fix: Fix task getting stuck after recovery (#42114)

Submit tasks into the global scheduler after recovery.

issue: https://github.com/milvus-io/milvus/issues/42046

---------

Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
This commit is contained in:
yihao.dai 2025-05-28 12:46:28 +08:00 committed by GitHub
parent 08a53c56b1
commit 79b51cbb73
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
9 changed files with 107 additions and 7 deletions

View File

@ -51,8 +51,9 @@ func (ai *analyzeInspector) Stop() {
func (ai *analyzeInspector) reloadFromMeta() { func (ai *analyzeInspector) reloadFromMeta() {
analyzeTasks := ai.mt.analyzeMeta.GetAllTasks() analyzeTasks := ai.mt.analyzeMeta.GetAllTasks()
for _, t := range analyzeTasks { for _, t := range analyzeTasks {
if t.GetState() == indexpb.JobState_JobStateFinished || if t.GetState() != indexpb.JobState_JobStateInit &&
t.GetState() == indexpb.JobState_JobStateFailed { t.GetState() != indexpb.JobState_JobStateRetry &&
t.GetState() != indexpb.JobState_JobStateInProgress {
continue continue
} }
ai.scheduler.Enqueue(newAnalyzeTask( ai.scheduler.Enqueue(newAnalyzeTask(

View File

@ -553,6 +553,7 @@ func (c *compactionInspector) submitTask(t CompactionTask) error {
func (c *compactionInspector) restoreTask(t CompactionTask) { func (c *compactionInspector) restoreTask(t CompactionTask) {
c.executingGuard.Lock() c.executingGuard.Lock()
c.executingTasks[t.GetTaskProto().GetPlanID()] = t c.executingTasks[t.GetTaskProto().GetPlanID()] = t
c.scheduler.Enqueue(t)
c.executingGuard.Unlock() c.executingGuard.Unlock()
metrics.DataCoordCompactionTaskNum.WithLabelValues(fmt.Sprintf("%d", t.GetTaskProto().GetNodeID()), t.GetTaskProto().GetType().String(), metrics.Executing).Inc() metrics.DataCoordCompactionTaskNum.WithLabelValues(fmt.Sprintf("%d", t.GetTaskProto().GetNodeID()), t.GetTaskProto().GetType().String(), metrics.Executing).Inc()
} }

View File

@ -75,6 +75,7 @@ func (s *CompactionPlanHandlerSuite) TestScheduleEmpty() {
} }
func (s *CompactionPlanHandlerSuite) generateInitTasksForSchedule() { func (s *CompactionPlanHandlerSuite) generateInitTasksForSchedule() {
s.handler.scheduler.(*task.MockGlobalScheduler).EXPECT().Enqueue(mock.Anything).Return()
task1 := &mixCompactionTask{ task1 := &mixCompactionTask{
meta: s.mockMeta, meta: s.mockMeta,
} }
@ -407,6 +408,8 @@ func (s *CompactionPlanHandlerSuite) TestRemoveTasksByChannel() {
s.SetupTest() s.SetupTest()
ch := "ch1" ch := "ch1"
s.handler.scheduler.(*task.MockGlobalScheduler).EXPECT().Enqueue(mock.Anything).Return()
t1 := newMixCompactionTask(&datapb.CompactionTask{ t1 := newMixCompactionTask(&datapb.CompactionTask{
PlanID: 19530, PlanID: 19530,
Type: datapb.CompactionType_MixCompaction, Type: datapb.CompactionType_MixCompaction,

View File

@ -62,6 +62,7 @@ func NewImportInspector(ctx context.Context, meta *meta, importMeta ImportMeta,
} }
func (s *importInspector) Start() { func (s *importInspector) Start() {
s.reloadFromMeta()
log.Ctx(s.ctx).Info("start import inspector") log.Ctx(s.ctx).Info("start import inspector")
ticker := time.NewTicker(Params.DataCoordCfg.ImportScheduleInterval.GetAsDuration(time.Second)) ticker := time.NewTicker(Params.DataCoordCfg.ImportScheduleInterval.GetAsDuration(time.Second))
defer ticker.Stop() defer ticker.Stop()
@ -82,6 +83,21 @@ func (s *importInspector) Close() {
}) })
} }
func (s *importInspector) reloadFromMeta() {
jobs := s.importMeta.GetJobBy(s.ctx)
sort.Slice(jobs, func(i, j int) bool {
return jobs[i].GetJobID() < jobs[j].GetJobID()
})
for _, job := range jobs {
tasks := s.importMeta.GetTaskBy(s.ctx, WithJob(job.GetJobID()))
for _, task := range tasks {
if task.GetState() == datapb.ImportTaskStateV2_InProgress {
s.scheduler.Enqueue(task)
}
}
}
}
func (s *importInspector) inspect() { func (s *importInspector) inspect() {
jobs := s.importMeta.GetJobBy(s.ctx) jobs := s.importMeta.GetJobBy(s.ctx)
sort.Slice(jobs, func(i, j int) bool { sort.Slice(jobs, func(i, j int) bool {

