[
https://issues.apache.org/jira/browse/KAFKA-4473?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15744756#comment-15744756
]
ASF GitHub Bot commented on KAFKA-4473:
---------------------------------------
GitHub user dguy opened a pull request:
https://github.com/apache/kafka/pull/2249
KAFKA-4473: RecordCollector should handle retriable exceptions more strictly
The `RecordCollectorImpl` currently drops messages on the floor if an
exception is non-null in the producer callback. This will result in message
loss and violates at-least-once processing.
Rather than just log an error in the callback, save the exception in a
field. On subsequent calls to `send`, `flush`, `close`, first check for the
existence of an exception and throw a `StreamsException` if it is non-null.
Also, in the callback, if an exception has already occurred, the `offsets` map
should not be updated.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/dguy/kafka kafka-4473
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/kafka/pull/2249.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #2249
----
commit 9ceaa67a9967edfa7522bb9daea64c1bc44738c0
Author: Damian Guy <damian....@gmail.com>
Date: 2016-12-12T17:47:27Z
throw exception in record collector instead of just dropping messages on
the floor
----
> RecordCollector should handle retriable exceptions more strictly
> ----------------------------------------------------------------
>
> Key: KAFKA-4473
> URL: https://issues.apache.org/jira/browse/KAFKA-4473
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Affects Versions: 0.10.1.0
> Reporter: Thomas Schulz
> Assignee: Damian Guy
> Priority: Critical
> Labels: architecture
>
> see: https://groups.google.com/forum/#!topic/confluent-platform/DT5bk1oCVk8
> There is probably a bug in the RecordCollector as described in my detailed
> Cluster test published in the aforementioned post.
> The class RecordCollector has the following behavior:
> - if there is no exception, add the message offset to a map
> - otherwise, do not add the message offset and instead log the above statement
> Is it possible that this offset map contains the latest offset to commit? If
> so, a message that fails might be overriden be a successful (later) message
> and the consumer commits every message up to the latest offset?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
