[KAFKA-5427] Transactional producer cannot find coordinator when trying to abort transaction after error

• Type: Sub-task
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Fix Version/s: 0.11.0.0
• Component/s: clients, core, producer
• Labels: exactly-once
• Assignee: Jason Gustafson
• Reporter: Jason Gustafson
• Created: 10/Jun/17 22:24
• Updated: 12/Jun/17 22:07
• Resolved: 12/Jun/17 22:06

Description

It can happen that we receive an abortable error while we are already aborting a transaction. In this case, we have an EndTxnRequest queued for sending when we transition to ABORTABLE_ERROR. It could be that we need to find the coordinator before sending this EndTxnRequest. The problem is that we will fail even the FindCoordinatorRequest because we are in an error state. This causes the following endless loop:

[2017-06-10 19:29:33,436] DEBUG [TransactionalId my-first-transactional-id] Enqueuing transactional request (type=FindCoordinatorRequest, coordinatorKey=my-fi
rst-transactional-id, coordinatorType=TRANSACTION) (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,436] DEBUG [TransactionalId my-first-transactional-id] Enqueuing transactional request (type=EndTxnRequest, transactionalId=my-first-tran
sactional-id, producerId=1000, producerEpoch=0, result=ABORT) (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,536] TRACE [TransactionalId my-first-transactional-id] Not sending transactional request (type=FindCoordinatorRequest, coordinatorKey=my-
first-transactional-id, coordinatorType=TRANSACTION) because we are in an error state (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,637] TRACE [TransactionalId my-first-transactional-id] Request (type=EndTxnRequest, transactionalId=my-first-transactional-id, producerId
=1000, producerEpoch=0, result=ABORT) dequeued for sending (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,637] DEBUG [TransactionalId my-first-transactional-id] Enqueuing transactional request (type=FindCoordinatorRequest, coordinatorKey=my-fi
rst-transactional-id, coordinatorType=TRANSACTION) (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,637] DEBUG [TransactionalId my-first-transactional-id] Enqueuing transactional request (type=EndTxnRequest, transactionalId=my-first-tran
sactional-id, producerId=1000, producerEpoch=0, result=ABORT) (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,737] TRACE [TransactionalId my-first-transactional-id] Not sending transactional request (type=FindCoordinatorRequest, coordinatorKey=my-
first-transactional-id, coordinatorType=TRANSACTION) because we are in an error state (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,837] TRACE [TransactionalId my-first-transactional-id] Request (type=EndTxnRequest, transactionalId=my-first-transactional-id, producerId
=1000, producerEpoch=0, result=ABORT) dequeued for sending (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,838] DEBUG [TransactionalId my-first-transactional-id] Enqueuing transactional request (type=FindCoordinatorRequest, coordinatorKey=my-fi
rst-transactional-id, coordinatorType=TRANSACTION) (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,838] DEBUG [TransactionalId my-first-transactional-id] Enqueuing transactional request (type=EndTxnRequest, transactionalId=my-first-tran
sactional-id, producerId=1000, producerEpoch=0, result=ABORT) (org.apache.kafka.clients.producer.internals.TransactionManager)
[2017-06-10 19:29:33,938] TRACE [TransactionalId my-first-transactional-id] Not sending transactional request (type=FindCoordinatorRequest, coordinatorKey=my-
first-transactional-id, coordinatorType=TRANSACTION) because we are in an error state (org.apache.kafka.clients.producer.internals.TransactionManager)

A couple suggested improvements:

1. We should allow FindCoordinator requests regardless of the transaction state.
2. It is a bit confusing that we allow EndTxnRequest to be sent in both the ABORTABLE_ERROR and the ABORTING_TRANSACTION states. Perhaps we should only allow EndTxnRequest to be sent in ABORTING_TRANSACTION. If we hit an abortable error and we are already aborting, then we should just stay in ABORTING_TRANSACTION and perhaps log a warning.

Issue Links

• links to: GitHub Pull Request #3297