fix(task): prevent RecoverStuckTasks from re-enqueueing in-flight tasks

RecoverStuckTasks scans for tasks with updated_at > 5min ago and
re-enqueues them. This incorrectly matched tasks actively being
processed by the worker (e.g. slow downloads), causing
double-processing.

Add inflight sync.Map to track taskIDs currently inside ProcessTask.
RecoverStuckTasks skips tasks found in inflight. On server restart
inflight is empty (in-memory), so genuinely stuck tasks are still
correctly recovered.

Also: increase taskCh buffer 16→10000, add periodic RecoverStuckTasks
goroutine in TaskPoller (every 5min), and add status guard in
ProcessTask as defense-in-depth against duplicate enqueues.
This commit is contained in:
dailz
2026-04-21 17:19:10 +08:00
parent 8955e513aa
commit 435ab285c1
3 changed files with 103 additions and 7 deletions

View File

@@ -8,12 +8,15 @@ import (
"go.uber.org/zap" "go.uber.org/zap"
) )
// TaskPollable defines the interface for refreshing stale task statuses. // TaskPollable defines the interface for refreshing stale task statuses
// and recovering stuck tasks.
type TaskPollable interface { type TaskPollable interface {
RefreshStaleTasks(ctx context.Context) error RefreshStaleTasks(ctx context.Context) error
RecoverStuckTasks(ctx context.Context)
} }
// TaskPoller periodically polls Slurm for task status updates via TaskPollable. // TaskPoller periodically polls Slurm for task status updates and recovers
// stuck tasks via TaskPollable.
type TaskPoller struct { type TaskPoller struct {
taskSvc TaskPollable taskSvc TaskPollable
interval time.Duration interval time.Duration
@@ -31,9 +34,11 @@ func NewTaskPoller(taskSvc TaskPollable, interval time.Duration, logger *zap.Log
} }
} }
// Start launches the background goroutine that periodically refreshes stale tasks. // Start launches background goroutines that periodically refresh stale tasks
// and recover stuck tasks.
func (p *TaskPoller) Start(ctx context.Context) { func (p *TaskPoller) Start(ctx context.Context) {
ctx, p.cancel = context.WithCancel(ctx) ctx, p.cancel = context.WithCancel(ctx)
p.wg.Add(1) p.wg.Add(1)
go func() { go func() {
defer p.wg.Done() defer p.wg.Done()
@@ -50,9 +55,24 @@ func (p *TaskPoller) Start(ctx context.Context) {
} }
} }
}() }()
p.wg.Add(1)
go func() {
defer p.wg.Done()
ticker := time.NewTicker(5 * time.Minute)
defer ticker.Stop()
for {
select {
case <-ctx.Done():
return
case <-ticker.C:
p.taskSvc.RecoverStuckTasks(ctx)
}
}
}()
} }
// Stop cancels the background goroutine and waits for it to finish. // Stop cancels the background goroutines and waits for them to finish.
func (p *TaskPoller) Stop() { func (p *TaskPoller) Stop() {
if p.cancel != nil { if p.cancel != nil {
p.cancel() p.cancel()

View File

@@ -25,6 +25,8 @@ func (m *mockTaskPollable) RefreshStaleTasks(ctx context.Context) error {
return nil return nil
} }
func (m *mockTaskPollable) RecoverStuckTasks(ctx context.Context) {}
func (m *mockTaskPollable) getCallCount() int { func (m *mockTaskPollable) getCallCount() int {
m.mu.Lock() m.mu.Lock()
defer m.mu.Unlock() defer m.mu.Unlock()

