[ 
https://issues.apache.org/jira/browse/FLINK-41466?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Anton Kalashnikov resolved FLINK-41466.
---------------------------------------
    Resolution: Fixed

> Restored savepoint's CheckpointProperties always report CANONICAL format, 
> regardless of the actual savepoint format
> -------------------------------------------------------------------------------------------------------------------
>
>                 Key: FLINK-41466
>                 URL: https://issues.apache.org/jira/browse/FLINK-41466
>             Project: Flink
>          Issue Type: Bug
>    Affects Versions: 2.3.0, 2.4.0
>            Reporter: Anton Kalashnikov
>            Assignee: Anton Kalashnikov
>            Priority: Major
>              Labels: pull-request-available
>             Fix For: 2.5.0
>
>
> h2. Problem
> When a job is restored from a savepoint, 
> \{{CheckpointCoordinator#restoreSavepoint}} builds the 
> \{{CheckpointProperties}} attached to the restored checkpoint. These 
> properties are then recorded through 
> \{{CheckpointStatsTracker#reportRestoredCheckpoint}} and exposed via 
> \{{RestoredCheckpointStats#getProperties()}}.
> In the two claim modes that treat the restored checkpoint as a savepoint, the 
> savepoint format in these properties is hardcoded to CANONICAL:
> * \{{NO_CLAIM}} (the default): 
> \{{CheckpointProperties.forUnclaimedSnapshot()}} uses 
> \{{SavepointType.savepoint(SavepointFormatType.CANONICAL)}}.
> * \{{LEGACY}}: \{{CheckpointProperties.forSavepoint(false, 
> SavepointFormatType.CANONICAL)}}, with the code comment "we do not care about 
> the format when restoring".
> As a result, a job restored from a NATIVE savepoint ends up with a restored 
> checkpoint whose properties say CANONICAL. The recorded type does not 
> describe what was actually restored. According to the code comment the 
> restore logic does not depend on the format, so only the recorded metadata is 
> wrong.
> h2. Why it matters
> The restored checkpoint's properties are the only place where its type is 
> recorded, and for NATIVE savepoints the recorded format is wrong. Anything 
> that reads them (REST handlers, metrics, custom code built on 
> \{{RestoredCheckpointStats}}) gets a misleading answer. In particular, the 
> format of the restored checkpoint cannot be surfaced (for example next to 
> \{{is_savepoint}} in the \{{latest.restored}} part of the checkpoint 
> statistics REST response) without exposing a wrong value.
> h2. How to reproduce (from code reading, not yet verified with a test)
> # Run a job with the RocksDB state backend and take a savepoint with 
> \{{--type native}}.
> # Restore from it with the default claim mode (\{{NO_CLAIM}}).
> # Inspect \{{RestoredCheckpointStats#getProperties().getCheckpointType()}}: 
> it is a \{{SavepointType}} whose format is CANONICAL.



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to