MM-10502: Only cluster master should run job schedulers. (#9174)

* MM-10502: Only cluster master should run job schedulers.

* Use sync.Map for thread safety.

* Fix tests.

app/app.go

@@ -64,10 +64,11 @@ type App struct {
 	Mfa              einterfaces.MfaInterface
 	Saml             einterfaces.SamlInterface
 
-	config          atomic.Value
+	config                 atomic.Value
-	envConfig       map[string]interface{}
+	envConfig              map[string]interface{}
-	configFile      string
+	configFile             string
-	configListeners map[string]func(*model.Config, *model.Config)
+	configListeners        map[string]func(*model.Config, *model.Config)
+	clusterLeaderListeners sync.Map
 
 	licenseValue       atomic.Value
 	clientLicenseValue atomic.Value
@@ -79,14 +80,15 @@ type App struct {
 
 	newStore func() store.Store
 
-	htmlTemplateWatcher  *utils.HTMLTemplateWatcher
+	htmlTemplateWatcher     *utils.HTMLTemplateWatcher
-	sessionCache         *utils.Cache
+	sessionCache            *utils.Cache
-	configListenerId     string
+	configListenerId        string
-	licenseListenerId    string
+	licenseListenerId       string
-	logListenerId        string
+	logListenerId           string
-	disableConfigWatch   bool
+	clusterLeaderListenerId string
-	configWatcher        *utils.ConfigWatcher
+	disableConfigWatch      bool
-	asymmetricSigningKey *ecdsa.PrivateKey
+	configWatcher           *utils.ConfigWatcher
+	asymmetricSigningKey    *ecdsa.PrivateKey
 
 	pluginCommands     []*PluginCommand
 	pluginCommandsLock sync.RWMutex
@@ -218,6 +220,10 @@ func New(options ...Option) (outApp *App, outErr error) {
 		app.initJobs()
 	})
 
+	app.clusterLeaderListenerId = app.AddClusterLeaderChangedListener(func() {
+		app.Jobs.Schedulers.HandleClusterLeaderChange(app.IsLeader())
+	})
+
 	subpath, err := utils.GetSubpathFromConfig(app.Config())
 	if err != nil {
 		return nil, errors.Wrap(err, "failed to parse SiteURL subpath")
@@ -270,6 +276,7 @@ func (a *App) Shutdown() {
 	a.RemoveConfigListener(a.configListenerId)
 	a.RemoveLicenseListener(a.licenseListenerId)
 	a.RemoveConfigListener(a.logListenerId)
+	a.RemoveClusterLeaderChangedListener(a.clusterLeaderListenerId)
 	mlog.Info("Server stopped")
 
 	a.DisableConfigWatch()
@@ -432,6 +439,8 @@ func (a *App) initJobs() {
 	if jobsMigrationsInterface != nil {
 		a.Jobs.Migrations = jobsMigrationsInterface(a)
 	}
+	a.Jobs.Workers = a.Jobs.InitWorkers()
+	a.Jobs.Schedulers = a.Jobs.InitSchedulers()
 }
 
 func (a *App) DiagnosticId() string {

app/cluster.go

@@ -0,0 +1,29 @@
+// Copyright (c) 2018-present Mattermost, Inc. All Rights Reserved.
+// See License.txt for license information.
+
+package app
+
+import "github.com/mattermost/mattermost-server/model"
+
+// Registers a given function to be called when the cluster leader may have changed. Returns a unique ID for the
+// listener which can later be used to remove it. If clustering is not enabled in this build, the callback will never
+// be called.
+func (a *App) AddClusterLeaderChangedListener(listener func()) string {
+	id := model.NewId()
+	a.clusterLeaderListeners.Store(id, listener)
+	return id
+}
+
+// Removes a listener function by the unique ID returned when AddConfigListener was called
+func (a *App) RemoveClusterLeaderChangedListener(id string) {
+	a.clusterLeaderListeners.Delete(id)
+}
+
+func (a *App) InvokeClusterLeaderChangedListeners() {
+	a.Go(func() {
+		a.clusterLeaderListeners.Range(func(_, listener interface{}) bool {
+			listener.(func())()
+			return true
+		})
+	})
+}

jobs/schedulers.go

@@ -13,12 +13,13 @@ import (
 )
 
 type Schedulers struct {
-	stop          chan bool
+	stop                 chan bool
-	stopped       chan bool
+	stopped              chan bool
-	configChanged chan *model.Config
+	configChanged        chan *model.Config
-	listenerId    string
+	clusterLeaderChanged chan bool
-	startOnce     sync.Once
+	listenerId           string
-	jobs          *JobServer
+	startOnce            sync.Once
+	jobs                 *JobServer
 
 	schedulers   []model.Scheduler
 	nextRunTimes []*time.Time
@@ -28,10 +29,11 @@ func (srv *JobServer) InitSchedulers() *Schedulers {
 	mlog.Debug("Initialising schedulers.")
 
 	schedulers := &Schedulers{
-		stop:          make(chan bool),
+		stop:                 make(chan bool),
-		stopped:       make(chan bool),
+		stopped:              make(chan bool),
-		configChanged: make(chan *model.Config),
+		configChanged:        make(chan *model.Config),
-		jobs:          srv,
+		clusterLeaderChanged: make(chan bool),
+		jobs:                 srv,
 	}
 
 	if srv.DataRetentionJob != nil {
@@ -114,6 +116,14 @@ func (schedulers *Schedulers) Start() *Schedulers {
 							schedulers.setNextRunTime(newCfg, idx, now, false)
 						}
 					}
+				case isLeader := <-schedulers.clusterLeaderChanged:
+					for idx := range schedulers.schedulers {
+						if !isLeader {
+							schedulers.nextRunTimes[idx] = nil
+						} else {
+							schedulers.setNextRunTime(schedulers.jobs.Config(), idx, now, false)
+						}
+					}
 				}
 			}
 		})
@@ -171,3 +181,11 @@ func (schedulers *Schedulers) handleConfigChange(oldConfig *model.Config, newCon
 	mlog.Debug("Schedulers received config change.")
 	schedulers.configChanged <- newConfig
 }
+
+func (schedulers *Schedulers) HandleClusterLeaderChange(isLeader bool) {
+	select {
+	case schedulers.clusterLeaderChanged <- isLeader:
+	default:
+		mlog.Debug("Did not send cluster leader change message to schedulers as no schedulers listening to notification channel.")
+	}
+}

jobs/server.go

@@ -50,11 +50,11 @@ func (srv *JobServer) Config() *model.Config {
 }
 
 func (srv *JobServer) StartWorkers() {
-	srv.Workers = srv.InitWorkers().Start()
+	srv.Workers = srv.Workers.Start()
 }
 
 func (srv *JobServer) StartSchedulers() {
-	srv.Schedulers = srv.InitSchedulers().Start()
+	srv.Schedulers = srv.Schedulers.Start()
 }
 
 func (srv *JobServer) StopWorkers() {