[
https://issues.apache.org/jira/browse/HBASE-2283?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12840957#action_12840957
]
Kannan Muthukkaruppan commented on HBASE-2283:
----------------------------------------------
Followup discussions on the hbase-dev:
JD wrote: <<< Indeed. The syncWal was taken back up in HRS as a way to optimize
batch Puts but the fact it's called after all the MemStore operations is indeed
a problem. I think we need to fix both (#1) and (#2) by ensuring we do only a
single append for whatever we have to put and then syncWAL once before
processing the MemStore. But, the other problem here is that the row locks have
to be taken out on all rows before everything else in the case of a Put[] else
we aren't atomic. And then I think some checks are ran under HRegion that we
would need to run before everything else.>>>
Ryan wrote: <<< Do we really need a single actual DFS atomic write operation?
If we had some kind of end-of-row marker, would that help instead?>>>
Yes, a marker or length-prefixed approach would suffice to recognize and ignore
incomplete transactions during recovery.
Ryan wrote: <<< But as you said, what happens if hlog append fails? The
obvious thing would be to remove the additions from the memstore. But how to
accomplish this easily?>>>
Wouldn't moving all the memstore updates to happen after the sync suffice?
> row level atomicity
> --------------------
>
> Key: HBASE-2283
> URL: https://issues.apache.org/jira/browse/HBASE-2283
> Project: Hadoop HBase
> Issue Type: Bug
> Reporter: Kannan Muthukkaruppan
> Priority: Blocker
>
> The flow during a HRegionServer.put() seems to be the following. [For now,
> let's just consider single row Put containing edits to multiple column
> families/columns.]
> HRegionServer.put() does a:
> HRegion.put();
> syncWal() (the HDFS sync call). /* this is assuming we have HDFS-200
> */
> HRegion.put() does a:
> for each column family
> {
> HLog.append(all edits to the colum family);
> write all edits to Memstore;
> }
> HLog.append() does a :
> foreach edit in a single column family {
> doWrite()
> }
> doWrite() does a:
> this.writer.append().
> There seems to be two related issues here that could result in
> inconsistencies.
> Issue #1: A put() does a bunch of HLog.append() calls. These in turn do a
> bunch of "write" calls on the underlying DFS stream. If we crash after
> having written out some append's to DFS, recovery will run and apply a
> partial transaction to memstore.
> Issue #2: The updates to memstore should happen after the sync rather than
> before. Otherwise, there is the danger that the write to DFS (sync) fails for
> some reason & we return an error to the client, but we have already taken
> edits to the memstore. So subsequent reads will serve uncommitted data.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
