[KAFKA-6119] Silent Data Loss in Kafka011 Transactional Producer

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 0.11.0.0, 0.11.0.1
• Fix Version/s: 0.11.0.2, 1.0.0
• Component/s: core, producer
• Labels: reliability
• Environment: openjdk version "1.8.0_144" ; OpenJDK Runtime Environment (Zulu 8.23.0.3-macosx) (build 1.8.0_144-b01) ; OpenJDK 64-Bit Server VM (Zulu 8.23.0.3-macosx) (build 25.144-b01, mixed mode)
• Assignee: Apurva Mehta
• Reporter: Gary Y.
• Created: 25/Oct/17 21:12
• Updated: 31/Oct/17 15:52
• Resolved: 27/Oct/17 06:27

Description

Kafka can lose data published by a transactional KafkaProducer under some circumstances, i.e., data that should be committed atomically may not be fully visible from a consumer with read_committed isolation level.

Steps to reproduce:

1. Set transaction.timeout.ms to a low value such as 100
2. Publish two messages in one transaction to different partitions of a topic with a sufficiently long time in-between the messages (e.g., 70 s).
3. Only the second message is visible with read_committed isolation level.

See
https://github.com/GJL/kafka011-transactional-producer-bug-demo/blob/master/src/main/java/com/garyyao/App.java for a full example. Detailed instructions can be found in the README.md: https://github.com/GJL/kafka011-transactional-producer-bug-demo

Why is this possible?
Because the transaction timeout is set to a low value, the transaction will be rolled back quickly after the first message is sent. Indeed, in the broker the following logs could be found:

[2017-10-25 22:54:58,224] INFO [Transaction Coordinator 0]: Initialized transactionalId test-producer-1508964897483 with producerId 5 and producer epoch 0 on partition __transaction_state-10 (kafka.coordinator.transaction.TransactionCoordinator)
[2017-10-25 22:55:24,260] INFO [Transaction Coordinator 0]: Completed rollback ongoing transaction of transactionalId: test-producer-1508964897483 due to timeout (kafka.coordinator.transaction.TransactionCoordinator)

After rollback, the second message is sent to a different partition than the first message.
Upon, transaction commit, org.apache.kafka.clients.producer.internals.TransactionManager may enqueue the request addPartitionsToTransactionHandler:

private TransactionalRequestResult beginCompletingTransaction(TransactionResult transactionResult) {
        if (!newPartitionsInTransaction.isEmpty())
            enqueueRequest(addPartitionsToTransactionHandler());
        EndTxnRequest.Builder builder = new EndTxnRequest.Builder(transactionalId, producerIdAndEpoch.producerId,
                producerIdAndEpoch.epoch, transactionResult);
        EndTxnHandler handler = new EndTxnHandler(builder);
        enqueueRequest(handler);
        return handler.result;
    }

As can be seen, the condition is fulfilled if newPartitionsInTransaction is non-empty. I suspect because the second message goes to a different partition, this condition is satisfied.

In KafkaApis.scala, I can see that handleAddPartitionToTxnRequest may eventually call TransactionMetadata#prepareAddPartitions:

 def prepareAddPartitions(addedTopicPartitions: immutable.Set[TopicPartition], updateTimestamp: Long): TxnTransitMetadata = {
    val newTxnStartTimestamp = state match {
      case Empty | CompleteAbort | CompleteCommit => updateTimestamp
      case _ => txnStartTimestamp
    }

    prepareTransitionTo(Ongoing, producerId, producerEpoch, txnTimeoutMs, (topicPartitions ++ addedTopicPartitions).toSet,
      newTxnStartTimestamp, updateTimestamp)
  }

Note that the method's first argument newState of is always Ongoing here. I suspect that this puts the transaction, which should be aborted, to Ongoing again.

Issue Links

• links to: GitHub Pull Request #4137