fix(v2): avoid starvation of failed jobs in the compaction scheduler (#3732)

Author: kolesnikovae

pkg/experiment/metastore/compaction/compaction.go

@@ -55,5 +55,6 @@ type Schedule interface {
 	// AssignJob is called on behalf of the worker to request a new job.
 	AssignJob() (*raft_log.AssignedCompactionJob, error)
 	// AddJob is called on behalf of the planner to add a new job to the schedule.
+	// The scheduler may decline the job by returning a nil state.
 	AddJob(*raft_log.CompactionJobPlan) *raft_log.CompactionJobState
 }

pkg/experiment/metastore/compaction/compactor/compactor_strategy.go

@@ -30,7 +30,7 @@ func DefaultStrategy() Strategy {
 		MaxBlocksPerLevel:  []uint{20, 10, 10},
 		MaxBlocksDefault:   10,
 		MaxLevel:           3,
-		MaxBatchAge:        3 * time.Minute.Nanoseconds(), //defaultMaxBlockBatchAge,
+		MaxBatchAge:        defaultMaxBlockBatchAge,
 		CleanupBatchSize:   2,
 		CleanupDelay:       15 * time.Minute,
 		CleanupJobMaxLevel: 1,

pkg/experiment/metastore/compaction/scheduler/schedule.go

@@ -21,7 +21,8 @@ type schedule struct {
 	// Read-only.
 	scheduler *Scheduler
 	// Uncommitted schedule updates.
-	updates map[string]*raft_log.CompactionJobState
+	updates   map[string]*raft_log.CompactionJobState
+	addedJobs int
 	// Modified copy of the job queue.
 	copied []priorityJobQueue
 	level  int
@@ -90,10 +91,29 @@ func (p *schedule) newStateForStatusReport(status *raft_log.CompactionJobStatusU
 	return nil
 }
 
-// AddJob creates a state for the new plan. The method must be called
-// after the last AssignJob and UpdateJob calls.
+// AddJob creates a state for the newly planned job.
+//
+// The method must be called after the last AssignJob and UpdateJob calls.
+// It returns an empty state if the queue size limit is reached.
+//
+// TODO(kolesnikovae): Implement displacement policy.
+// When the scheduler queue is full, no new jobs can be added. Currently,
+// it's possible that all jobs fail and can't be retried, and consequently,
+// can't leave the queue, blocking the entire compaction process until the
+// failure or queue limit is increased. Additionally, it's possible for a
+// job to never be completed and thus remain in the queue indefinitely.
+//
+// One way to implement this is to evict the job with the highest number of
+// failures (exceeding a configurable threshold, in addition to MaxFailures).
+// This way, we can easily remove the job least likely to succeed.
+// However, this needs to be handled explicitly in UpdateSchedule; at this
+// point, we can only identify candidates for eviction.
 func (p *schedule) AddJob(plan *raft_log.CompactionJobPlan) *raft_log.CompactionJobState {
-	// TODO(kolesnikovae): Job queue size limit.
+	if limit := p.scheduler.config.MaxQueueSize; limit > 0 {
+		if size := uint64(p.addedJobs + p.scheduler.queue.size()); size >= limit {
+			return nil
+		}
+	}
 	state := &raft_log.CompactionJobState{
 		Name:            plan.Name,
 		CompactionLevel: plan.CompactionLevel,
@@ -102,13 +122,19 @@ func (p *schedule) AddJob(plan *raft_log.CompactionJobPlan) *raft_log.Compaction
 		Token:           p.token,
 	}
 	p.updates[state.Name] = state
+	p.addedJobs++
 	return state
 }
 
