Apache9 commented on code in PR #6991:
URL: https://github.com/apache/hbase/pull/6991#discussion_r2095897676
##########
hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/MergeTableRegionsProcedure.java:
##########
@@ -450,6 +451,20 @@ private boolean prepareMergeRegion(final
MasterProcedureEnv env) throws IOExcept
"Skip merging regions " + regionNamesToLog + ", because we are
snapshotting " + tn);
}
+ /*
+ * Sometimes a ModifyTableProcedure has edited a table descriptor to
change the number of region
+ * replicas for a table, but it has not yet opened/closed the new
replicas. The
+ * ModifyTableProcedure assumes that nobody else will do the
opening/closing of the new
+ * replicas, but a concurrent MergeTableRegionProcedure would violate that
assumption. So, we
+ * need to wait for the ModifyTableProcedure to finish.
+ */
+ if (TableModificationProcedureUtils.isTableModificationInProgress(env,
getTableName())) {
+ setFailure(getClass().getSimpleName(),
+ new IOException("Skip merging regions " + regionNamesToLog
Review Comment:
> So then, should this merge request enter a retry loop/queue instead of
simply failing? I think that the system overall will be more reliable if gets
around to executing the requested procedure eventually.
In the current code base it is not easy to do so. We set holdLock = true for
split/merge procedure to avoid races, so we will hold the region exclusive lock
all the time even if the procedure is waiting, this will hurt availability
since TRSP can not get the region lock to assign it... Think of a region crash
happens...
And this may even introduce dead lock, for example, ModifyTableProcedure
needs to reopen all the regions, which replies on schedule TRSPs for all the
regions for a table, if split/merge procedure hold the region lock all the
time, at least one TRSP can not be executed and cause the ModifyTableProcedure
never end, but here the split/merge procedure needs to wait until the
ModifyTableProcedure finishes...
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
