[
https://issues.apache.org/jira/browse/IGNITE-17215?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17570796#comment-17570796
]
Maksim Timonin commented on IGNITE-17215:
-----------------------------------------
[~xtern] hi! I found the reason, this is actually a bug, introduced in
https://issues.apache.org/jira/browse/IGNITE-17272. I prepared a fix in
https://issues.apache.org/jira/browse/IGNITE-17408.
> Write ClusterSnapshotRecord to WAL
> ----------------------------------
>
> Key: IGNITE-17215
> URL: https://issues.apache.org/jira/browse/IGNITE-17215
> Project: Ignite
> Issue Type: New Feature
> Reporter: Maksim Timonin
> Assignee: Maksim Timonin
> Priority: Major
> Labels: IEP-89
> Fix For: 2.14
>
> Time Spent: 50m
> Remaining Estimate: 0h
>
> For PITR [1] process of recovering based on ClusterSnapshot + archived WALs.
> It's required to have a point in WAL which splits whole WAL on 2 areas:
> # Before this point all data changes are contained within ClusterSnapshot,
> and no need to recover them from WAL archived files.
> # After this point all data need to be recovered from WAL archived files.
> It's proposed to write ClusterSnapshotRecord while the checkpoint is running
> (cp#writeLock has acquired). ClusterSnapshot process guarantees:
> # there is no active transactions (or any data changes) in moment of running
> checkpoint.
> # ClusterSnapshot contains all data pages that will be persisted within this
> checkpoint process.
> Then every logical record after begin CheckointRecord doesn't belong to
> ClusterSnapshot. Then it's safe to write ClusterSnapshotRecord within the
> checkpoint process.
> [1]
> [https://cwiki.apache.org/confluence/pages/editpage.action?pageId=211884314]
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
