db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Mamta A. Satoor (JIRA)" <j...@apache.org>
Subject [jira] Updated: (DERBY-3926) Incorrect ORDER BY caused by index
Date Thu, 04 Jun 2009 06:28:07 GMT

     [ https://issues.apache.org/jira/browse/DERBY-3926?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel

Mamta A. Satoor updated DERBY-3926:

    Attachment: DERBY3926_patch6_060309_stat.txt

I am attaching DERBY3926_patch5_060309_diff.txt Hopefully, this is the final patch for this
jira entry. This patch takes care of the original problem query and one query from wisconsin
which was getting an unnecessary sort node on it with the previous patch (DERBY3926_patch5_052709_stat.txt).

Following are the files that were touched by the patch.
M      java\engine\org\apache\derby\impl\sql\compile\RowOrderingImpl.java
M      java\engine\org\apache\derby\impl\sql\compile\OrderByList.java
M      java\engine\org\apache\derby\impl\sql\compile\FromBaseTable.java
M      java\engine\org\apache\derby\impl\sql\compile\OptimizerImpl.java
M      java\engine\org\apache\derby\impl\sql\compile\PredicateList.java
M      java\engine\org\apache\derby\iapi\sql\compile\RowOrdering.java
M      java\engine\org\apache\derby\iapi\sql\compile\RequiredRowOrdering.java
M      java\engine\org\apache\derby\iapi\sql\compile\OptimizablePredicateList.java
M      java\testing\org\apache\derbyTesting\functionTests\tests\lang\wisc_setup.sql
M      java\testing\org\apache\derbyTesting\functionTests\tests\lang\_Suite.java
M      java\testing\org\apache\derbyTesting\functionTests\tests\lang\OrderByAndSortAvoidance.java
M      java\testing\org\apache\derbyTesting\functionTests\master\wisconsin.out

Following is the patch description.
The problem with the trunk codeline is that when optimizer goes through optimizables in a
join order, it only looks at those optimizables individually to decide whether sorting can
be avoided on them or not. That approach leaves out few queries which require sorting but
do not get sorted. The decision for avoiding sorting should also include relationship between
the optimizables in a given join order. Following query demonstrates the trunk problem
SELECT table1.id, table2.value, table3.value FROM --DERBY-PROPERTIES joinOrder=FIXED 
table3 -- DERBY-PROPERTIES index=nonUniqueOnValue_Table3 
, table2 -- DERBY-PROPERTIES index=nonUniqueOnValue_Table2 
, table1 
WHERE table1.id=table2.id AND table2.name='PageSequenceId' 
AND table1.id=table3.id 
AND table3.name='PostComponentId' 
AND table3.value='21857' ORDER BY table2.value; 

In the query above, when optimizer is considering [table3, table2, -1] join order, it determines
that sorting can be avoided on this join order because the order by column table2.value is
already covered by the index nonUniqueOnValue_Table2. It does not see that the outermost optimizable
table3 will qualify more than one row and hence it will be a multi-row resulset and for each
one of those rows, we will be doing a scan into table2. In other words, there will be multiple
scans into table2(and the rows returned by each one of those scans will be ordered on table2.value)
but the collective rows from those multiple scans are not necessarily going to be ordered
on table2.value. This patch is attempting to fix that problem.

Currently, in trunk, a column is marked always ordered during a query processing when the
optimizer finds that there is constant comparison predicate on the order by column. If the
column does not have a constant predicate (as in our example above), we next see if we are
using an index which will provide the required ordering on column (which is true in our case.
The required ordering on table2.value is provided by the index nonUniqueOnValue_Table2). But
as we can see in the query above, this index coverage is not enough to say that sorting is
not needed. We need to add 2 more conditions before we can decide to avoid the sorting. One
of those cases is 1)if the order by column does not belong to the outermost optimizable, then
check if the order by column's optimizable is a one-row resultset. If yes, then it will be
safe for the optimizer to avoid the sorting. The second case to consider is 2)if the order
by column does not belong to the outermost optimizable, then check if the order by column's
optimizable is multi-row resultset BUT all the outer optimizables are one-row resulsets. If
either of these 2 additional conditions are satisfied then optimizer can choose to avoid the
sorting. Otherwise
sorting should be added to the query plan. The example query above does not satisfy the 2
additional checks and hence sorting should be done as part of the query plan.

The changes for the 1)check above has gone into OrderbyByList.sortRequired(RowOrdering, JBitSet,
OptimizableList). The implementation of this change just required us to check the outer optimizables
to be one row since the order by column's optimizable is not one row. If outer optimizables
are all one-row, then we say that sorting can be avoided. Otherwise sorting is required.

