Major Description
When a connector or task tries to read the offsets from the offsets topic, it issues `OffsetStorageImpl#offsets` method. This method gets a Future from the underneath KafkaBackingStore. KafkaBackingStore invokes `KafkaBasedLog#readToEnd` method and passes the Callback. This method essentially adds the Callback to a Queue of callbacks that are being managed.
Within KafkaBasedLog, there's a WorkThread which keeps polling over the callback queue and executes them and it does this in an infinite loop. However, there is an enclosing try/catch block around the while loop. If there's an exception thrown which is not caught by any of the other catch blocks, the control goes to the outermost catch block and the WorkThread is terminated. However, the connectors/tasks are not aware of this and they would keep submitting callbacks to KafkaBasedLog with nobody processing them. This can be seen in the thread dumps as well:
"task-thread-connector-0" #6334 prio=5 os_prio=0 cpu=19.36ms elapsed=2092.93s tid=0x00007f8d9c037000 nid=0x5d00 waiting on condition [0x00007f8dc08cd000] java.lang.Thread.State: WAITING (parking) at jdk.internal.misc.Unsafe.park(java.base@11.0.15/Native Method) - parking to wait for <0x000000070345c9a8> (a java.util.concurrent.CountDownLatch$Sync) at java.util.concurrent.locks.LockSupport.park(java.base@11.0.15/LockSupport.java:194) at java.util.concurrent.locks.AbstractQueuedSynchronizer.parkAndCheckInterrupt(java.base@11.0.15/AbstractQueuedSynchronizer.java:885) at java.util.concurrent.locks.AbstractQueuedSynchronizer.doAcquireSharedInterruptibly(java.base@11.0.15/AbstractQueuedSynchronizer.java:1039) at java.util.concurrent.locks.AbstractQueuedSynchronizer.acquireSharedInterruptibly(java.base@11.0.15/AbstractQueuedSynchronizer.java:1345) at java.util.concurrent.CountDownLatch.await(java.base@11.0.15/CountDownLatch.java:232) at org.apache.kafka.connect.util.ConvertingFutureCallback.get(ConvertingFutureCallback.java:98) at org.apache.kafka.connect.storage.OffsetStorageReaderImpl.offsets(OffsetStorageReaderImpl.java:101) at org.apache.kafka.connect.storage.OffsetStorageReaderImpl.offset(OffsetStorageReaderImpl.java:63)
We need a mechanism to fail all such offset read requests. That is because even if we restart the thread, chances are it will still fail with the same error so the offset fetch would be stuck perennially.
As already explained, this scenario happens mainly when the exception thrown is such that it isn't caught by any of the catch blocks and the control lands up in the outermost catch block. In my experience, I have seen this situation happening on a few occasions, when the exception thrown is:
[2022-11-20 09:00:59,307] ERROR Unexpected exception in Thread[KafkaBasedLog Work Thread - connect-offsets,5,main] (org.apache.kafka.connect.util.KafkaBasedLog:440)org.apache.kafka.connect.errors.ConnectException: Error while getting end offsets for topic 'connect-offsets' on brokers at XXX at org.apache.kafka.connect.util.TopicAdmin.endOffsets(TopicAdmin.java:695) at org.apache.kafka.connect.util.KafkaBasedLog.readEndOffsets(KafkaBasedLog.java:371) at org.apache.kafka.connect.util.KafkaBasedLog.readToLogEnd(KafkaBasedLog.java:332) at org.apache.kafka.connect.util.KafkaBasedLog.access$400(KafkaBasedLog.java:75) at org.apache.kafka.connect.util.KafkaBasedLog$WorkThread.run(KafkaBasedLog.java:406) Caused by: java.util.concurrent.ExecutionException: org.apache.kafka.common.errors.SslAuthenticationException: SSL handshake failed at org.apache.kafka.common.internals.KafkaFutureImpl.wrapAndThrow(KafkaFutureImpl.java:45) at org.apache.kafka.common.internals.KafkaFutureImpl.access$000(KafkaFutureImpl.java:32) at org.apache.kafka.common.internals.KafkaFutureImpl$SingleWaiter.await(KafkaFutureImpl.java:89) at org.apache.kafka.common.internals.KafkaFutureImpl.get(KafkaFutureImpl.java:260) at org.apache.kafka.connect.util.TopicAdmin.endOffsets(TopicAdmin.java:672) ... 4 more
At this point, the WorkThread is dead once the control goes out of the catch block
and we can find the following line `Unexpected exception in` in the logs.
Another example could be when the worker is already OOM and in such cases as well the work thread would die. This is not a good example because once the worker is OOM, we can't make any progress anyways but adding this example for brevity's sake.