Ignore obsolete dangling indices #37824
Conversation
For a non-data, non-master node we now warn about dangling indices and will otherwise ignore them. This avoids import of old indices with a following inevitable red cluster status. Issue elastic#27073
henningandersen
added
>bug
:Distributed Coordination/Allocation
|
Pinging @elastic/es-distributed |
There was a problem hiding this comment.
LGTM. Should we also log a warning at start-up that the node has metadata / data which should possibly be cleaned up? Another small follow-up would be to completely disable DanglingIndicesState on non-master non-data nodes on 7.x, by only conditionally calling clusterService.addListener(this); in the constructor.
Now warn about both left-behind data and metadata for non-data or non-data and non-master nodes. Disable dangling indices check completely for coordinating only nodes (non-data and non-master).
Fix checkstyle issues.
Improved messaging.
|
Thanks for your comments, @ywelsch, I have added those in this same PR. Seems like a better solution than the original. Since it is more or less a complete rewrite, I am requesting another review. GatewayIndexStateIT did not change since your last review |
Fix checkstyle issue.
There was a problem hiding this comment.
Thanks @henningandersen. I like aligning this as close as possible with the 7.0 code. I've left two smaller comments, but looks good o.w.
| try { | ||
| ensureNoIndexMetaData(nodePaths); | ||
| } catch (IllegalStateException e) { | ||
| logger.warn(e.getMessage() + ", this should be cleaned up (will refuse to start in 7.0). Beware of data-loss."); |
There was a problem hiding this comment.
I think we should use the deprecation logger here (see DeprecationLogger class). That one will log to a different log file. This will help users in finding all things to address before upgrading
There was a problem hiding this comment.
I'm not sure about the "Beware of data-loss." part of the message. That's not really actionable. Perhaps it could say something along the lines of "create a backup copy before removing."
Improved messaging and log to the deprecation logger.
For a non-data, non-master node we now warn about dangling indices and
will otherwise ignore them. This avoids import of old indices with a
following inevitable red cluster status.
Issue #27073