[
https://issues.apache.org/jira/browse/HDFS-14941?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16963518#comment-16963518
]
Konstantin Shvachko commented on HDFS-14941:
--------------------------------------------
Great find [~vagarychen].
Yes this is _*existing problem in current code*_, but it is exacerbated by
fast journal tailing since small tail-edits.period increases the probability
of the race condition. In fact we occasionally see missing blocks after failing
over to SBN on our clusters with pre-CRS code.
Want to suggest another _*possible solution 4*_.
4. Delay incrementing the global {{generationStamp}} on standby until it
actually sees the block with that stamp. That is, when {{OP_SET_GENSTAMP_V2}}
comes to SBN it records the new value in a new variable
{{lastReportedGenStamp}}. When {{OP_ADD_BLOCK}} with the genStamp that equals
{{lastReportedGenStamp}} comes the global {{generationStamp}} is set to
{{lastReportedGenStamp}}. This should also solve the race condition.
We were looking at {{updateBlockForPipeline()}} and found out that it could be
_*another source of missing blocks on SBN*_, because it only increments the
global {{generationStamp}}, but does not update the generation stamp of that
block. The new gen stamp will be eventually updated by subsequent {{OP_ADD}},
or {{OP_CLOSE}}, or {{OP_UPDATE_BLOCKS}}. But the race condition with IBRs will
be still present. If an IBR comes after incrementing the global genStamp the
replica will not be in the future, but since the block genStamp has not be yet
updated by the subsequent {{OP_ADD}} or such, this replica will be invalidated
on SBN.
> Potential editlog race condition can cause corrupted file
> ---------------------------------------------------------
>
> Key: HDFS-14941
> URL: https://issues.apache.org/jira/browse/HDFS-14941
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: namenode
> Reporter: Chen Liang
> Assignee: Chen Liang
> Priority: Major
>
> Recently we encountered an issue that, after a failover, NameNode complains
> corrupted file/missing blocks. The blocks did recover after full block
> reports, so the blocks are not actually missing. After further investigation,
> we believe this is what happened:
> First of all, on SbN, it is possible that it receives block reports before
> corresponding edit tailing happened. In which case SbN postpones processing
> the DN block report, handled by the guarding logic below:
> {code:java}
> if (shouldPostponeBlocksFromFuture &&
> namesystem.isGenStampInFuture(iblk)) {
> queueReportedBlock(storageInfo, iblk, reportedState,
> QUEUE_REASON_FUTURE_GENSTAMP);
> continue;
> }
> {code}
> Basically if reported block has a future generation stamp, the DN report gets
> requeued.
> However, in {{FSNamesystem#storeAllocatedBlock}}, we have the following code:
> {code:java}
> // allocate new block, record block locations in INode.
> newBlock = createNewBlock();
> INodesInPath inodesInPath = INodesInPath.fromINode(pendingFile);
> saveAllocatedBlock(src, inodesInPath, newBlock, targets);
> persistNewBlock(src, pendingFile);
> offset = pendingFile.computeFileSize();
> {code}
> The line
> {{newBlock = createNewBlock();}}
> Would log an edit entry {{OP_SET_GENSTAMP_V2}} to bump generation stamp on
> Standby
> while the following line
> {{persistNewBlock(src, pendingFile);}}
> would log another edit entry {{OP_ADD_BLOCK}} to actually add the block on
> Standby.
> Then the race condition is that, imagine Standby has just processed
> {{OP_SET_GENSTAMP_V2}}, but not yet {{OP_ADD_BLOCK}} (if they just happen to
> be in different setment). Now a block report with new generation stamp comes
> in.
> Since the genstamp bump has already been processed, the reported block may
> not be considered as future block. So the guarding logic passes. But
> actually, the block hasn't been added to blockmap, because the second edit is
> yet to be tailed. So, the block then gets added to invalidate block list and
> we saw messages like:
> {code:java}
> BLOCK* addBlock: block XXX on node XXX size XXX does not belong to any file
> {code}
> Even worse, since this IBR is effectively lost, the NameNode has no
> information about this block, until the next full block report. So after a
> failover, the NN marks it as corrupt.
> This issue won't happen though, if both of the edit entries get tailed all
> together, so no IBR processing can happen in between. But in our case, we set
> edit tailing interval to super low (to allow Standby read), so when under
> high workload, there is a much much higher chance that the two entries are
> tailed separately, causing the issue.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
