/**
* Plan for some additional workers to be launched.
*
* @param numWorkers The number of workers to allocate.
*/
@Override
protected void requestNewWorkers(int numWorkers) {
try {
List<TaskMonitor.TaskGoalStateUpdated> toMonitor = new ArrayList<>(numWorkers);
List<LaunchableTask> toLaunch = new ArrayList<>(numWorkers);
// generate new workers into persistent state and launch associated actors
for (int i = 0; i < numWorkers; i++) {
MesosWorkerStore.Worker worker = MesosWorkerStore.Worker.newWorker(workerStore.newTaskID());
workerStore.putWorker(worker);
workersInNew.put(extractResourceID(worker.taskID()), worker);
LaunchableMesosWorker launchable = createLaunchableMesosWorker(worker.taskID());
LOG.info(
"Scheduling Mesos task {} with ({} MB, {} cpus).",
launchable.taskID().getValue(),
launchable.taskRequest().getMemory(),
launchable.taskRequest().getCPUs());
toMonitor.add(new TaskMonitor.TaskGoalStateUpdated(extractGoalState(worker)));
toLaunch.add(launchable);
}
// tell the task router about the new plans
for (TaskMonitor.TaskGoalStateUpdated update : toMonitor) {
taskRouter.tell(update, self());
}
// tell the launch coordinator to launch the new tasks
if (toLaunch.size() >= 1) {
launchCoordinator.tell(new LaunchCoordinator.Launch(toLaunch), self());
}
} catch (Exception ex) {
fatalError("unable to request new workers", ex);
}
}
/**
* Accept offers as advised by the launch coordinator.
*
* <p>Acceptance is routed through the RM to update the persistent state before forwarding the
* message to Mesos.
*/
private void acceptOffers(AcceptOffers msg) {
try {
List<TaskMonitor.TaskGoalStateUpdated> toMonitor = new ArrayList<>(msg.operations().size());
// transition the persistent state of some tasks to Launched
for (Protos.Offer.Operation op : msg.operations()) {
if (op.getType() != Protos.Offer.Operation.Type.LAUNCH) {
continue;
}
for (Protos.TaskInfo info : op.getLaunch().getTaskInfosList()) {
MesosWorkerStore.Worker worker = workersInNew.remove(extractResourceID(info.getTaskId()));
assert (worker != null);
worker = worker.launchWorker(info.getSlaveId(), msg.hostname());
workerStore.putWorker(worker);
workersInLaunch.put(extractResourceID(worker.taskID()), worker);
LOG.info(
"Launching Mesos task {} on host {}.",
worker.taskID().getValue(),
worker.hostname().get());
toMonitor.add(new TaskMonitor.TaskGoalStateUpdated(extractGoalState(worker)));
}
}
// tell the task router about the new plans
for (TaskMonitor.TaskGoalStateUpdated update : toMonitor) {
taskRouter.tell(update, self());
}
// send the acceptance message to Mesos
schedulerDriver.acceptOffers(msg.offerIds(), msg.operations(), msg.filters());
} catch (Exception ex) {
fatalError("unable to accept offers", ex);
}
}
/**
* Extracts the Mesos task goal state from the worker information.
*
* @param worker the persistent worker information.
* @return goal state information for the {@Link TaskMonitor}.
*/
static TaskMonitor.TaskGoalState extractGoalState(MesosWorkerStore.Worker worker) {
switch (worker.state()) {
case New:
return new TaskMonitor.New(worker.taskID());
case Launched:
return new TaskMonitor.Launched(worker.taskID(), worker.slaveID().get());
case Released:
return new TaskMonitor.Released(worker.taskID(), worker.slaveID().get());
default:
throw new IllegalArgumentException("unsupported worker state");
}
}
/** Recover framework/worker information persisted by a prior incarnation of the RM. */
private void recoverWorkers() throws Exception {
// if this application master starts as part of an ApplicationMaster/JobManager recovery,
// then some worker tasks are most likely still alive and we can re-obtain them
final List<MesosWorkerStore.Worker> tasksFromPreviousAttempts = workerStore.recoverWorkers();
if (!tasksFromPreviousAttempts.isEmpty()) {
LOG.info("Retrieved {} TaskManagers from previous attempt", tasksFromPreviousAttempts.size());
List<Tuple2<TaskRequest, String>> toAssign =
new ArrayList<>(tasksFromPreviousAttempts.size());
List<LaunchableTask> toLaunch = new ArrayList<>(tasksFromPreviousAttempts.size());
for (final MesosWorkerStore.Worker worker : tasksFromPreviousAttempts) {
LaunchableMesosWorker launchable = createLaunchableMesosWorker(worker.taskID());
switch (worker.state()) {
case New:
workersInNew.put(extractResourceID(worker.taskID()), worker);
toLaunch.add(launchable);
break;
case Launched:
workersInLaunch.put(extractResourceID(worker.taskID()), worker);
toAssign.add(new Tuple2<>(launchable.taskRequest(), worker.hostname().get()));
break;
case Released:
workersBeingReturned.put(extractResourceID(worker.taskID()), worker);
break;
}
taskRouter.tell(new TaskMonitor.TaskGoalStateUpdated(extractGoalState(worker)), self());
}
// tell the launch coordinator about prior assignments
if (toAssign.size() >= 1) {
launchCoordinator.tell(new LaunchCoordinator.Assign(toAssign), self());
}
// tell the launch coordinator to launch any new tasks
if (toLaunch.size() >= 1) {
launchCoordinator.tell(new LaunchCoordinator.Launch(toLaunch), self());
}
}
}
/** Plan for the removal of the given worker. */
private void releaseWorker(MesosWorkerStore.Worker worker) {
try {
LOG.info("Releasing worker {}", worker.taskID());
// update persistent state of worker to Released
worker = worker.releaseWorker();
workerStore.putWorker(worker);
workersBeingReturned.put(extractResourceID(worker.taskID()), worker);
taskRouter.tell(new TaskMonitor.TaskGoalStateUpdated(extractGoalState(worker)), self());
if (worker.hostname().isDefined()) {
// tell the launch coordinator that the task is being unassigned from the host, for planning
// purposes
launchCoordinator.tell(
new LaunchCoordinator.Unassign(worker.taskID(), worker.hostname().get()), self());
}
} catch (Exception ex) {
fatalError("unable to release worker", ex);
}
}
