger := slog.New(slog.DiscardHandler)
+
+	var starts atomic.Int64
+	process := &ProcessMock{
+		StartMock: func(_ *slog.Logger) error {
+			starts.Add(1)
+
+			return nil
+		},
+		StopMock:    func(_ *slog.Logger) error { return nil },
+		HealthyMock: func(_ *slog.Logger) bool { return true },
+	}
+
+	// A concurrency of 2 makes the drain acquire one slot on top of the one the
+	// triggering task hands over. Fill the semaphore so it never can, mimicking
+	// a concurrent task that never completes.
+	ps := NewProcessSupervisor(logger, "test", process, 1, 0, 2, 0).(*processSupervisor)
+	ps.semaphore <- struct{}{}
+	ps.semaphore <- struct{}{}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 50*time.Millisecond)
+	defer cancel()
+
+	err := ps.doRestartLocked(ctx, restartReasonMaxRequests)
+	if err == nil {
+		t.Fatal("expected the drain to fail on the context deadline")
+	}
+
+	if ps.isRestarting.Load() {
+		t.Fatal("expected isRestarting to be cleared after a failed drain")
+	}
+
+	if starts.Load() != 0 {
+		t.Fatalf("expected no restart attempt after a failed drain but got %d", starts.Load())
+	}
+}
+
+// TestProcessSupervisor_maybeRestartAfterTask_Bounded verifies that the eager
+// restart runs under a deadline. A concurrent task that never completes blocks
+// the drain, and without a bound the restart goroutine would wait forever with
+// isRestarting pinned, leaving Healthy() reporting a planned restart for good.
+func TestProcessSupervisor_maybeRestartAfterTask_Bounded(t *testing.T) {
+	logger := slog.New(slog.DiscardHandler)
+
+	process := &ProcessMock{
+		StartMock:   func(_ *slog.Logger) error { return nil },
+		StopMock:    func(_ *slog.Logger) error { return nil },
+		HealthyMock: func(_ *slog.Logger) bool { return true },
+	}
+
+	ps := NewProcessSupervisor(logger, "test", process, 1, 0, 2, 0).(*processSupervisor)
+	ps.eagerRestartTimeout = 100 * time.Millisecond
+	ps.firstStart.Store(true)
+
+	// Wedge one slot so the drain, which needs one on top of the slot the
+	// triggering task hands over, can never complete.
+	ps.semaphore <- struct{}{}
+
+	err := ps.Run(context.Background(), logger, func() error { return nil })
+	if err != nil {
+		t.Fatalf("expected no error but got: %v", err)
+	}
+
+	waitFor := func(what string, want bool) {
+		t.Helper()
+
+		deadline := time.Now().Add(5 * time.Second)
+		for ps.isRestarting.Load() != want {
+			if time.Now().After(deadline) {
+				t.Fatalf("timed out waiting for the eager restart to %s", what)
+			}
+			time.Sleep(5 * time.Millisecond)
+		}
+	}
+
+	waitFor("start", true)
+	waitFor("give up on its deadline", false)
+}
+
 // TestProcessSupervisor_Healthy_CachesPositiveResult verifies that a
 // successful probe is cached for [healthCheckCacheTTL] so subsequent
 // supervisor.Healthy() calls do not re-issue the underlying process