[
https://issues.apache.org/jira/browse/HBASE-8877?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dave Latham updated HBASE-8877:
-------------------------------
Attachment: HBASE-8877-refCounts.patch
Alright, here's a patch with reentrant row locks and reference counts.
This patch updates the row lock interface on HRegion to the following methods:
- RowLock HRegion.getRowLock(final byte[] row, boolean waitForLock): acquire
lock, optionally wait for lock
- RowLock HRegion.getRowLock(final byte[] row): acquire lock, always waits for
lock
- RowLock.release(): releases the row lock
The same thread can acquire multiple locks on the same row, but must release
them all. doMiniBatchMutate now keeps a List of row locks (one for each row)
and releases them all.
While reviewing the clients I noticed that HRegion.increment and the duplicated
code in HRegion.append both had possible paths to fail to release their row
locks so I added a try/finally block to ensure they always release their locks.
All the tests pass locally. I will run the microbenchmark too and post the
results.
> Reentrant row locks
> -------------------
>
> Key: HBASE-8877
> URL: https://issues.apache.org/jira/browse/HBASE-8877
> Project: HBase
> Issue Type: Bug
> Components: Coprocessors, regionserver
> Reporter: Dave Latham
> Assignee: Dave Latham
> Fix For: 0.95.2
>
> Attachments: hbase-8877-0.94-microbenchmark.txt,
> HBASE-8877-0.94.patch, HBASE-8877-0.94-v2.patch, HBASE-8877.patch,
> HBASE-8877-refCounts.patch, HBASE-8877-v2.patch, HBASE-8877-v3.patch,
> hbase-8877-v4-microbenchmark.txt, HBASE-8877-v4.patch, HBASE-8877-v5.patch,
> HBASE-8877-v6.patch, HBASE-8877-v7.patch
>
>
> HBASE-8806 revealed performance problems with batch mutations failing to
> reacquire the same row locks. It looks like HBASE-8806 will use a less
> intrusive change for 0.94 to have batch mutations track their own row locks
> and not attempt to reacquire them. Another approach will be to support
> reentrant row locks directly. This allows simplifying a great deal of
> calling code to no longer track and pass around lock ids.
> One affect this change will have is changing the RegionObserver coprocessor's
> methods preBatchMutate and postBatchMutate from taking a
> {{MiniBatchOperationInProgress<Pair<Mutation, Integer>> miniBatchOp}} to
> taking a {{MiniBatchOperationInProgress<Mutation> miniBatchOp}}. I don't
> believe CPs should be relying on these lock ids, but that's a potential
> incompatibility.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
