[
https://issues.apache.org/jira/browse/KAFKA-966?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13703755#comment-13703755
]
Joel Koshy commented on KAFKA-966:
----------------------------------
One way to accomplish this is to turn off autocommit and checkpoint offsets
only after a message (or batch of messages) have been written to the DB.
One caveat though is that rebalances (e.g., if a new consumer instance shows
up) will result in offsets being committed so there would be an issue if the DB
is unavailable and a rebalance occurs simultaneously and there are unprocessed
messages that have already been pulled out of the iterator.
> Allow high level consumer to 'nak' a message and force Kafka to close the
> KafkaStream without losing that message
> -----------------------------------------------------------------------------------------------------------------
>
> Key: KAFKA-966
> URL: https://issues.apache.org/jira/browse/KAFKA-966
> Project: Kafka
> Issue Type: Improvement
> Components: consumer
> Affects Versions: 0.8
> Reporter: Chris Curtin
> Assignee: Neha Narkhede
> Priority: Minor
>
> Enhancement request.
> The high level consumer is very close to handling a lot of situations a
> 'typical' client would need. Except for when the message received from Kafka
> is valid, but the business logic that wants to consume it has a problem.
> For example if I want to write the value to a MongoDB or Cassandra database
> and the database is not available. I won't know until I go to do the write
> that the database isn't available, but by then it is too late to NOT read the
> message from Kafka. Thus if I call shutdown() to stop reading, that message
> is lost since the offset Kafka writes to ZooKeeper is the next offset.
> Ideally I'd like to be able to tell Kafka: close the KafkaStream but set the
> next offset to read for this partition to this message when I start up again.
> And if there are any messages in the BlockingQueue for other partitions, find
> the lowest # and use it for that partitions offset since I haven't consumed
> them yet.
> Thus I can cleanly shutdown my processing, resolve whatever the issue is and
> restart the process.
> Another idea might be to allow a 'peek' into the next message and if I
> succeed in writing to the database call 'next' to remove it from the queue.
> I understand this won't deal with a 'kill -9' or hard failure of the JVM
> leading to the latest offsets not being written to ZooKeeper but it addresses
> a likely common scenario for consumers. Nor will it add true transactional
> support since the ZK update could fail.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
