[
https://issues.apache.org/jira/browse/KAFKA-1633?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14142124#comment-14142124
]
Jun Rao commented on KAFKA-1633:
--------------------------------
If both brokers are shut down normally, there shouldn't be any data loss since
during the normal shutdown, we force unflushed data to disk.
The hard killing case is probably a bit different. Depending on the OS and file
system, unflushed data could be lost during a hard kill. If that's the case,
when both brokers are hard killed, some previously acked messages (but not
flushed yet) could be lost.
> Data loss if broker is killed
> -----------------------------
>
> Key: KAFKA-1633
> URL: https://issues.apache.org/jira/browse/KAFKA-1633
> Project: Kafka
> Issue Type: Bug
> Components: producer
> Affects Versions: 0.8.1.1
> Environment: centos 6.3, open jdk 7
> Reporter: gautham varada
> Assignee: Jun Rao
>
> We have a 2 node kafka cluster, we experienced data loss when we did a kill
> -9 on the brokers. We also found a work around to prevent this loss.
> Replication factor :2, 4 partitions
> Steps to reproduce
> 1. Create a 2 node cluster with replication factor 2, num partitions 4
> 2. We used Jmeter to pump events
> 3. We used kafka web console to inspect the log size after the test
> During the test, we simultaneously killed the brokers using kill -9 and we
> tallied the metrics reported by jmeter and the size we observed in the web
> console, we lost tons of messages.
> We went back and set the Producer retry to 1 instead of the default 3 and
> repeated the above tests and we did not loose a single message.
> We repeated the above tests with the Producer retry set to 3 and 1 with a
> single broker and we observed data loss when the retry was 3 and no loss when
> the retry was 1
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
