[AURORA-121] Make the preemptor more efficient

• Type: Story
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.7.0
• Component/s: Scheduler
• Labels: O3
• Sprint: Twitter Aurora Q4 Sprint 5
• Assignee: Bill Farner
• Reporter: Bill Farner
• Created: 25/Jan/14 18:13
• Updated: 26/Jan/15 22:58
• Resolved: 03/Dec/14 18:52

Description

When TaskSchedulerImpl fails to find an open slot for a task, it falls back to the preemptor:

if (!offerQueue.launchFirst(getAssignerFunction(taskId, task))) {
  // Task could not be scheduled.
  maybePreemptFor(taskId);
  return TaskSchedulerResult.TRY_AGAIN;
}

This can be problematic when the task store is large (O(10k tasks)) and there is a steady supply of PENDING tasks not satisfied by open slots. This will manifest as an overall degraded/slow scheduler, and logs of slow queries used for preemption:

I0125 17:47:36.970 THREAD23 org.apache.aurora.scheduler.storage.mem.MemTaskStore.fetchTasks: Query took 107 ms: TaskQuery(owner:null, environment:null, jobName:null,
taskIds:null, statuses:[KILLING, ASSIGNED, STARTING, RUNNING, RESTARTING], slaveHost:null, instanceIds:null)

Several approaches come to mind to improve this situation (not mutually exclusive):

• (easy) More aggressively back off on tasks that cannot be satisfied
• (easy) Fall back to preemption less frequently
• (easy) Gather the list of slaves from AttributeStore rather than TaskStore. This breaks the operation up into many smaller queries and reduces the amount of work in cases where a match is found. However, this would actually create more work when a match is not found, so this approach is probably not helpful by itself.
• (harder) Scan for preemption candidates asynchronously, freeing up the TaskScheduler thread and the storage write lock. Scans could be kicked off by the task scheduler, ideally in a way that doesn't dogpile. This could also be done in a weakly-consistent way to minimally contribute to storage contention.