[
https://issues.apache.org/jira/browse/HDFS-1186?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12882365#action_12882365
]
Todd Lipcon commented on HDFS-1186:
-----------------------------------
Hi Sam, thanks for taking a look. I think you're right that in some really
weird timing scenarios we might have a problem:
{noformat}
writer writes offset 1 and syncs, gs=1
NN recovery starts:
- interrupts writer, gets metadata (len 1)
- recovering DN hangs for a little bit
writer recovery starts, picks a different primary DN:
- interrupts writer (noop)
- gets metadata (len 1)
- gets new GS=2
- syncs blocks to GS=2 len=1
- restarts pipeline
- writes and syncs some more data to block with GS=2
NN-directed recovery proceeds:
- gets new GS=3 (this has to be at least 10 seconds after above due to
lastRecoveryTime check)
- calls updateBlock on all DNs, which truncates files
{noformat}
I think the issue here is that the genstamp can be incremented in between
startBlockRecovery() and updateBlock(), and thus updateBlock is allowing an
update based on stale recovery info. If we simply added a check in
tryUpdateBlock() that oldblock.getGenerationStamp() == oldgs, I think we'd be
safe. What do you think?
> 0.20: DNs should interrupt writers at start of recovery
> -------------------------------------------------------
>
> Key: HDFS-1186
> URL: https://issues.apache.org/jira/browse/HDFS-1186
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: data-node
> Affects Versions: 0.20-append
> Reporter: Todd Lipcon
> Assignee: Todd Lipcon
> Priority: Blocker
> Attachments: hdfs-1186.txt
>
>
> When block recovery starts (eg due to NN recovering lease) it needs to
> interrupt any writers currently writing to those blocks. Otherwise, an old
> writer (who hasn't realized he lost his lease) can continue to write+sync to
> the blocks, and thus recovery ends up truncating data that has been sync()ed.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