View File

@ -258,6 +258,77 @@ func (s *ImportInspectorSuite) TestProcessFailed() {
s.Equal(0, len(task.(*importTask).GetSegmentIDs())) s.Equal(0, len(task.(*importTask).GetSegmentIDs()))
} }
func (s *ImportInspectorSuite) TestReloadFromMeta() {
// Test case 1: No jobs and tasks
s.catalog.EXPECT().ListImportJobs(mock.Anything).Return(nil, nil)
s.catalog.EXPECT().ListPreImportTasks(mock.Anything).Return(nil, nil)
s.catalog.EXPECT().ListImportTasks(mock.Anything).Return(nil, nil)
s.inspector.reloadFromMeta()
// Test case 2: Jobs with in-progress tasks
jobProto := &datapb.ImportJob{
JobID: 1,
CollectionID: s.collectionID,
TimeoutTs: math.MaxUint64,
Schema: &schemapb.CollectionSchema{},
}
job := &importJob{
ImportJob: jobProto,
tr: timerecord.NewTimeRecorder("import job"),
}
s.catalog.EXPECT().SaveImportJob(mock.Anything, mock.Anything).Return(nil)
err := s.importMeta.AddJob(context.TODO(), job)
s.NoError(err)
// Add an in-progress pre-import task
inprogressPreImportTask := &preImportTask{
importMeta: s.importMeta,
tr: timerecord.NewTimeRecorder("preimport task"),
}
inprogressPreImportTask.task.Store(&datapb.PreImportTask{
JobID: 1,
TaskID: 1,
CollectionID: s.collectionID,
State: datapb.ImportTaskStateV2_InProgress,
})
s.catalog.EXPECT().SavePreImportTask(mock.Anything, mock.Anything).Return(nil)
err = s.importMeta.AddTask(context.TODO(), inprogressPreImportTask)
s.NoError(err)
// Add an in-progress import task
inprogressImportTask := &importTask{
importMeta: s.importMeta,
tr: timerecord.NewTimeRecorder("import task"),
}
inprogressImportTask.task.Store(&datapb.ImportTaskV2{
JobID: 1,
TaskID: 2,
CollectionID: s.collectionID,
State: datapb.ImportTaskStateV2_InProgress,
})
s.catalog.EXPECT().SaveImportTask(mock.Anything, mock.Anything).Return(nil)
err = s.importMeta.AddTask(context.TODO(), inprogressImportTask)
s.NoError(err)
// Add an pending import task
pendingImportTask := &importTask{
importMeta: s.importMeta,
tr: timerecord.NewTimeRecorder("import task"),
}
pendingImportTask.task.Store(&datapb.ImportTaskV2{
JobID: 1,
TaskID: 3,
CollectionID: s.collectionID,
State: datapb.ImportTaskStateV2_Pending,
})
s.catalog.EXPECT().SaveImportTask(mock.Anything, mock.Anything).Return(nil)
err = s.importMeta.AddTask(context.TODO(), pendingImportTask)
// Mock scheduler expectations
s.inspector.scheduler.(*task2.MockGlobalScheduler).EXPECT().Enqueue(mock.Anything).Times(2)
s.inspector.reloadFromMeta()
}
func TestImportInspector(t *testing.T) { func TestImportInspector(t *testing.T) {
suite.Run(t, new(ImportInspectorSuite)) suite.Run(t, new(ImportInspectorSuite))
} }

View File

@ -200,8 +200,9 @@ func (i *indexInspector) reloadFromMeta() {
segments := i.meta.GetAllSegmentsUnsafe() segments := i.meta.GetAllSegmentsUnsafe()
for _, segment := range segments { for _, segment := range segments {
for _, segIndex := range i.meta.indexMeta.GetSegmentIndexes(segment.GetCollectionID(), segment.ID) { for _, segIndex := range i.meta.indexMeta.GetSegmentIndexes(segment.GetCollectionID(), segment.ID) {
if segIndex.IsDeleted || segIndex.IndexState == commonpb.IndexState_Finished || if segIndex.IsDeleted || (segIndex.IndexState != commonpb.IndexState_Unissued &&
segIndex.IndexState == commonpb.IndexState_Failed { segIndex.IndexState != commonpb.IndexState_Retry &&
segIndex.IndexState != commonpb.IndexState_InProgress) {
continue continue
} }

View File

@ -98,8 +98,9 @@ func (si *statsInspector) Stop() {
func (si *statsInspector) reloadFromMeta() { func (si *statsInspector) reloadFromMeta() {
tasks := si.mt.statsTaskMeta.GetAllTasks() tasks := si.mt.statsTaskMeta.GetAllTasks()
for _, st := range tasks { for _, st := range tasks {
if st.GetState() == indexpb.JobState_JobStateFinished || if st.GetState() != indexpb.JobState_JobStateInit &&
st.GetState() == indexpb.JobState_JobStateFailed { st.GetState() != indexpb.JobState_JobStateRetry &&
st.GetState() != indexpb.JobState_JobStateInProgress {
continue continue
} }
segment := si.mt.GetHealthySegment(si.ctx, st.GetSegmentID()) segment := si.mt.GetHealthySegment(si.ctx, st.GetSegmentID())

