Cannot exit and close when there are rebalancing storm

[sangreal]

Version: 0.5.2.5 (+ stale container fix (#623))

Scenario: When using, some of the health check logic on our end will trigger rebalancing storm by re-initialization. The closing is constantly interrupted by wake-up interrupt by poller. And we met with below exception.

java.lang.InterruptedException: null
    at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.reportInterruptAfterWait(AbstractQueuedSynchronizer.java:2014)
    at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.awaitNanos(AbstractQueuedSynchronizer.java:2088)
    at java.util.concurrent.ThreadPoolExecutor.awaitTermination(ThreadPoolExecutor.java:1475)
    at io.confluent.parallelconsumer.internal.AbstractParallelEoSStreamProcessor.doClose(AbstractParallelEoSStreamProcessor.java:523)
    at io.confluent.parallelconsumer.internal.AbstractParallelEoSStreamProcessor.controlLoop(AbstractParallelEoSStreamProcessor.java:706)
    at io.confluent.parallelconsumer.internal.AbstractParallelEoSStreamProcessor.lambda$supervisorLoop$5(AbstractParallelEoSStreamProcessor.java:648)
    at java.util.concurrent.FutureTask.run(FutureTask.java:266)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
    at java.lang.Thread.run(Thread.java:748)

Solution:

1. We will improve the health check and eliminate the rebalancing storm.
2. Meanwhile, the exit logic should also improve to handle these kind of scenario, as should calculate the waiting time for shutting down the workerThreadPool. Should not just reset the flag. as https://github.com/confluentinc/parallel-consumer/blob/master/parallel-consumer-core/src/main/java/io/confluent/parallelconsumer/internal/AbstractParallelEoSStreamProcessor.java#L6573.

I will draft a pr when I have time.

[rkolesnev]

Hi @sangreal - can you please elaborate a bit more - what gets interrupted by which wake-up?

Are you observing - that workerThreadPool.get().awaitTermination(toSeconds(timeout), SECONDS) is interrupted by consumerManager.wakeup() calls in BrokerPoller?

[sangreal]

Hi @rkolesnev since we were experiencing rebalancing storm. I think this should be from assigning partition. But since we don't enable debug level log. I am not sure. https://github.com/confluentinc/parallel-consumer/blob/master/parallel-consumer-core/src/main/java/io/confluent/parallelconsumer/internal/AbstractParallelEoSStreamProcessor.java#L446

[sangreal]

But it is workerThreadPool.get().awaitTermination(toSeconds(timeout), SECONDS) be interrupted.

[rkolesnev]

Ok, i see, thanks.

The interrupt / blocking thread handling always makes my head hurt a bit :) So - i guess we can either add a check to notifySomethingToDo() method - if in Closing state (or not polling mailbox right now) - don't interrupt control thread - actually it used to have that check - there is still the atomicBoolean - currentlyPollingWorkCompleteMailBox but it is no longer being checked prior to interrupting the control thread.

Alternatively - wrap the awaitTermination() call in the loop with try / catch for interrupted exception to prevent finishing the wait early.

https://github.com/confluentinc/parallel-consumer/blob/cdaf7ccabc05bacbe0f25a81bbde021ac1459827/parallel-consumer-core/src/main/java/io/confluent/parallelconsumer/internal/AbstractParallelEoSStreamProcessor.java#L1402C17-L1402C36

[sangreal]

@rkolesnev I think checking Closing state in notifySomethingToDo is the neat way to handle this. If you don't mind I will draft a pr for this one.

[rkolesnev]

Closing as #789 is merged.