[
https://issues.apache.org/jira/browse/KAFKA-1074?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13826647#comment-13826647
]
Jun Rao commented on KAFKA-1074:
--------------------------------
Removing the old replica log from the disk itself is simple. What we need to
make sure is that all potential outstanding operations on the deleted log are
handled properly. In particular, we don't want any potential IOException due to
the missing log causes the broker to halt.
1. All read operations are ok since we already handle unexpected exceptions in
KafkaApi.
2. Writing to the log by the producer request, the replica fetcher or the log
flusher: We need to make sure that after the log is deleted, no more
writes/flushes will be attempted on the log. This can be achieved by:
2.1 For producer requests, the delete partition operation will synchronize on
the leaderAndIsrUpdate lock.
2.2 For replica fetcher, this is already handled since the fetcher is removed
before the log is deleted.
2.3 For log flusher, the flush and the delete will now synchronize on a delete
lock.
> Reassign partitions should delete the old replicas from disk
> ------------------------------------------------------------
>
> Key: KAFKA-1074
> URL: https://issues.apache.org/jira/browse/KAFKA-1074
> Project: Kafka
> Issue Type: Bug
> Components: core
> Affects Versions: 0.8
> Reporter: Jun Rao
> Fix For: 0.8.1
>
> Attachments: KAFKA-1074.patch
>
>
> Currently, after reassigning replicas to other brokers, the old replicas are
> not removed from disk and have to be deleted manually.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
