[
https://issues.apache.org/jira/browse/DERBY-4274?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Knut Anders Hatlen updated DERBY-4274:
--------------------------------------
Attachment: d4274-1a.diff
The exclusive lock is obtained in
AlterTableConstantAction.executeConstantAction():
// now do the real work
// get an exclusive lock of the heap, to avoid deadlock on rows
of
// SYSCOLUMNS etc datadictionary tables and phantom table
// descriptor, in which case table shape could be changed by a
// concurrent thread doing add/drop column.
.
.
.
lockTableForDDL(tc, tableConglomerateId, true);
When this code is called, the index cardinality statistics have already been
updated, just not committed, so there's no need for the lock. It looks to me as
if the update statistics work is done, and it goes on to doing "normal" alter
table work, which is nothing in the case of update statistics. Looking at the
code for in-place compress, it returns from executeConstantAction() before it
gets to this point, and I believe that update statistics should do the same.
The attached patch changes AlterTableConstantAction.executeConstantAction() so
that it returns immediately when it has finished updating the statistics. It
also adds a test case to UpdateStatisticsTest where SYSCS_UPDATE_STATISTICS is
called on a table locked by another transaction in shared mode. It fails with a
timeout without the fix and passes with the fix.
Running the full regression test suite now.
> SYSCS_UPDATE_STATISTICS takes unnecessary table lock
> ----------------------------------------------------
>
> Key: DERBY-4274
> URL: https://issues.apache.org/jira/browse/DERBY-4274
> Project: Derby
> Issue Type: Bug
> Components: SQL
> Affects Versions: 10.5.1.1
> Reporter: Knut Anders Hatlen
> Assignee: Knut Anders Hatlen
> Attachments: d4274-1a.diff
>
>
> I'm using SYSCS_UTIL.SYSCS_UPDATE_STATISTICS in an application and see
> deadlocks similar to this one from time to time:
> Caused by: java.sql.SQLTransactionRollbackException: A lock could not be
> obtained due to a deadlock, cycle of locks and waiters is:
> Lock : ROW, SYSSTATISTICS, (2,20)
> Waiting XID : {20137, S} , APP, SELECT CS.REVISION, A.NAME, CS.TIME,
> CS.MESSAGE, F2.PATH FROM CHANGESETS CS, FILECHANGES FC, REPOSITORIES R, FILES
> F, AUTHORS A, FILECHANGES FC2, FILES F2 WHERE R.PATH = ? AND F.PATH LIKE ?
> ESCAPE '#' AND F.REPOSITORY = R.ID AND A.REPOSITORY = R.ID AND CS.ID =
> FC.CHANGESET AND R.ID = CS.REPOSITORY AND FC.FILE = F.ID AND A.ID = CS.AUTHOR
> AND CS.ID = FC2.CHANGESET AND FC2.FILE = F2.ID ORDER BY CS.ID DESC
> Granted XID : {20134, X}
> Lock : TABLE, CHANGESETS, Tablelock
> Waiting XID : {20134, X} , APP, alter table "APP"."CHANGESETS" all update
> statistics
> Granted XID : {20137, IS}
> . The selected victim is XID : 20137.
> Here, a select statement is being re-prepared because update statistics has
> invalidated it, but it has to wait for update statistics to finish in order
> to read the new statistics from SYSSTATISTICS. Then update statistics
> attempts to obtain an exclusive lock on the table whose indexes are being
> updated, but it has to wait because the select statement is holding an
> intention lock on the table. Both transactions wait for each other, so we
> have a deadlock.
> Since SYSCS_UPDATE_STATISTICS does not update the table, only its entries in
> SYSSTATISTICS, there is no need to take an exclusive table lock. If it didn't
> lock the table exclusively, this deadlock situation would go away.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
