[
https://issues.apache.org/jira/browse/CASSANDRA-8805?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14569384#comment-14569384
]
Benedict commented on CASSANDRA-8805:
-------------------------------------
The rollback logic in the event of CIE is, unfortunately, likely to break
things. Once a reader replaces another, it needs to be published or itself
replaced. So we should either make visible the newly sampled readers (excluding
the ones we haven't completed), or we should construct a new instance with the
old summaries that are publised.
Since the operation can also be very expensive, we should probably extract the
resampling into a method that can itself check for the {{isStopRequested}},
otherwise it could take a long time to acquiesce.
> runWithCompactionsDisabled only cancels compactions, which is not the only
> source of markCompacted
> --------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-8805
> URL: https://issues.apache.org/jira/browse/CASSANDRA-8805
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: Benedict
> Assignee: Carl Yeksigian
> Fix For: 2.1.x
>
> Attachments: 8805-2.1.txt
>
>
> Operations like repair that may operate over all sstables cancel compactions
> before beginning, and fail if there are any files marked compacting after
> doing so. Redistribution of index summaries is not a compaction, so is not
> cancelled by this action, but does mark sstables as compacting, so such an
> action will fail to initiate if there is an index summary redistribution in
> progress. It seems that IndexSummaryManager needs to register itself as
> interruptible along with compactions (AFAICT no other actions that may
> markCompacting are not themselves compactions).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
