[
http://issues.apache.org/jira/browse/DERBY-634?page=comments#action_12427670 ]
Satheesh Bandaram commented on DERBY-634:
-----------------------------------------
I propose twhere Table2.x=Table3.x and y > 100)
o address this issue in phases. Here is my current thinking:
Background information:
-----------------------
A performance optimization was introduced in Cloudscape before open sourcing as
Apache Derby.
Before this optimization was introduced, a query like:
Select a, b, c from Table1 where a> 5 and b in
(select x from Table2, Table3
would take longer time time to execute than needed. This becomes worse as the
complexity of the
subquery increases. Basic problem was that for every qualified value of 'b',
the subquery was
getting executed, recreating the results multiple times. Cloudscape, at that
time, had the ability
to materialize results of a subquery only if a single row is returned from the
subquery.
(where subquery is of the form select max(i) from Table2)
A performance optimization was introduced that allowed for some "small number"
of rows greater than
1 to be cached dynamically at runtime. As the subquery was executed first time,
results of the
subquery were cached until MAX_MEMORY_PER_TABLE is reached, which was 1MG by
default. If the results
of the subquery could be fit into memory less than this configurable size, a
nested union resultset
would be built to cache the results.
Future invocations of the subquery would simply return results from this
subquery cache without
actually evaluating subquery. This resulted in performance boost for a customer
query from 10 minutes
to a few seconds.
Side effect of this optimization:
---------------------------------
While the optimization worked well for the customer query, it is causing issues
for the query in
question here. If the subquery were returning just an integer, the optimization
could attempt to cache
1MG/4, about 250,000 rows in nested union resultsets. Nesting of this deep
would cause stack overflow.
Jeff Lichman also identified several other issues mentioned in the description
of this entry.
Proposed Fix:
-------------
First, I think it is important to fix regression caused by this optimization.
The optimization was
intended to cache small number of subquery results. Instead of caching single
row result of a subquery,
this optimization could be adjusted to cache a small number of results.
Second, caching results of subquery in nested union resultsets is not needed.
This can be rewritten to
generate a linear resultset, which would save runtime stack growth.
Third, as Jeff pointed out, a global subquery optimization that is performed
during optimization is
the best approach. He pointed out subquery materialization based on hash joins
decided during
optimization phase is the ideal solution. Fortunately, Army's optimizer
enhancements introduced
recently builds a subquery materialization capability to Derby and this could
be extended to handle
this case as well.
I propose to address the regression first by caching only small number of rows
dynamically. If number
of subquery results could cross 512, I think this optimization should be
dynamically disabled.
I will also file another Improvement that would suggest reworking the original
optimization to be
built on Army's subquery materialization framework. My current impression is
that Army's subquery
work needs to be more generalized and stabilized before extending to cover
other cases.
Let me know if anyone has any comments.
> Subquery materialization can cause stack overflow
> -------------------------------------------------
>
> Key: DERBY-634
> URL: http://issues.apache.org/jira/browse/DERBY-634
> Project: Derby
> Issue Type: Bug
> Components: SQL
> Affects Versions: 10.1.1.1
> Reporter: Jeff Lichtman
> Fix For: 10.2.0.0
>
>
> A performance optimization in subquery processing can cause a stack overflow.
> The optimization materializes a subquery ResultSet in memory where it thinks
> the rows will fit in memory. The materialization is done as a set of nested
> unions of constant rows (UnionResultSets and RowResultSets). If there are a
> lot of rows this can cause a stack overflow when fetching a row.
> The obvious fix is to make it use an iterative technique rather than a
> recursive one for storing and returning the rows. See the method
> BaseActivation.materializeResultSetIfPossible() in the language execution
> code.
> There are some other issues with this performance optimization that should be
> looked at:
> 1) The optimization can backfire, making the query run much slower. For
> example, in the query:
> select * from one_row_table where column1 not in
> (select column2 from million_row_table)
> reading million_row_table into memory is an expensive operation. If there is
> an index on million_row_table.column2, the query should return a result very
> quickly despite the large size of million_row_table by doing a single probe
> into million_row_table via the index.
> Since in-memory materialization can be an expensive operation, the decision
> about whether to do it should be made based on query optimizer cost
> estimates. See SubqueryNode.generateExpression().
> 2) It may not be wise to cache partial query results in memory at all.
> Although this can help performance in some cases, it also chews up memory.
> This is different from a limited-size cache with a backing store (like what
> the store uses for page caching). The language has no way to limit the total
> amount of memory used in this type of processing. Note that hash joins
> originally used in-memory hash tables with no backing store, and that a
> backing store was added later.
> 3) The implementation of this optimization has some problems. The decision to
> materialize the subquery results in memory is made during code generation -
> all such decisions should be made during the compilation phase. It's not
> clear to me why materializeResultSetIfPossible() is in BaseActivation - I
> would expect the of materialization to be done by a type of ResultSet, not by
> a method in BaseActivation. Also, this method calls getMaxMemoryPerTable() in
> the OptimizerFactory - nothing in the execution code should refer to anything
> in the compilation code (perhaps getMaxMemoryPerTable() should be moved
> somewhere else).
--
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators:
http://issues.apache.org/jira/secure/Administrators.jspa
-
For more information on JIRA, see: http://www.atlassian.com/software/jira
