Backport of Multiple instances of a periodic job are run simultaneously, when prohibit_overlap is true into release/1.5.x

nomad/leader.go

@@ -774,16 +774,46 @@ func (s *Server) restorePeriodicDispatcher() error {
 			continue
 		}
 
+		// We skip if the job doesn't allow overlap and there are already
+		// instances running
+		needed, err := s.isNewEvalNeeded(job)
+		if err != nil {
+			return fmt.Errorf("failed to get job status: %v", err)
+		}
+		if !needed {
+			continue
+		}
+
 		if _, err := s.periodicDispatcher.ForceEval(job.Namespace, job.ID); err != nil {
 			logger.Error("force run of periodic job failed", "job", job.NamespacedID(), "error", err)
 			return fmt.Errorf("force run of periodic job %q failed: %v", job.NamespacedID(), err)
 		}
-		logger.Debug("periodic job force runned during leadership establishment", "job", job.NamespacedID())
+
+		logger.Debug("periodic job force run during leadership establishment", "job", job.NamespacedID())
 	}
 
 	return nil
 }
 
+// isNewEvalNeeded checks if the job allows for overlap and if there are already
+// instances of the job running in order to determine if a new evaluation needs to
+// be created upon periodic dispatcher restore
+func (s *Server) isNewEvalNeeded(job *structs.Job) (bool, error) {
+
+	if job.Periodic.ProhibitOverlap {
+		running, err := s.periodicDispatcher.dispatcher.RunningChildren(job)
+		if err != nil {
+			return false, fmt.Errorf("failed to determine if periodic job has running children %q error %q", job.NamespacedID(), err)
+		}
+
+		if running {
+			return false, nil
+		}
+	}
+
+	return true, nil
+}
+
 // schedulePeriodic is used to do periodic job dispatch while we are leader
 func (s *Server) schedulePeriodic(stopCh chan struct{}) {
 	evalGC := time.NewTicker(s.config.EvalGCInterval)

nomad/leader_test.go

@@ -411,7 +411,7 @@ func TestLeader_PeriodicDispatcher_Restore_NoEvals(t *testing.T) {
 	now := time.Now()
 
 	// Sleep till after the job should have been launched.
-	time.Sleep(3 * time.Second)
+	time.Sleep(5 * time.Second)
 
 	// Restore the periodic dispatcher.
 	s1.periodicDispatcher.SetEnabled(true)
@@ -438,13 +438,35 @@ func TestLeader_PeriodicDispatcher_Restore_NoEvals(t *testing.T) {
 	}
 }
 
+type mockJobEvalDispatcher struct {
+	forceEvalCalled, children bool
+	evalToReturn              *structs.Evaluation
+	JobEvalDispatcher
+}
+
+func (mjed *mockJobEvalDispatcher) DispatchJob(_ *structs.Job) (*structs.Evaluation, error) {
+	mjed.forceEvalCalled = true
+	return mjed.evalToReturn, nil
+}
+
+func (mjed *mockJobEvalDispatcher) RunningChildren(_ *structs.Job) (bool, error) {
+	return mjed.children, nil
+}
+
+func testPeriodicJob_OverlapEnabled(times ...time.Time) *structs.Job {
+	job := testPeriodicJob(times...)
+	job.Periodic.ProhibitOverlap = true
+	return job
+}
+
 func TestLeader_PeriodicDispatcher_Restore_Evals(t *testing.T) {
 	ci.Parallel(t)
 
 	s1, cleanupS1 := TestServer(t, func(c *Config) {
 		c.NumSchedulers = 0
 	})
 	defer cleanupS1()
+
 	testutil.WaitForLeader(t, s1.RPC)
 
 	// Inject a periodic job that triggered once in the past, should trigger now
@@ -465,7 +487,15 @@ func TestLeader_PeriodicDispatcher_Restore_Evals(t *testing.T) {
 	}
 
 	// Create an eval for the past launch.
-	s1.periodicDispatcher.createEval(job, past)
+	eval, _ := s1.periodicDispatcher.createEval(job, past)
+
+	md := &mockJobEvalDispatcher{
+		children:          false,
+		evalToReturn:      eval,
+		JobEvalDispatcher: s1,
+	}
+
+	s1.periodicDispatcher.dispatcher = md
 
 	// Flush the periodic dispatcher, ensuring that no evals will be created.
 	s1.periodicDispatcher.SetEnabled(false)
@@ -475,6 +505,7 @@ func TestLeader_PeriodicDispatcher_Restore_Evals(t *testing.T) {
 
 	// Restore the periodic dispatcher.
 	s1.periodicDispatcher.SetEnabled(true)
+
 	s1.restorePeriodicDispatcher()
 