 func (p *schedule) nextAssignment() *raft_log.CompactionJobState {
 	// We don't need to check the job ownership here: the worker asks
 	// for a job assigment (new ownership).
 	for p.level < len(p.scheduler.queue.levels) {
+		// We evict the job from our copy of the queue: each job is only
+		// accessible once. When we reach the bottom of the queue (the first
+		// failed job, or the last job in the queue), we move to the next
+		// level. Note that we check all in-progress jobs if there are not
+		// enough unassigned jobs in the queue.
 		pq := p.queueLevelCopy(p.level)
 		if pq.Len() == 0 {
 			p.level++
@@ -130,15 +156,17 @@ func (p *schedule) nextAssignment() *raft_log.CompactionJobState {
 			return p.assignJob(job)
 
 		case metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS:
-			if p.shouldReassign(job) {
+			if p.isFailed(job) {
+				// We reached the bottom of the queue: only failed jobs left.
+				p.level++
+				continue
+			}
+			if p.isAbandoned(job) {
 				state := p.assignJob(job)
 				state.Failures++
 				return state
 			}
 		}
-
-		// If no jobs can be assigned at this level.
-		p.level++
 	}
 
 	return nil
@@ -156,11 +184,13 @@ func (p *schedule) assignJob(e *jobEntry) *raft_log.CompactionJobState {
 	return job
 }
 
-func (p *schedule) shouldReassign(job *jobEntry) bool {
-	abandoned := p.now.UnixNano() > job.LeaseExpiresAt
+func (p *schedule) isAbandoned(job *jobEntry) bool {
+	return p.now.UnixNano() > job.LeaseExpiresAt
+}
+
+func (p *schedule) isFailed(job *jobEntry) bool {
 	limit := p.scheduler.config.MaxFailures
-	faulty := limit > 0 && uint64(job.Failures) >= limit
-	return abandoned && !faulty
+	return limit > 0 && uint64(job.Failures) >= limit
 }
 
 // The queue must not be modified by the assigner. Therefore, we're copying the

pkg/experiment/metastore/compaction/scheduler/schedule_test.go

@@ -123,10 +123,12 @@ func TestSchedule_Assign(t *testing.T) {
 	}
 
 	scheduler := NewScheduler(config, store, nil)
+	// The job plans are accessed when it's getting assigned.
+	// Their content is not important for the test.
 	plans := []*raft_log.CompactionJobPlan{
-		{Name: "2", Tenant: "A", Shard: 1, CompactionLevel: 0, SourceBlocks: []string{"d", "e", "f"}},
-		{Name: "3", Tenant: "A", Shard: 1, CompactionLevel: 0, SourceBlocks: []string{"j", "h", "i"}},
-		{Name: "1", Tenant: "A", Shard: 1, CompactionLevel: 1, SourceBlocks: []string{"a", "b", "c"}},
+		{Name: "2", CompactionLevel: 0},
+		{Name: "3", CompactionLevel: 0},
+		{Name: "1", CompactionLevel: 1},
 	}
 	for _, p := range plans {
 		store.On("GetJobPlan", mock.Anything, p.Name).Return(p, nil)
@@ -169,37 +171,55 @@ func TestSchedule_ReAssign(t *testing.T) {
 
 	scheduler := NewScheduler(config, store, nil)
 	plans := []*raft_log.CompactionJobPlan{
-		{Name: "1", Tenant: "A", Shard: 1, SourceBlocks: []string{"a", "b", "c"}},
-		{Name: "2", Tenant: "A", Shard: 1, SourceBlocks: []string{"d", "e", "f"}},
-		{Name: "3", Tenant: "A", Shard: 1, SourceBlocks: []string{"j", "h", "i"}},
+		{Name: "1"},
+		{Name: "2"},
+		{Name: "3"},
+		{Name: "4"},
+		{Name: "5"},
+		{Name: "6"},
 	}
 	for _, p := range plans {
 		store.On("GetJobPlan", mock.Anything, p.Name).Return(p, nil)
 	}
 
