After upgrading a fairly busy broker from 0.10.2.0 to 2.1.0, it locked up within a few minutes (by "locked up" I mean that all request handler threads were busy, and other brokers reported that they couldn't communicate with it). I restarted it a few times and it did the same thing each time. After downgrading to 0.10.2.0, the broker was stable. I attached a threaddump.txt from the last attempt on 2.1.0 that shows lots of kafka-request-handler- threads trying to acquire the leaderIsrUpdateLock lock in kafka.cluster.Partition.
It jumps out that there are two threads that already have some read lock (can't tell which one) and are trying to acquire a second one (on two different read locks: 0x0000000708184b88 and 0x000000070821f188): kafka-request-handler-1 and kafka-request-handler-4. Both are handling a produce request, and in the process of doing so, are calling Partition.fetchOffsetSnapshot while trying to complete a DelayedFetch. At the same time, both of those locks have writers from other threads waiting on them (kafka-request-handler-2 and kafka-scheduler-6). Neither of those locks appear to have writers that hold them (if only because no threads in the dump are deep enough in inWriteLock to indicate that).
ReentrantReadWriteLock in nonfair mode prioritizes waiting writers over readers. Is it possible that kafka-request-handler-1 and kafka-request-handler-4 are each trying to read-lock the partition that is currently locked by the other one, and they're both parked waiting for kafka-request-handler-2 and kafka-scheduler-6 to get write locks, which they never will, because the former two threads own read locks and aren't giving them up?