DieterDP-ng commented on code in PR #7582: URL: https://github.com/apache/hbase/pull/7582#discussion_r2709210795
########## hbase-backup/src/main/java/org/apache/hadoop/hbase/backup/impl/IncrementalBackupManager.java: ########## Review Comment: So the scenario would be that a RS host X would (just before receiving the log roll command) start a shutdown, causing it so there's no updated entry for that host in `readRegionServerLastLogRollResult`. Then, there's at least one RS Y that completes it's roll so quick, while having some time skew, that it would appear that X went offline after the the log roll of Y, causing it to not be included. It's a bit of a far-fetched scenario, but given the complexity of having 2 possible logroll procedures, I can't easily say whether it's possible or not. But I admit that looking at the WAL files instead avoids this complexity of reasoning. So, I guess I'm convinced that the suggested WAL-file approach is fine. I would appreciate if the code is condensed & structured a bit extra though. Just out of curiosity: did you actually have evidence of clock skew causing issues? I thought clock-sync methods were sufficient to avoid such issues. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