+	now := int64(5)
 	states := []*raft_log.CompactionJobState{
-		{Name: "1", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 0},
-		{Name: "2", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 0},
-		{Name: "3", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 0},
+		// Jobs with expired leases (now > LeaseExpiresAt).
+		{Name: "1", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 1},
+		{Name: "2", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 1},
+		{Name: "3", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 1},
+		// This job can't be reassigned as its lease is still valid.
+		{Name: "4", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 10},
+		// The job has already failed in the past.
+		{Name: "5", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 1, Failures: 1},
+		// The job has already failed in the past and exceeded the error threshold.
+		{Name: "6", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 1, LeaseExpiresAt: 1, Failures: 3},
 	}
 	for _, s := range states {
 		scheduler.queue.put(s)
 	}
 
+	lease := now + int64(config.LeaseDuration)
+	expected := []*raft_log.CompactionJobState{
+		{Name: "1", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 2, LeaseExpiresAt: lease, Failures: 1},
+		{Name: "2", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 2, LeaseExpiresAt: lease, Failures: 1},
+		{Name: "3", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 2, LeaseExpiresAt: lease, Failures: 1},
+		{Name: "5", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, Token: 2, LeaseExpiresAt: lease, Failures: 2},
+	}
+
 	test.AssertIdempotent(t, func(t *testing.T) {
-		s := scheduler.NewSchedule(nil, &raft.Log{Index: 2, AppendedAt: time.Unix(0, 1)})
-		for j := range plans {
+		s := scheduler.NewSchedule(nil, &raft.Log{Index: 2, AppendedAt: time.Unix(0, now)})
+		assigned := make([]*raft_log.CompactionJobState, 0, len(expected))
+		for {
 			update, err := s.AssignJob()
 			require.NoError(t, err)
-			assert.Equal(t, plans[j], update.Plan)
-			assert.Equal(t, metastorev1.CompactionJobStatus_COMPACTION_STATUS_IN_PROGRESS, update.State.Status)
-			assert.Equal(t, int64(config.LeaseDuration)+1, update.State.LeaseExpiresAt)
-			assert.Equal(t, uint64(2), update.State.Token)
+			if update == nil {
+				break
+			}
+			assigned = append(assigned, update.State)
 		}
 
-		update, err := s.AssignJob()
-		require.NoError(t, err)
-		assert.Nil(t, update)
+		assert.Equal(t, expected, assigned)
 	})
 }
 
@@ -212,9 +232,9 @@ func TestSchedule_UpdateAssign(t *testing.T) {
 
 	scheduler := NewScheduler(config, store, nil)
 	plans := []*raft_log.CompactionJobPlan{
-		{Name: "1", Tenant: "A", Shard: 1, SourceBlocks: []string{"a", "b", "c"}},
-		{Name: "2", Tenant: "A", Shard: 1, SourceBlocks: []string{"d", "e", "f"}},
-		{Name: "3", Tenant: "A", Shard: 1, SourceBlocks: []string{"j", "h", "i"}},
+		{Name: "1"},
+		{Name: "2"},
+		{Name: "3"},
 	}
 	for _, p := range plans {
 		store.On("GetJobPlan", mock.Anything, p.Name).Return(p, nil)
@@ -301,9 +321,9 @@ func TestSchedule_Add(t *testing.T) {
 
 	scheduler := NewScheduler(config, store, nil)
 	plans := []*raft_log.CompactionJobPlan{
-		{Name: "1", Tenant: "A", Shard: 1, SourceBlocks: []string{"a", "b", "c"}},
-		{Name: "2", Tenant: "A", Shard: 1, SourceBlocks: []string{"d", "e", "f"}},
-		{Name: "3", Tenant: "A", Shard: 1, SourceBlocks: []string{"j", "h", "i"}},
+		{Name: "1"},
+		{Name: "2"},
+		{Name: "3"},
 	}
 