View File

@@ -29,12 +29,29 @@ type TaskService struct {
logger *zap.Logger logger *zap.Logger
// async processing // async processing
taskCh chan int64 // buffered channel, cap=16 taskCh chan int64 // buffered channel for task IDs awaiting processing
cancelFn context.CancelFunc cancelFn context.CancelFunc
wg sync.WaitGroup wg sync.WaitGroup
mu sync.Mutex // protects taskCh from send-on-closed mu sync.Mutex // protects taskCh from send-on-closed
started bool // prevent double-start started bool // prevent double-start
stopped bool stopped bool
// inflight tracks task IDs currently being processed by the worker goroutine.
//
// Why it exists: taskCh is an in-memory Go channel — all pending taskIDs are
// lost when the server restarts. RecoverStuckTasks is responsible for
// recovering those lost tasks from the DB. However, GetStuckTasks uses a
// broad query (status NOT IN completed/failed AND updated_at < 5min ago) that
// also matches tasks being actively processed by the worker (e.g. a slow
// download). Without inflight, RecoverStuckTasks would reset those tasks to
// "submitted" and re-enqueue them, causing double-processing.
//
// How it works:
// - ProcessTask stores the taskID on entry, deletes on exit (via defer).
// - RecoverStuckTasks checks inflight before re-enqueueing; in-flight tasks
// are skipped.
// - On server restart inflight is empty (in-memory), so all genuinely stuck
// tasks are correctly recovered without false negatives.
inflight sync.Map // map[int64]struct{}
} }
func NewTaskService( func NewTaskService(
@@ -56,7 +73,7 @@ func NewTaskService(
jobSvc: jobSvc, jobSvc: jobSvc,
workDirBase: workDirBase, workDirBase: workDirBase,
logger: logger, logger: logger,
taskCh: make(chan int64, 16), taskCh: make(chan int64, 10000),
} }
} }
@@ -169,6 +186,9 @@ func (s *TaskService) CreateTask(ctx context.Context, req *model.CreateTaskReque
// ProcessTask runs the full synchronous processing pipeline for a task. // ProcessTask runs the full synchronous processing pipeline for a task.
func (s *TaskService) ProcessTask(ctx context.Context, taskID int64) error { func (s *TaskService) ProcessTask(ctx context.Context, taskID int64) error {
s.inflight.Store(taskID, struct{}{})
defer s.inflight.Delete(taskID)
// 1. Fetch task // 1. Fetch task
task, err := s.taskStore.GetByID(ctx, taskID) task, err := s.taskStore.GetByID(ctx, taskID)
if err != nil { if err != nil {
@@ -178,6 +198,24 @@ func (s *TaskService) ProcessTask(ctx context.Context, taskID int64) error {
return fmt.Errorf("task %d not found", taskID) return fmt.Errorf("task %d not found", taskID)
} }
// Defense-in-depth against duplicate processing. When the same taskID enters
// taskCh multiple times (e.g. submitted normally + RecoverStuckTasks also
// enqueues it before the worker picks up the first copy), the worker processes
// them sequentially. The first invocation changes status from "submitted" to
// "preparing"; the second invocation reads the latest DB status, sees
// non-submitted, and safely skips.
//
// This does NOT block retries: processWithRetry sets status back to "submitted"
// before re-enqueueing, so the retried invocation passes this check and
// continues from the saved currentStep.
if task.Status != model.TaskStatusSubmitted {
s.logger.Debug("skipping task with non-submitted status",
zap.Int64("task_id", taskID),
zap.String("status", string(task.Status)),
)
return nil
}
fail := func(step, msg string) error { fail := func(step, msg string) error {
_ = s.taskStore.UpdateStatus(ctx, taskID, model.TaskStatusFailed, msg) _ = s.taskStore.UpdateStatus(ctx, taskID, model.TaskStatusFailed, msg)
_ = s.taskStore.UpdateRetryState(ctx, taskID, model.TaskStatusFailed, step, task.RetryCount) _ = s.taskStore.UpdateRetryState(ctx, taskID, model.TaskStatusFailed, step, task.RetryCount)
@@ -664,7 +702,7 @@ func (s *TaskService) StopProcessor() {
s.mu.Lock() s.mu.Lock()
drainCh := s.taskCh drainCh := s.taskCh
s.taskCh = make(chan int64, 16) s.taskCh = make(chan int64, 10000)
s.mu.Unlock() s.mu.Unlock()
for taskID := range drainCh { for taskID := range drainCh {
@@ -698,6 +736,36 @@ func (s *TaskService) processWithRetry(ctx context.Context, taskID int64) {
} }
func (s *TaskService) RecoverStuckTasks(ctx context.Context) { func (s *TaskService) RecoverStuckTasks(ctx context.Context) {
// RecoverStuckTasks recovers tasks that are "stuck" — they exist in the DB
// with a non-terminal status but are not being processed.
//
// Scenarios that create stuck tasks:
//
// 1. Server restart: taskCh is an in-memory Go channel, all pending IDs are
// lost on process exit. Tasks that were queued but never picked up by the
// worker remain in "submitted" status in DB with no one to process them.
//
// 2. Server crash mid-processing: the worker had advanced a task to
// "preparing"/"downloading" and then died. The task sits in that
// intermediate state with no SlurmJobID and no worker to continue.
//
// 3. Channel full: SubmitAsync dropped a task because taskCh was at
// capacity. The task stays "submitted" but was never enqueued.
//
// The bug this fix addresses:
//
// GetStuckTasks queries: status NOT IN (completed, failed) AND updated_at <
// 5min ago. This also matches tasks currently being processed by the worker
// whose step is slow (>5 min, e.g. downloading large files) and hasn't
// refreshed updated_at. Without the inflight check below, this function
// would reset such a task to "submitted" and re-enqueue it, causing the
// same task to be processed by two concurrent invocations of ProcessTask.
//
// Fix: the inflight sync.Map tracks taskIDs currently inside ProcessTask.
// Tasks found in inflight are skipped here. On server restart inflight is
// empty (it's in-memory), so all genuinely stuck tasks from scenarios 1-3
// above are correctly recovered.
tasks, err := s.taskStore.GetStuckTasks(ctx, 5*time.Minute) tasks, err := s.taskStore.GetStuckTasks(ctx, 5*time.Minute)
if err != nil { if err != nil {
s.logger.Error("failed to get stuck tasks", zap.Error(err)) s.logger.Error("failed to get stuck tasks", zap.Error(err))
@@ -711,6 +779,12 @@ func (s *TaskService) RecoverStuckTasks(ctx context.Context) {
) )
continue continue
} }
if _, ok := s.inflight.Load(tasks[i].ID); ok {
s.logger.Debug("skipping in-flight task",
zap.Int64("taskID", tasks[i].ID),
)
continue
}
_ = s.taskStore.UpdateStatus(ctx, tasks[i].ID, model.TaskStatusSubmitted, "") _ = s.taskStore.UpdateStatus(ctx, tasks[i].ID, model.TaskStatusSubmitted, "")
s.mu.Lock() s.mu.Lock()
if !s.stopped { if !s.stopped {