[
https://issues.apache.org/jira/browse/PHOENIX-6387?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17340445#comment-17340445
]
ASF GitHub Bot commented on PHOENIX-6387:
-----------------------------------------
tkhurana commented on a change in pull request #1215:
URL: https://github.com/apache/phoenix/pull/1215#discussion_r627780432
##########
File path:
phoenix-core/src/main/java/org/apache/phoenix/hbase/index/IndexRegionObserver.java
##########
@@ -1138,10 +1329,10 @@ private void
doPre(ObserverContext<RegionCoprocessorEnvironment> c, BatchMutateC
metricSource.updatePreIndexUpdateFailureTime(dataTableName,
EnvironmentEdgeManager.currentTimeMillis() - start);
metricSource.incrementPreIndexUpdateFailures(dataTableName);
- // Remove all locks as they are already unlocked. There is no need
to unlock them again later when
- // postBatchMutateIndispensably() is called
- removePendingRows(context);
- context.rowLocks.clear();
+ // Re-acquire all locks since we released them before making index
updates
+ // Removal of reference counts and locks for the rows of this batch
will be
Review comment:
All cleanup now happens in the `postBatchMutateIndispensably` function
which is always called irrespective of whether the preBatchMutate() and doPre()
which it internally calls succeed or throw an exception. That function works
under the assumption that rows have been locked so that it can safely
manipulate the batch mutate context object. Because we released the row locks
before calling doPre() and acquire them again in the success path, we need to
do the same in the exception path.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> Conditional updates on tables with indexes
> ------------------------------------------
>
> Key: PHOENIX-6387
> URL: https://issues.apache.org/jira/browse/PHOENIX-6387
> Project: Phoenix
> Issue Type: Improvement
> Affects Versions: 5.0.0, 4.15.0
> Reporter: Kadir OZDEMIR
> Assignee: Tanuj Khurana
> Priority: Major
>
> For a row update done by using the UPSERT VALUES statement, the exact values
> of the columns to be updated are specified within the UPSERT statement.
> Regardless of whether a given row exists or not, after the update, we know
> what the content will be for these columns. However, this is not the case
> when the ON DUPLICATE KEY clause is added the UPSERT VALUES statement. This
> clause makes the update conditional and the end result is determined based on
> the conditions stated within the clause and the current state of the row at
> the time the update is done. Also, this clause makes the UPSERT VALUES
> statement atomic.
> Conditional updates are supported for the tables without indexes currently.
> The current design leverages an HBase atomic operation and cannot be expanded
> to support tables with indexes since the design requires holding (HBase
> level) row locks while doing index table updates over RPCs. This results in
> cluster wide deadlocks. This jira is to redesign conditional updates using
> Phoenix level row locks instead of using HBase level row locks to also
> support tables with indexes by leveraging the design of PHOENIX-6160 which
> simplifies the concurrent mutation handling on tables with indexes.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
