[jira] Resolved: (DERBY-642) SELECT MAX doesn't use indices optimally

["Knut Anders Hatlen (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/DERBY-642?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Knut Anders Hatlen resolved DERBY-642.
--------------------------------------

       Resolution: Fixed
    Fix Version/s: 10.8.0.0

I'm marking this issue as fixed in 10.8 so that it turns up in the release notes.

There's still the unaddressed issue with how to prevent phantom reads in serializable transactions with row lock granularity. Since BTreeMaxScan always uses table locks when the isolation level is serializable, it won't cause any problems currently, so I believe it is sufficient to address this issue when we add full support for backward scans. I'll add a comment about it in DERBY-884.

> SELECT MAX doesn't use indices optimally
> ----------------------------------------
>
>                 Key: DERBY-642
>                 URL: https://issues.apache.org/jira/browse/DERBY-642
>             Project: Derby
>          Issue Type: Improvement
>          Components: Store
>    Affects Versions: 10.2.1.6
>            Reporter: Knut Anders Hatlen
>            Assignee: Knut Anders Hatlen
>            Priority: Minor
>             Fix For: 10.8.0.0
>
>         Attachments: derby-642-1a.diff, derby-642-1b-withTests.diff, derby-642-2a-test-serializable.diff, derby-642-3a-waitBeforeRetry.diff
>
>
> I tried running SELECT MAX on an indexed column in a big (8 GB)
> table. It took 12 minutes, which is about 12 minutes more than I
> expected.
> After a bit of investigation, I found out that a full index scan was
> performed because all the rows referenced from the rightmost B-tree
> node were actually deleted.
> Possible improvements:
>   1. Implement backwards scan in the B-trees (this is also suggested
>      in the comments in BTreeMaxScan).
>   2. Go up to the parent node and down to the next leaf node on the
>      left side, and continue until a valid max value is found. In
>      Derby, traversing up in a B-tree is not allowed, but would it be
>      ok to go up if the latches were kept on the higher-level nodes in
>      the tree? Of course, this would have negative impact on
>      concurrency.
>   3. Right-to-left traversal on the leaf level is possible (using
>      ControlRow.getLeftSibling()), but will throw an exception if the
>      page cannot be latched without waiting. It is therefore possible
>      to try a right-to-left search for the max value, and only fall
>      back to the left-to-right approach if a conflict arises.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira