[
https://issues.apache.org/jira/browse/HADOOP-2012?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12537517
]
eric baldeschwieler commented on HADOOP-2012:
---------------------------------------------
A couple of comments:
1) the idea of not keeping scan times strikes me as "really bad". Also
randomizing the scan order weirds me out. I do think Raghu's point about
modeling the meta files may be valid. Why not simply always scan blocks in
numeric order starting from zero and log actions. On restart we can tail this
log to find the last block validated and start from there. Inserted blocks are
recent by definition, so it is ok if we don't get around to them til the next
scan.
2) It seems to me it might be better to try to repair the block if possible,
rather then just delete it. This avoids bad corner cases. It adds complexity
though. Thoughts? A simple variant is just to copy a new version locally.
3) Throttling might simply entail spacing and scheduling when you scan the next
block to complete within roughly two weeks. This would imply that we want to
persist when the current scan started. If we do that, the penalty of scanning
quickly might be fairly ignorable, when you consider the other variations in
work load a DN is exposed to anyway. You'd want some rule like always wait 10x
the time it took you to validate a block between blocks to avoid wierd corner
cases where the node gets a huge number of blocks added near the end of the
time period.
> Periodic verification at the Datanode
> -------------------------------------
>
> Key: HADOOP-2012
> URL: https://issues.apache.org/jira/browse/HADOOP-2012
> Project: Hadoop
> Issue Type: New Feature
> Components: dfs
> Reporter: Raghu Angadi
> Assignee: Raghu Angadi
> Fix For: 0.16.0
>
> Attachments: HADOOP-2012.patch, HADOOP-2012.patch, HADOOP-2012.patch,
> HADOOP-2012.patch
>
>
> Currently on-disk data corruption on data blocks is detected only when it is
> read by the client or by another datanode. These errors are detected much
> earlier if datanode can periodically verify the data checksums for the local
> blocks.
> Some of the issues to consider :
> - How should we check the blocks ( no more often than once every couple of
> weeks ?)
> - How do we keep track of when a block was last verfied ( there is a .meta
> file associcated with each lock ).
> - What action to take once a corruption is detected
> - Scanning should be done as a very low priority with rest of the datanode
> disk traffic in mind.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
