func (r *Orchestrator) handleTaskChange(ctx context.Context, t *api.Task) {
// If we already set the desired state past TaskStateRunning, there is no
// further action necessary.
if t.DesiredState > api.TaskStateRunning {
return
}
var (
n *api.Node
service *api.Service
)
r.store.View(func(tx store.ReadTx) {
if t.NodeID != "" {
n = store.GetNode(tx, t.NodeID)
}
if t.ServiceID != "" {
service = store.GetService(tx, t.ServiceID)
}
})
if !orchestrator.IsReplicatedService(service) {
return
}
if t.Status.State > api.TaskStateRunning ||
(t.NodeID != "" && invalidNode(n)) {
r.restartTasks[t.ID] = struct{}{}
}
}
func (r *Orchestrator) handleServiceEvent(ctx context.Context, event events.Event) {
switch v := event.(type) {
case state.EventDeleteService:
if !orchestrator.IsReplicatedService(v.Service) {
return
}
orchestrator.DeleteServiceTasks(ctx, r.store, v.Service)
r.restarts.ClearServiceHistory(v.Service.ID)
case state.EventCreateService:
if !orchestrator.IsReplicatedService(v.Service) {
return
}
r.reconcileServices[v.Service.ID] = v.Service
case state.EventUpdateService:
if !orchestrator.IsReplicatedService(v.Service) {
return
}
r.reconcileServices[v.Service.ID] = v.Service
}
}
func (r *Orchestrator) initServices(readTx store.ReadTx) error {
services, err := store.FindServices(readTx, store.All)
if err != nil {
return err
}
for _, s := range services {
if orchestrator.IsReplicatedService(s) {
r.reconcileServices[s.ID] = s
}
}
return nil
}
func (r *Orchestrator) tickTasks(ctx context.Context) {
if len(r.restartTasks) > 0 {
_, err := r.store.Batch(func(batch *store.Batch) error {
for taskID := range r.restartTasks {
err := batch.Update(func(tx store.Tx) error {
// TODO(aaronl): optimistic update?
t := store.GetTask(tx, taskID)
if t != nil {
if t.DesiredState > api.TaskStateRunning {
return nil
}
service := store.GetService(tx, t.ServiceID)
if !orchestrator.IsReplicatedService(service) {
return nil
}
// Restart task if applicable
if err := r.restarts.Restart(ctx, tx, r.cluster, service, *t); err != nil {
return err
}
}
return nil
})
if err != nil {
log.G(ctx).WithError(err).Errorf("Orchestrator task reaping transaction failed")
}
}
return nil
})
if err != nil {
log.G(ctx).WithError(err).Errorf("orchestrator task removal batch failed")
}
r.restartTasks = make(map[string]struct{})
}
}
func (r *Orchestrator) restartTasksByNodeID(ctx context.Context, nodeID string) {
var err error
r.store.View(func(tx store.ReadTx) {
var tasks []*api.Task
tasks, err = store.FindTasks(tx, store.ByNodeID(nodeID))
if err != nil {
return
}
for _, t := range tasks {
if t.DesiredState > api.TaskStateRunning {
continue
}
service := store.GetService(tx, t.ServiceID)
if orchestrator.IsReplicatedService(service) {
r.restartTasks[t.ID] = struct{}{}
}
}
})
if err != nil {
log.G(ctx).WithError(err).Errorf("failed to list tasks to remove")
}
}
func (r *Orchestrator) handleTaskEvent(ctx context.Context, event events.Event) {
switch v := event.(type) {
case state.EventDeleteNode:
r.restartTasksByNodeID(ctx, v.Node.ID)
case state.EventCreateNode:
r.handleNodeChange(ctx, v.Node)
case state.EventUpdateNode:
r.handleNodeChange(ctx, v.Node)
case state.EventDeleteTask:
if v.Task.DesiredState <= api.TaskStateRunning {
service := r.resolveService(ctx, v.Task)
if !orchestrator.IsReplicatedService(service) {
return
}
r.reconcileServices[service.ID] = service
}
r.restarts.Cancel(v.Task.ID)
case state.EventUpdateTask:
r.handleTaskChange(ctx, v.Task)
case state.EventCreateTask:
r.handleTaskChange(ctx, v.Task)
}
}
func (r *Orchestrator) initTasks(ctx context.Context, readTx store.ReadTx) error {
tasks, err := store.FindTasks(readTx, store.All)
if err != nil {
return err
}
for _, t := range tasks {
if t.NodeID != "" {
n := store.GetNode(readTx, t.NodeID)
if invalidNode(n) && t.Status.State <= api.TaskStateRunning && t.DesiredState <= api.TaskStateRunning {
r.restartTasks[t.ID] = struct{}{}
}
}
}
_, err = r.store.Batch(func(batch *store.Batch) error {
for _, t := range tasks {
if t.ServiceID == "" {
continue
}
// TODO(aluzzardi): We should NOT retrieve the service here.
