[
https://issues.apache.org/jira/browse/SPARK-54121?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anish Shrigondekar resolved SPARK-54121.
----------------------------------------
Fix Version/s: 4.2.0
Resolution: Fixed
Issue resolved by pull request 52821
[https://github.com/apache/spark/pull/52821]
> Automatic Snapshot Repair for State store
> -----------------------------------------
>
> Key: SPARK-54121
> URL: https://issues.apache.org/jira/browse/SPARK-54121
> Project: Spark
> Issue Type: Task
> Components: Structured Streaming
> Affects Versions: 4.1.0
> Reporter: B. Micheal Okutubo
> Assignee: B. Micheal Okutubo
> Priority: Major
> Labels: pull-request-available
> Fix For: 4.2.0
>
>
> Today, the engine currently treats both the changelog and snapshot files with
> the same importance, so when a state store needs to be loaded it reads the
> latest snapshot and applies the subsequent changes on it. If the snapshot is
> bad or corrupted, then the query will fail and be completely down and
> blocked, needing manual intervention. This leads to user clearing their query
> checkpoint and having to do full recomputation.
> This shouldn’t be the case. The changelog should be treated as the “source of
> truth” and the snapshot is just a disposable materialization of the log.
> Introducing Automatic snapshot repair, which will automatically repair the
> checkpoint by skipping bad snapshots and rebuilding the current state from
> the last good snapshot (works even if there’s none) and applying the
> changelogs on it. This eliminates the need for manual intervention and
> unblocks the pipeline to keep it running.
> Also emit metrics about number of state stores that were auto repaired in a
> given batch, so that you can build alert and dashboard for it.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