 	states := []*raft_log.CompactionJobState{
@@ -319,3 +339,31 @@ func TestSchedule_Add(t *testing.T) {
 		}
 	})
 }
+
+func TestSchedule_QueueSizeLimit(t *testing.T) {
+	store := new(mockscheduler.MockJobStore)
+	config := Config{
+		MaxFailures:   3,
+		LeaseDuration: 10 * time.Second,
+		MaxQueueSize:  2,
+	}
+
+	scheduler := NewScheduler(config, store, nil)
+	plans := []*raft_log.CompactionJobPlan{
+		{Name: "1"},
+		{Name: "2"},
+		{Name: "3"},
+	}
+
+	states := []*raft_log.CompactionJobState{
+		{Name: "1", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_UNSPECIFIED, AddedAt: 1, Token: 1},
+		{Name: "2", Status: metastorev1.CompactionJobStatus_COMPACTION_STATUS_UNSPECIFIED, AddedAt: 1, Token: 1},
+	}
+
+	test.AssertIdempotent(t, func(t *testing.T) {
+		s := scheduler.NewSchedule(nil, &raft.Log{Index: 1, AppendedAt: time.Unix(0, 1)})
+		assert.Equal(t, states[0], s.AddJob(plans[0]))
+		assert.Equal(t, states[1], s.AddJob(plans[1]))
+		assert.Nil(t, s.AddJob(plans[2]))
+	})
+}

pkg/experiment/metastore/compaction/scheduler/scheduler.go

@@ -45,11 +45,13 @@ type JobStore interface {
 type Config struct {
 	MaxFailures   uint64        `yaml:"compaction_max_failures" doc:""`
 	LeaseDuration time.Duration `yaml:"compaction_job_lease_duration" doc:""`
+	MaxQueueSize  uint64        `yaml:"compaction_max_job_queue_size" doc:""`
 }
 
 func (c *Config) RegisterFlagsWithPrefix(prefix string, f *flag.FlagSet) {
 	f.Uint64Var(&c.MaxFailures, prefix+"compaction-max-failures", 3, "")
 	f.DurationVar(&c.LeaseDuration, prefix+"compaction-job-lease-duration", 15*time.Second, "")
+	f.Uint64Var(&c.MaxQueueSize, prefix+"compaction-max-job-queue-size", 2000, "")
 }
 
 type Scheduler struct {

pkg/experiment/metastore/compaction/scheduler/scheduler_queue.go

@@ -45,6 +45,16 @@ func (q *schedulerQueue) delete(name string) *raft_log.CompactionJobState {
 	return nil
 }
 
+func (q *schedulerQueue) size() int {
+	var size int
+	for _, level := range q.levels {
+		if level != nil {
+			size += level.jobs.Len()
+		}
+	}
+	return size
+}
+
 func (q *schedulerQueue) level(x uint32) *jobQueue {
 	s := x + 1 // Levels are 0-based.
 	if s >= uint32(len(q.levels)) {
@@ -136,7 +146,9 @@ func compareJobs(a, b *jobEntry) int {
 	if a.Status != b.Status {
 		return int(a.Status) - int(b.Status)
 	}
-	// Faulty jobs should wait.
+	// Faulty jobs should wait. Our aim is to put them at the
+	// end of the queue, after all the jobs we may consider
+	// for assigment.
 	if a.Failures != b.Failures {
 		return int(a.Failures) - int(b.Failures)
 	}

pkg/experiment/metastore/compaction_raft_handler.go

@@ -117,8 +117,14 @@ func (h *CompactionCommandHandler) GetCompactionPlanUpdate(
 			// No more jobs to create.
 			break
 		}
+		state := scheduler.AddJob(plan)
+		if state == nil {
+			// Scheduler declined the job. The only case when this may happen
+			// is when the scheduler queue is full.
+			break
+		}
 		p.NewJobs = append(p.NewJobs, &raft_log.NewCompactionJob{
-			State: scheduler.AddJob(plan),
+			State: state,
 			Plan:  plan,
 		})
 	}