[
https://issues.apache.org/jira/browse/FLINK-26079?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17492094#comment-17492094
]
Roman Khachatryan commented on FLINK-26079:
-------------------------------------------
> shouldn't claim mode with changelog state backend wrap it first and then be
> added to the checkpoint store?
As per the discussion above, this doesn't seem a good option because JM should
NOT be aware of how specific backends build their state handles.
> 3. Wrap the materialized state with Changelog handles on JM, during recovery
> (not an option IMO because JM shouldn't be aware of that)
> Conceptually, when claiming: shouldn't we claim as a changelog checkpoint if
> we are using changelog state-backend after restoring?
What do you mean by "claim as a changelog checkpoint"?
> But I agree that's something we need to figure out after code freeze.
I personally meant that after the feature freeze we'll probably fix the root
cause.
This issue of enforcing the limitation should be merged ASAP IMO, so that it
can be tested as part of the release process. Although formally it can be
merged after the feature freeze too.
> [Changelog] Disallow recovery from non-changelog checkpoints
> ------------------------------------------------------------
>
> Key: FLINK-26079
> URL: https://issues.apache.org/jira/browse/FLINK-26079
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Configuration, Runtime / State Backends
> Reporter: Roman Khachatryan
> Assignee: Roman Khachatryan
> Priority: Blocker
> Labels: pull-request-available
> Fix For: 1.15.0
>
>
> Extracted from FLINK-25872.
> The issue is with the CLAIM mode:
> > Because discarding an initial checkpoint will invalidate its "private"
> > state which might be in use by future checkpoints.
> > Normally, changelog backend wraps it and registers with tjhe
> > SharedStateRegistry.
> > But when recovering from non-changelog checkpoint, it is first added to the
> > Checkpoint store, and wrapping in subsequent checkpoints doesn't help.
> NO_CLAIM mode is not supported.
> LEGACY could work.
> But it's difficult to differentiate between the modes on TM, where backend
> type is reliably known (see the discussion below).
> CANONICAL non-changelog savepoints must still be supported.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
