[
https://issues.apache.org/jira/browse/HBASE-29442?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18004876#comment-18004876
]
Vinayak Hegde commented on HBASE-29442:
---------------------------------------
After evaluating the concerns around ordering and consistency in the context of
our continuous backup implementation, we've concluded that serial replication
is not required.
Originally, we considered serial replication to avoid out-of-order WAL edits
(e.g., Delete arriving before Put), which can cause inconsistencies in live
replication scenarios. However, in our case:
* We're replicating to a backup store, not a live cluster, so there's no flush
or compaction occurring between the Delete and Put.
* WAL order is preserved, and tools like WALPlayer can handle the ordering.
* Each RS is uniquely identified by host+port+startcode, so restarted RSes are
treated as new sources, and we don’t risk duplicate or conflicting streams.
* The replication framework already isolates WAL queues by RS and handles
recovery cleanly via separate threads.
* PITR implements its own checkpoint logic, and while stale RS entries could
skew calculations, we've identified how to address this separately without
requiring serial replication.
Hence, normal replication suffices for our backup use case, and the risks that
serial replication addresses don’t apply here.
We’ll proceed to document this reasoning and can consider closing the ticket.
> Evaluate Need for Serial Replication in Continuous Backup
> ---------------------------------------------------------
>
> Key: HBASE-29442
> URL: https://issues.apache.org/jira/browse/HBASE-29442
> Project: HBase
> Issue Type: Task
> Components: backup&restore
> Reporter: Vinayak Hegde
> Assignee: Vinayak Hegde
> Priority: Major
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
