[
https://issues.apache.org/jira/browse/KAFKA-3096?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15096465#comment-15096465
]
ASF GitHub Bot commented on KAFKA-3096:
---------------------------------------
GitHub user ijuma opened a pull request:
https://github.com/apache/kafka/pull/765
KAFKA-3096; Leader is not set to -1 when it is shutdown if followers are
down
* Update leader to -1 before throwing `NoReplicaOnlineException` in
`OfflinePartitionLeaderSelector` as suggested by Guozhang. This fixes the test,
which is great, but it seems a bit out of place. Would it be better to change
the code somewhere else?
* Fix bug in `waitUntilLeaderIsElectedOrChanged` and simplify result type.
The bug was for the following case:
```scala
leader.isDefined && oldLeaderOpt.isEmpty && newLeaderOpt.isDefined &&
newLeaderOpt.get != leader.get
```
We would consider it a successful election even though we should not. I
also changed the result type as we never return `None` (we throw an exception
instead). Fixing this bug is what uncovered the leader issue being solved in
this PR.
Also included:
* Various mechanical clean-ups found while trying to understand the code
(as usual, this is in separate commits, so I can submit them separately if
desired).
* Make logging more regular in `PartitionLeaderSelector.scala`
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/ijuma/kafka
kafka-3096-leader-should-be-set-to--1
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/kafka/pull/765.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 #765
----
commit a8cbf88b37665c9669f0e9beae608794d82a810c
Author: Ismael Juma <[email protected]>
Date: 2016-01-08T16:48:47Z
Fix bug in `waitUntilLeaderIsElectedOrChanged` and simplify result type
The bug was for the following case:
`leader.isDefined && oldLeaderOpt.isEmpty && newLeaderOpt.isDefined &&
newLeaderOpt.get != leader.get`
We would consider it a successful election even though we should not.
I also changed the result type is we never return `None` (we throw an
exception instead).
commit f2266e42b1d3fc7058574f16b0c0d8de08b3a3f6
Author: Ismael Juma <[email protected]>
Date: 2016-01-13T15:23:19Z
Update leader to -1 before throwing `NoReplicaOnlineException`
Suggested by Guozhang.
commit 21568618d8ac82d3f801307b0145baa75240b0a4
Author: Ismael Juma <[email protected]>
Date: 2016-01-13T14:25:40Z
Various mechanical clean-ups
commit 7b608c44f17536274606654f6faa2cc45d890c3c
Author: Ismael Juma <[email protected]>
Date: 2016-01-13T15:46:50Z
Make logging more regular in `PartitionLeaderSelector.scala` and other
minor clean-ups
----
> Leader is not set to -1 when it is shutdown if followers are down
> -----------------------------------------------------------------
>
> Key: KAFKA-3096
> URL: https://issues.apache.org/jira/browse/KAFKA-3096
> Project: Kafka
> Issue Type: Bug
> Affects Versions: 0.9.0.0
> Reporter: Ismael Juma
> Assignee: Ismael Juma
> Fix For: 0.9.1.0
>
>
> Assuming a cluster with 2 brokers with unclear leader election disabled:
> 1. Start brokers 0 and 1
> 2. Perform partition assignment
> 3. Broker 0 is elected leader
> 4. Produce message and wait until metadata is propagated
> 6. Shutdown follower
> 7. Produce message
> 8. Shutdown leader
> 9. Start follower
> 10. Wait for leader election
> Expected: leader is -1
> Actual: leader is 0
> We have a test for this, but a bug in `waitUntilLeaderIsElectedOrChanged`
> means that `newLeaderOpt` is not being checked.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
