db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "A B (JIRA)" <derby-...@db.apache.org>
Subject [jira] Updated: (DERBY-558) Optimizer hangs with query that uses more than 6 tables and does subquery flattening.
Date Mon, 12 Sep 2005 16:00:46 GMT
     [ http://issues.apache.org/jira/browse/DERBY-558?page=all ]

A B updated DERBY-558:

    Attachment: d558.patch

Attaching a patch for this problem.  The patch does the following:

1) Fixes the logic in OptimizerImpl.java that was causing the hang (an indirect infinite loop).
2) Adds some comments describing the "JUMPING" logic that is in OptimizerImpl so that developers
looking at the code can (hopefully) figure out what's going on more quickly in the future.
3) Adds a test case to the lang/subqueryFlattening.sql test for verification of the fix. 
The test case is based on the repro attached to this issue.  NOTE: I had to set the "derby.optimizer.noTimeout"
property to true for this entire test--I think this is okay since everything still passes
(on my machine), but if anyone feels otherwise, please let me know...

I ran derbyall on Windows 2000 w/ Sun jdk 1.4.2 and saw no failures.  If someone could review
this, I'd be grateful..

> Optimizer hangs with query that uses more than 6 tables and does subquery flattening.
> -------------------------------------------------------------------------------------
>          Key: DERBY-558
>          URL: http://issues.apache.org/jira/browse/DERBY-558
>      Project: Derby
>         Type: Bug
>     Versions:,,,,,
>  Environment: Running query in "ij"  with derby.optimizer.noTimeout=true
>     Reporter: A B
>     Assignee: A B
>      Fix For:,
>  Attachments: d558.patch, repro.sql
> I was running a query that has a large number (hundreds) of tables in it and I set the
derby property "derby.optimizer.noTimeout" to true to see what plan Derby would choose as
the _best_ plan for the query.  When doing so, I ran into a situation where the optimizer
hung forever--which is wrong.  I expect that setting "noTimeout" to true might cause the query
to run more slowly (since it has to evaluate ALL possible join orders for all of the tables
in question), but it should _not_ cause the optimizer to hang forever.
> I noticed that "subquery flattening" is peformed on the query, which introduces dependencies
between the various tables and thus restricts the possible join orders that the optimizer
can choose (see http://db.apache.org/derby/docs/10.1/tuning/ctuntransform25868.html).  I was
eventually able to track the problem down to code in OptimizerImpl where, for queries with
more than 6 tables, a certain "jumping" algorithm is used to try to allow the optimizer to
find a better plan more quickly.
> Long story short, there is logic in the "jumping" mechanism that tries to put the tables
into a legal join order, but in certain (rare) cases where multiple join order dependencies
have to be enforced, the jump logic can end up looping indefinitely, causing the "hang" in
the optimizer.

This message is automatically generated by JIRA.
If you think it was sent incorrectly contact one of the administrators:
For more information on JIRA, see:

View raw message