Description
Right now, the OverseerCollectionProcessor is single threaded i.e submitting anything long running would have it block processing of other mutually exclusive tasks.
When OCP tasks become optionally async (SOLR-5477), it'd be good to have truly non-blocking behavior by multi-threading the OCP itself.
For example, a ShardSplit call on Collection1 would block the thread and thereby, not processing a create collection task (which would stay queued in zk) though both the tasks are mutually exclusive.
Here are a few of the challenges:
- Mutual exclusivity: Only let mutually exclusive tasks run in parallel. An easy way to handle that is to only let 1 task per collection run at a time.
- ZK Distributed Queue to feed tasks: The OCP consumes tasks from a queue. The task from the workQueue is only removed on completion so that in case of a failure, the new Overseer can re-consume the same task and retry. A queue is not the right data structure in the first place to look ahead i.e. get the 2nd task from the queue when the 1st one is in process. Also, deleting tasks which are not at the head of a queue is not really an 'intuitive' thing.
Proposed solutions for task management:
- Task funnel and peekAfter(): The parent thread is responsible for getting and passing the request to a new thread (or one from the pool). The parent method uses a peekAfter(last element) instead of a peek(). The peekAfter returns the task after the 'last element'. Maintain this request information and use it for deleting/cleaning up the workQueue.
- Another (almost duplicate) queue: While offering tasks to workQueue, also offer them to a new queue (call it volatileWorkQueue?). The difference is, as soon as a task from this is picked up for processing by the thread, it's removed from the queue. At the end, the cleanup is done from the workQueue.