[
https://issues.apache.org/jira/browse/KUDU-3191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17199832#comment-17199832
]
ASF subversion and git services commented on KUDU-3191:
-------------------------------------------------------
Commit fcceb8b1a20afff30e15b6248a56ab3e06b61e79 in kudu's branch
refs/heads/master from Andrew Wong
[ https://gitbox.apache.org/repos/asf?p=kudu.git;h=fcceb8b ]
KUDU-3191: fail replicas when KUDU-2233 is detected
Despite the longstanding fixes that stop bad KUDU-2233 compactions,
users still see the results of already corrupted data, particularly when
upgrading to newer versions that may compact more aggressively than
older versions.
Rather than crashing when hitting a KUDU-2233 failure, this patch
updates the behavior to fail the replica. Similar to disk failures or
CFile checksum corruption, this will trigger re-replication to happen,
and eviction will only happen if there is a healthy majority.
The hope is that fewer users will see this corruption cause problems, as
the corruption will henceforth not crash servers, and only tablets with
a majority corrupted will be unavailable.
Change-Id: I43570b961dfd5eb8518328121585255d32cf2ebb
Reviewed-on: http://gerrit.cloudera.org:8080/16471
Tested-by: Kudu Jenkins
Reviewed-by: Alexey Serbin <aser...@cloudera.com>
> Fail tablet replicas that suffer from KUDU-2233 instead of crashing
> -------------------------------------------------------------------
>
> Key: KUDU-3191
> URL: https://issues.apache.org/jira/browse/KUDU-3191
> Project: Kudu
> Issue Type: Task
> Components: compaction
> Reporter: Andrew Wong
> Assignee: Andrew Wong
> Priority: Major
>
> KUDU-2233 results in persisted corruption that causes a broken invariant,
> leading to a server crash. The recovery process for this corruption is
> arduous, especially if there are multiple tablet replicas in a given server
> that suffer from it -- users typically start the server, see the crash,
> remove the affected replica manually via tooling, and restart, repeatedly
> until the server comes up healthily.
> Instead, we should consider treating this as we do CFile block-level
> corruption[1] and fail the tablet replica. At best, we end up recovering from
> a non-corrupted replica. At worst, we'd end up with multiple corrupted
> replicas, which is still better than what we have today, which is multiple
> corrupted replicas and unavailable servers that lead to excessive
> re-replication.
> [1]
> https://github.com/apache/kudu/commit/cf6927cb153f384afb649b664de1d4276bd6d83f
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
