db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Mike Matrigali (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (DERBY-6317) Optmizer can choose the wrong path when BTreeCostController.java returns an estimate cost and row count of 0.0
Date Tue, 27 Aug 2013 20:28:51 GMT

    [ https://issues.apache.org/jira/browse/DERBY-6317?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13751680#comment-13751680
] 

Mike Matrigali commented on DERBY-6317:
---------------------------------------

i do see a lot of cost *= cost and rowcount = rowcount1 * rowcount2 kinds of code in the optimizer,
so even more
reason not to return 0 out of the store costs.  

I guess the optmizer model might not be wrong if it is actually the case that if there were
0 rows in the outer table that
it would actually not do the full scan.  But since these are all estimates seems like there
always should be some cost that
is linearly related to number of rows scanned if there are actually not 0 rows.  better to
be a little pessimistic and return
1 row rather 0.  

There may be ways to make that btree code be smarter when it is returning a very small number
of rows, especially if the
start search and stop search are actually the same page.
                
> Optmizer can choose the wrong path when BTreeCostController.java returns an estimate
cost and row count of 0.0
> --------------------------------------------------------------------------------------------------------------
>
>                 Key: DERBY-6317
>                 URL: https://issues.apache.org/jira/browse/DERBY-6317
>             Project: Derby
>          Issue Type: Bug
>          Components: SQL
>    Affects Versions: 10.8.2.2
>         Environment: Derby 10.8.2.2 on Oracle Solaris 10 
>            Reporter: Brett Bergquist
>
> The optimizer can chose the wrong path when BTreeCostController.java returns an estimate
cost and row count of 0.0.  
> Assume that you have two tables that are being joined like:
> SELECT * FROM T1, T0
> WHERE T1.ID = T0.F_ID and
> T0.ID = 3;
> Also assume that T0 has two columns, ID and F_ID and F_ID is a foreign key on T1.ID.
  Assume that T1.ID is the primary key of T1 and (T0.F_ID, T0.ID) is the primary key on T0.
 Assume that there is a non-unique index on T0.ID.
> The correct query plan for this should be to query T0 using the non-unique index on T0.ID
and then use the foreign key value in those rows to do query T1 using the primary key on T1.
> With some values of T0.ID in the above query this query plan is chosen and works.  With
other values of T0.ID , the query plan does an query on T0 using the non-unique index on T0.ID
and then does a table scan on T1.
> For example, in my case the query:
> SELECT * FROM T1, T0
> WHERE T1.ID = T0.F_ID and
> T0.ID = 22112129;
> has this query plan.   
> The problem appears to be in BTreeCostController.java.  When this returns the same value
for the "left_of_start" and the "left_of_stop" (which is being used to estimate the number
of rows and cost), then the estimate cost and row count becomes 0.0.   When this is used in
the join order of T0, T1, then the cost of the table scan for T1 becomes 0.0 as well.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Mime
View raw message