 	// Ensure the job is tracked.
@@ -495,6 +526,133 @@ func TestLeader_PeriodicDispatcher_Restore_Evals(t *testing.T) {
 	if last.Launch == past {
 		t.Fatalf("restorePeriodicDispatcher did not force launch")
 	}
+
+	must.True(t, md.forceEvalCalled, must.Sprint("failed to force job evaluation"))
+}
+
+func TestLeader_PeriodicDispatcher_No_Overlaps_No_Running_Job(t *testing.T) {
+	ci.Parallel(t)
+
+	s1, cleanupS1 := TestServer(t, func(c *Config) {
+		c.NumSchedulers = 0
+	})
+	defer cleanupS1()
+	testutil.WaitForLeader(t, s1.RPC)
+
+	// Inject a periodic job that triggered once in the past, should trigger now
+	// and once in the future.
+	now := time.Now()
+	past := now.Add(-1 * time.Second)
+	future := now.Add(10 * time.Second)
+
+	job := testPeriodicJob_OverlapEnabled(past, now, future)
+	req := structs.JobRegisterRequest{
+		Job: job,
+		WriteRequest: structs.WriteRequest{
+			Namespace: job.Namespace,
+		},
+	}
+	_, _, err := s1.raftApply(structs.JobRegisterRequestType, req)
+	must.NoError(t, err)
+
+	// Create an eval for the past launch.
+	eval, _ := s1.periodicDispatcher.createEval(job, past)
+
+	md := &mockJobEvalDispatcher{
+		children:     false,
+		evalToReturn: eval,
+	}
+
+	s1.periodicDispatcher.dispatcher = md
+
+	// Flush the periodic dispatcher, ensuring that no evals will be created.
+	s1.periodicDispatcher.SetEnabled(false)
+
+	// Sleep till after the job should have been launched.
+	time.Sleep(3 * time.Second)
+
+	// Restore the periodic dispatcher.
+	s1.periodicDispatcher.SetEnabled(true)
+	must.NoError(t, s1.restorePeriodicDispatcher())
+
+	// Ensure the job is tracked.
+	tuple := structs.NamespacedID{
+		ID:        job.ID,
+		Namespace: job.Namespace,
+	}
+	must.MapContainsKey(t, s1.periodicDispatcher.tracked, tuple, must.Sprint("periodic job not restored"))
+
+	// Check that an eval was made.
+	ws := memdb.NewWatchSet()
+	last, err := s1.fsm.State().PeriodicLaunchByID(ws, job.Namespace, job.ID)
+	must.NoError(t, err)
+	must.NotNil(t, last)
+
+	if last.Launch == past {
+		t.Fatalf("restorePeriodicDispatcher did not force launch")
+	}
+
+	if md.forceEvalCalled != true {
+		t.Fatalf("failed to force job evaluation")
+	}
+	t.Fail()
+}
+
+func TestLeader_PeriodicDispatcher_No_Overlaps_Running_Job(t *testing.T) {
+	ci.Parallel(t)
+
+	s1, cleanupS1 := TestServer(t, func(c *Config) {
+		c.NumSchedulers = 0
+	})
+	defer cleanupS1()
+	testutil.WaitForLeader(t, s1.RPC)
+
+	// Inject a periodic job that triggered once in the past, should trigger now
+	// and once in the future.
+	now := time.Now()
+	past := now.Add(-1 * time.Second)
+	future := now.Add(10 * time.Second)
+
+	job := testPeriodicJob_OverlapEnabled(past, now, future)
+	req := structs.JobRegisterRequest{
+		Job: job,
+		WriteRequest: structs.WriteRequest{
+			Namespace: job.Namespace,
+		},
+	}
+	_, _, err := s1.raftApply(structs.JobRegisterRequestType, req)
+	if err != nil {
+		t.Fatalf("err: %v", err)
+	}
+
+	// Create an eval for the past launch.
+	eval, _ := s1.periodicDispatcher.createEval(job, past)
+
+	md := &mockJobEvalDispatcher{
+		children:     true,
+		evalToReturn: eval,
+	}
+
+	s1.periodicDispatcher.dispatcher = md
+
+	// Flush the periodic dispatcher, ensuring that no evals will be created.
+	s1.periodicDispatcher.SetEnabled(false)
+
+	// Sleep till after the job should have been launched.
+	time.Sleep(3 * time.Second)
+
+	// Restore the periodic dispatcher.
+	s1.periodicDispatcher.SetEnabled(true)
+	must.NoError(t, s1.restorePeriodicDispatcher())
+
+	// Ensure the job is tracked.
+	tuple := structs.NamespacedID{
+		ID:        job.ID,
+		Namespace: job.Namespace,
+	}
+	must.MapContainsKey(t, s1.periodicDispatcher.tracked, tuple, must.Sprint("periodic job not restored"))
+
+	must.False(t, md.forceEvalCalled, must.Sprint("evaluation forced with job already running"))
 }
 
 func TestLeader_PeriodicDispatch(t *testing.T) {

nomad/periodic.go

@@ -278,10 +278,10 @@ func (p *PeriodicDispatch) removeLocked(jobID structs.NamespacedID) error {
 // subsequent eval.
 func (p *PeriodicDispatch) ForceEval(namespace, jobID string) (*structs.Evaluation, error) {
 	p.l.Lock()
+	defer p.l.Unlock()
 
 	// Do nothing if not enabled
 	if !p.enabled {
-		p.l.Unlock()
 		return nil, fmt.Errorf("periodic dispatch disabled")
 	}
 
@@ -291,11 +291,9 @@ func (p *PeriodicDispatch) ForceEval(namespace, jobID string) (*structs.Evaluati
 	}
 	job, tracked := p.tracked[tuple]
 	if !tracked {
-		p.l.Unlock()
 		return nil, fmt.Errorf("can't force run non-tracked job %q (%s)", jobID, namespace)
 	}
 
-	p.l.Unlock()
 	return p.createEval(job, time.Now().In(job.Periodic.GetLocation()))
 }