[
https://issues.apache.org/jira/browse/SAMZA-1012?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15547072#comment-15547072
]
Jake Maes commented on SAMZA-1012:
----------------------------------
I looked for this when I reviewed the patch. Since we persist the mapping to
the coordinator stream, this ordering only applies in the cases where
partitions are added or the mapping is lost completely (e.g. coordinator stream
wiped). So it really just adds determinism for the case where we need to create
or amend the mapping.
> Generated changelog mappings are not consistent
> -----------------------------------------------
>
> Key: SAMZA-1012
> URL: https://issues.apache.org/jira/browse/SAMZA-1012
> Project: Samza
> Issue Type: Bug
> Affects Versions: 0.10.1
> Reporter: Tommy Becker
> Assignee: Tommy Becker
> Attachments: 0001-Fix-SAMZA-1012.patch
>
>
> Samza generates a changelog partition mapping for stateful jobs the first
> time the job is started. The mapping is then persisted in the coordinator
> stream. However, the way the mapping is generated does not guarantee the same
> mapping will always be produced for the same inputs. Although this is not
> strictly necessary (since the mapping is persisted) it can help in certain
> situations. This bit us recently when we lost the contents of a checkpoint
> topic. The job generated a new and different changelog mapping when restarted
> and began producing very strange results which took a lot of effort to track
> down.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
