[
https://issues.apache.org/jira/browse/HBASE-8458?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15197849#comment-15197849
]
Esther Kundin commented on HBASE-8458:
--------------------------------------
I actually have the same problem and attempted to fix with a coprocessor. An
observer coprocessor is unable to do the appropriate locking if writes are
happening from different clients. You cannot run a checkAndPut from within the
coprocessor as by the time you are in the preBatchMutate, the row is locked for
writing and a checkAndPut will be unable to run on the same row. If you
implement it as an endpoint observer, then it should work, though I haven't
tried it, but it will do a read/write lock on the row, which is a stricter
locking than checkAndPut would do, so it would hurt performance more than a
batch checkAndPut which will not block reads for as long. So it would seem
that a batch checkAndPut would be the best solution.
> Support for batch version of checkAndPut() and checkAndDelete()
> ---------------------------------------------------------------
>
> Key: HBASE-8458
> URL: https://issues.apache.org/jira/browse/HBASE-8458
> Project: HBase
> Issue Type: Improvement
> Components: Client, regionserver
> Affects Versions: 0.95.0
> Reporter: Hari Mankude
>
> The use case is that the user has multiple threads loading hundreds of keys
> into a hbase table. Occasionally there are collisions in the keys being
> uploaded by different threads. So for correctness, it is required to do
> checkAndPut() instead of a put(). However, doing a checkAndPut() rpc for
> every key update is non optimal. It would be good to have a batch version of
> checkAndPut() similar to batch put(). The client can partition the keys on
> region boundaries.
> The jira is NOT looking for any type of cross-row locking or multi-row
> atomicity with checkAndPut()
> Batch version of checkAndDelete() is a similar requirement.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