service := store.GetService(readTx, t.ServiceID)
if service == nil {
// Service was deleted
err := batch.Update(func(tx store.Tx) error {
return store.DeleteTask(tx, t.ID)
})
if err != nil {
log.G(ctx).WithError(err).Error("failed to set task desired state to dead")
}
continue
}
// TODO(aluzzardi): This is shady. We should have a more generic condition.
if t.DesiredState != api.TaskStateReady || !orchestrator.IsReplicatedService(service) {
continue
}
restartDelay := orchestrator.DefaultRestartDelay
if t.Spec.Restart != nil && t.Spec.Restart.Delay != nil {
var err error
restartDelay, err = gogotypes.DurationFromProto(t.Spec.Restart.Delay)
if err != nil {
log.G(ctx).WithError(err).Error("invalid restart delay")
restartDelay = orchestrator.DefaultRestartDelay
}
}
if restartDelay != 0 {
timestamp, err := gogotypes.TimestampFromProto(t.Status.Timestamp)
if err == nil {
restartTime := timestamp.Add(restartDelay)
calculatedRestartDelay := restartTime.Sub(time.Now())
if calculatedRestartDelay < restartDelay {
restartDelay = calculatedRestartDelay
}
if restartDelay > 0 {
_ = batch.Update(func(tx store.Tx) error {
t := store.GetTask(tx, t.ID)
// TODO(aluzzardi): This is shady as well. We should have a more generic condition.
if t == nil || t.DesiredState != api.TaskStateReady {
return nil
}
r.restarts.DelayStart(ctx, tx, nil, t.ID, restartDelay, true)
return nil
})
continue
}
} else {
log.G(ctx).WithError(err).Error("invalid status timestamp")
}
}
// Start now
err := batch.Update(func(tx store.Tx) error {
return r.restarts.StartNow(tx, t.ID)
})
if err != nil {
log.G(ctx).WithError(err).WithField("task.id", t.ID).Error("moving task out of delayed state failed")
}
}
return nil
})
return err
}
// Restart initiates a new task to replace t if appropriate under the service's
// restart policy.
func (r *Supervisor) Restart(ctx context.Context, tx store.Tx, cluster *api.Cluster, service *api.Service, t api.Task) error {
// TODO(aluzzardi): This function should not depend on `service`.
// Is the old task still in the process of restarting? If so, wait for
// its restart delay to elapse, to avoid tight restart loops (for
// example, when the image doesn't exist).
r.mu.Lock()
oldDelay, ok := r.delays[t.ID]
if ok {
if !oldDelay.waiter {
oldDelay.waiter = true
go r.waitRestart(ctx, oldDelay, cluster, t.ID)
}
r.mu.Unlock()
return nil
}
r.mu.Unlock()
// Sanity check: was the task shut down already by a separate call to
// Restart? If so, we must avoid restarting it, because this will create
// an extra task. This should never happen unless there is a bug.
if t.DesiredState > api.TaskStateRunning {
return errors.New("Restart called on task that was already shut down")
}
t.DesiredState = api.TaskStateShutdown
err := store.UpdateTask(tx, &t)
if err != nil {
log.G(ctx).WithError(err).Errorf("failed to set task desired state to dead")
return err
}
if !r.shouldRestart(ctx, &t, service) {
return nil
}
var restartTask *api.Task
if orchestrator.IsReplicatedService(service) {
restartTask = orchestrator.NewTask(cluster, service, t.Slot, "")
} else if orchestrator.IsGlobalService(service) {
restartTask = orchestrator.NewTask(cluster, service, 0, t.NodeID)
} else {
log.G(ctx).Error("service not supported by restart supervisor")
return nil
}
n := store.GetNode(tx, t.NodeID)
restartTask.DesiredState = api.TaskStateReady
var restartDelay time.Duration
// Restart delay is not applied to drained nodes
if n == nil || n.Spec.Availability != api.NodeAvailabilityDrain {
if t.Spec.Restart != nil && t.Spec.Restart.Delay != nil {
var err error
restartDelay, err = ptypes.Duration(t.Spec.Restart.Delay)
if err != nil {
log.G(ctx).WithError(err).Error("invalid restart delay; using default")
restartDelay = orchestrator.DefaultRestartDelay
}
} else {
restartDelay = orchestrator.DefaultRestartDelay
}
}
waitStop := true
// Normally we wait for the old task to stop running, but we skip this
// if the old task is already dead or the node it's assigned to is down.
if (n != nil && n.Status.State == api.NodeStatus_DOWN) || t.Status.State > api.TaskStateRunning {
waitStop = false
}
if err := store.CreateTask(tx, restartTask); err != nil {
log.G(ctx).WithError(err).WithField("task.id", restartTask.ID).Error("task create failed")
return err
}
r.recordRestartHistory(restartTask)
r.DelayStart(ctx, tx, &t, restartTask.ID, restartDelay, waitStop)
return nil
}