The changes for the 2)check above has gone into FromBaseTable.nextAccessPath(Optimizer optimizer,
OptimizablePredicateList predList, RowOrdering rowOrdering) The implementation of this change
requires us to see if the order by column is involved in equijoin with outer optimizable's
indexed column. If yes, then we know that since outer optimizable is ordered, the rows qualified
via the equijoin will also be ordered and hence sorting can be avoided. But if this is not
true, then we can't rely on outer optimizables' rows to be ordered on the order by column.
To avoid sorting, we need to identify this case 2) as another case when the column can be
marked as always ordered and that is when there is an equijoin predicate on the order by column
with some other column
which is already known to be always ordered. Taking the query from wisconsin as an example
will explain this behavior
select * from --DERBY-PROPERTIES joinOrder=FIXED 
where TENKTUP1.unique1 = TENKTUP2.unique1 
order by TENKTUP1.unique1, TENKTUP2.unique1; 

For the above query, as per the current trunk codeline, none of the order by columns are marked
as always ordered because there is no constant comparison predicate on them. But, for the
given join order, with TENKTUP1 as the outermost resultset and with the index TK1UNIQUE1,
we know that the current row ordering at this point is going to ensure that rows from TENKTUP1
are ordered on UNIQUE1. Next, when we process TEKTUP2 in the 2nd join order position, we find
that there is no constant predicate on TENKTUP2.unique1 and hence we conculde that the rows
from TENKTUP2 are not going to be ordered and we decide to force a sort node on the top of
the query. But in reality, even though the outer optimizable is not a single row resultset,
it is ordered on TENKTUP1.unique1 and hence all those rows from outer optimizable are going
to be ordered on TENKTUP1.unique1 and the inner optimizable has an equality join on 

TENKTUP1.unique1 using the order by column TENKTUP2.unique1 What that translates to is that
even if there will be multiple scans into TENKTUP2, the rows qualified are going to be all
ordered because of the equijoin between the outer and inner optimizables on the order by columns.
So, with my latest patch, I have expanded the notion of always ordered columns to include
both constant comparison predicates AND ordered column that has equijoin with an outer optimizable's
ordered column. 

I think this patch is also improving the existing queries to include a better path than what
it was picking up before. Following is an example of one such query from wisconsin.
select * from TENKTUP1, TENKTUP2 
where TENKTUP1.unique1 = TENKTUP2.unique1 
and TENKTUP2.unique1 < 100
order by TENKTUP1.unique1; 
For this query, the trunk currently decides to use TENKTUP1 as the outermost optimizable using
the TK1UNIQUE1 index and then those rows are filtered using TENKTUP2.unique1 < 100. Each
of the 2 tables involved in the query have 10000 rows each. So we are going through 10000
qualified indexed rows from TENKTUP1 and then applying TENKTUP2.unique1 < 100 on them.
With the attached patch, we use TENKTUP2 as the outermost optimizable with the index TK2UNIQUE1
and only gets the indexed rows which satisfy TENKTUP2.unique1 < 100 and then on them, we
use the equlity join to fetch qualified rows from TENKTUP1. 

I hope the above explanation helps understand the patch. I would appreciate if someone can
take the time to go through the patch and provide any feedback they may have. If I don't hear
anything by early next week, I will go ahead and commit the patch.

> Incorrect ORDER BY caused by index
> ----------------------------------
>                 Key: DERBY-3926
>                 URL: https://issues.apache.org/jira/browse/DERBY-3926
>             Project: Derby
>          Issue Type: Bug
>          Components: SQL
>    Affects Versions:,,,
>            Reporter: Tars Joris
>            Assignee: Mamta A. Satoor
>         Attachments: d3926_repro.sql, derby-reproduce.zip, DERBY3926_notforcheckin_patch1_051109_diff.txt,
DERBY3926_notforcheckin_patch1_051109_stat.txt, DERBY3926_notforcheckin_patch2_051109_diff.txt,
DERBY3926_patch3_051509_diff.txt, DERBY3926_patch3_051509_stat.txt, DERBY3926_patch4_051519_diff.txt,
DERBY3926_patch4_051519_stat.txt, DERBY3926_patch5_052709_diff.txt, DERBY3926_patch5_052709_stat.txt,
DERBY3926_patch6_060309_diff.txt, DERBY3926_patch6_060309_stat.txt, script3.sql, script3WithUserFriendlyIndexNames.sql,
> I think I found a bug in Derby that is triggered by an index on a large column: VARCHAR(1024).
I know it  is generally not a good idea to have an index on such a large column.
> I have a table (table2) with a column "value", my query orders on this column but the
result is not sorted. It is sorted if I remove the index on that column.
> The output of the attached script is as follows (results should be ordered on the middle
> ID                  |VALUE        |VALUE
> ----------------------------------------------
> 2147483653          |000002       |21857
> 2147483654          |000003       |21857
> 4294967297          |000001       |21857
> While I would expect:
> ID                  |VALUE        |VALUE
> ----------------------------------------------
> 4294967297          |000001       |21857
> 2147483653          |000002       |21857
> 2147483654          |000003       |21857
> This is the definition:
> CREATE INDEX key1 ON table1(id);
> CREATE TABLE table2 (id BIGINT NOT NULL, name VARCHAR(40) NOT NULL, value VARCHAR(1024),
PRIMARY KEY(id, name));
> CREATE UNIQUE INDEX key2 ON table2(id, name);
> CREATE INDEX key3 ON table2(value);
> This is the query:
> SELECT table1.id, m0.value, m1.value
> FROM table1, table2 m0, table2 m1
> WHERE table1.id=m0.id
> AND m0.name='PageSequenceId'
> AND table1.id=m1.id
> AND m1.name='PostComponentId'
> AND m1.value='21857'
> ORDER BY m0.value;
> The bug can be reproduced by just executing the attached script with the ij-tool.
> Note that the result of the query becomes correct when enough data is changed. This prevented
me from creating a smaller example.
> See the attached file "derby-reproduce.zip" for sysinfo, derby.log and script.sql.
> Michael Segel pointed out:
> "It looks like its hitting the index ordering on id,name from table 2 and is ignoring
the order by clause."

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message