View File

@ -69,10 +69,11 @@ func (s *globalTaskScheduler) Enqueue(task Task) {
case taskcommon.Init: case taskcommon.Init:
task.SetTaskTime(taskcommon.TimeQueue, time.Now()) task.SetTaskTime(taskcommon.TimeQueue, time.Now())
s.pendingTasks.Push(task) s.pendingTasks.Push(task)
case taskcommon.InProgress: case taskcommon.InProgress, taskcommon.Retry:
task.SetTaskTime(taskcommon.TimeStart, time.Now()) task.SetTaskTime(taskcommon.TimeStart, time.Now())
s.runningTasks.Insert(task.GetTaskID(), task) s.runningTasks.Insert(task.GetTaskID(), task)
} }
log.Ctx(s.ctx).Info("task enqueued", WrapTaskLog(task)...)
} }
func (s *globalTaskScheduler) AbortAndRemoveTask(taskID int64) { func (s *globalTaskScheduler) AbortAndRemoveTask(taskID int64) {

View File

@ -41,6 +41,7 @@ func TestGlobalScheduler_Enqueue(t *testing.T) {
task := NewMockTask(t) task := NewMockTask(t)
task.EXPECT().GetTaskID().Return(1) task.EXPECT().GetTaskID().Return(1)
task.EXPECT().GetTaskState().Return(taskcommon.Init) task.EXPECT().GetTaskState().Return(taskcommon.Init)
task.EXPECT().GetTaskType().Return(taskcommon.Compaction)
task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return() task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return()
scheduler.Enqueue(task) scheduler.Enqueue(task)
assert.Equal(t, 1, len(scheduler.(*globalTaskScheduler).pendingTasks.TaskIDs())) assert.Equal(t, 1, len(scheduler.(*globalTaskScheduler).pendingTasks.TaskIDs()))
@ -50,6 +51,7 @@ func TestGlobalScheduler_Enqueue(t *testing.T) {
task = NewMockTask(t) task = NewMockTask(t)
task.EXPECT().GetTaskID().Return(2) task.EXPECT().GetTaskID().Return(2)
task.EXPECT().GetTaskState().Return(taskcommon.InProgress) task.EXPECT().GetTaskState().Return(taskcommon.InProgress)
task.EXPECT().GetTaskType().Return(taskcommon.Compaction)
task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return() task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return()
scheduler.Enqueue(task) scheduler.Enqueue(task)
assert.Equal(t, 1, scheduler.(*globalTaskScheduler).runningTasks.Len()) assert.Equal(t, 1, scheduler.(*globalTaskScheduler).runningTasks.Len())
@ -64,6 +66,7 @@ func TestGlobalScheduler_AbortAndRemoveTask(t *testing.T) {
task := NewMockTask(t) task := NewMockTask(t)
task.EXPECT().GetTaskID().Return(1) task.EXPECT().GetTaskID().Return(1)
task.EXPECT().GetTaskState().Return(taskcommon.Init) task.EXPECT().GetTaskState().Return(taskcommon.Init)
task.EXPECT().GetTaskType().Return(taskcommon.Compaction)
task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return() task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return()
task.EXPECT().DropTaskOnWorker(mock.Anything).Return() task.EXPECT().DropTaskOnWorker(mock.Anything).Return()
scheduler.Enqueue(task) scheduler.Enqueue(task)
@ -74,6 +77,7 @@ func TestGlobalScheduler_AbortAndRemoveTask(t *testing.T) {
task = NewMockTask(t) task = NewMockTask(t)
task.EXPECT().GetTaskID().Return(2) task.EXPECT().GetTaskID().Return(2)
task.EXPECT().GetTaskState().Return(taskcommon.InProgress) task.EXPECT().GetTaskState().Return(taskcommon.InProgress)
task.EXPECT().GetTaskType().Return(taskcommon.Compaction)
task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return() task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return()
task.EXPECT().DropTaskOnWorker(mock.Anything).Return() task.EXPECT().DropTaskOnWorker(mock.Anything).Return()
scheduler.Enqueue(task) scheduler.Enqueue(task)
@ -140,6 +144,7 @@ func TestGlobalScheduler_TestSchedule(t *testing.T) {
task.EXPECT().GetTaskID().Return(1) task.EXPECT().GetTaskID().Return(1)
task.EXPECT().GetTaskType().Return(taskcommon.Compaction) task.EXPECT().GetTaskType().Return(taskcommon.Compaction)
task.EXPECT().GetTaskState().Return(taskcommon.Init) task.EXPECT().GetTaskState().Return(taskcommon.Init)
task.EXPECT().GetTaskType().Return(taskcommon.Compaction)
task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return() task.EXPECT().SetTaskTime(mock.Anything, mock.Anything).Return()
task.EXPECT().GetTaskSlot().Return(1) task.EXPECT().GetTaskSlot().Return(1)
return task return task