db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Dyre Tjeldvoll (JIRA)" <j...@apache.org>
Subject [jira] Commented: (DERBY-827) Performance can be improved by re-using language ResultSets across Activation executions.
Date Mon, 19 Mar 2007 12:26:32 GMT

    [ https://issues.apache.org/jira/browse/DERBY-827?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12482065
] 

Dyre Tjeldvoll commented on DERBY-827:
--------------------------------------

I tried adding the following to BulkTableScanResultSet.close():

		// Set the candidate back to NULL to prepare for another open
		DataValueDescriptor[] c = candidate.getRowArray();
		if (c != null) {
			for (int i = 0; i < c.length; ++i) {
				if (c[i] == null) continue;
				c[i] = c[i].getNewNull();
			}
		}

I would have preferred to use c[i].setToNull(), since that would
reuse the DataValueDescriptor objects, (ideally one should be
able to reuse BulkTableScanResultSet.rowArray and associated
obejcts as well), but this doesn't work for all implementations
of DataValueDescriptor (e.g. HeapRowLoaction which will trigger
an ASSERT when doing this). 

Another alternative would be to switch on the real type of the
DataValueDescriptor and only call setToNull() for types that have
a stream (brittle and not very clean IMHO). Yet another
possibility would be null out the DVDs in the openCore() method
before cloning. This naturally leads to the question of why
openCore() wants a clone rather than a null copy...?

With this change suites.All has 5 failures and 0 errors. 1 of
those failures are the testOnceResultSet failure mentioned
previously.

> Performance can be improved by re-using language ResultSets across Activation executions.
> -----------------------------------------------------------------------------------------
>
>                 Key: DERBY-827
>                 URL: https://issues.apache.org/jira/browse/DERBY-827
>             Project: Derby
>          Issue Type: Improvement
>          Components: Performance
>            Reporter: Daniel John Debrunner
>         Attachments: derby827_draft_reuse_result_sets.txt, derby827_update920.txt, rsfromps.v1.diff,
rsfromps.v1.stat, rsfromps_prelim.diff, rsfromps_prelim2.diff
>
>
> >Shouldn't DistinctScalarAggregateRS implement a close or a finish method
> >>(not sure what the difference is) and close the scan controller there.
> The close() and finish() methods are actually explained in their javadoc
> in the language org.apache.derby.iapi.sql.ResultSet class.
> [note this is not a JDBC java.sql.ResultSet object]
> close() -  Tells the system that there will be no more calls to
> getNextRow() (until the next open() call)
> finish() - Tells the system that there will be no more access to any
> database information via this result set
> So close means the ResultSet may be opened again for more access, while
> finish means it will not be used again.
> However, their use in the code always doesn't match that, and that does
> cause confusion, at least to me.
> Language ResultSets (not JDBC ones) can be and are opened multiple
> times, for example when scanning a table multiple times within a join.
> An Activation, which represents the internal state of
> java.sql.PreparedStatement object & has the lifetime of the
> java.sql.PreparedStatement, contains a top-level language ResultSet.
> This top-level language ResultSet provides the execution of the SQL
> statement, DML, DDL or a query. The top-level ResultSet may contain
> other ResultSets and could be seen as a tree structure. For the simple
> case of a primary key lookup query like:
>    select name from customer where id = ?
> The activation would contain this:
> top result set
> ProjectRestrictRS << IndexRowToBaseRowRS << TableScanRS
> Now for some reason, even though the api of ResultSet say they can be
> re-used, and in some cases they are, this result set tree is thrown away
> after each execution. That is, the top result set has its finish()
> method called and then the activation removes its reference to it. Then
> on the next execution a new (identical) tree is set up.
> There is potential for a huge performance gain if this top level result
> set and its tree are re-used and have the same lifetime as the
> Activation. The saving comes in two forms, not having to create many
> objects on each execution, and not creating short-lived objects for the
> garbage collector to handle.
> I made a simple fix, it's a couple of lines of code, just calling close
> & finish at the correct times, and for the above simple primary key
> lookup query, the performance went from 17,300 to 24,000 selects per
> second (cached data, single user). I'll post a patch shortly as an
> indication of the direction, once I can separate it from other changes
> in my client.
> However, I'm running the Derby tests and there are some (maybe 25-30)
> failures, I think because not all the language ResultSet implementations
> are correctly written to be re-opened. Interestingly, the first failure
> I saw was in an aggregrate test, which goes back to the issue Manish saw.
> Even if derbyall passed I would be nervous about submitting this patch
> for real, because I don't think there's a lot of testing using repeat
> executions of PreparedStatements in the tests. The ij tests mainly use
> Statement, this is a single use of an activation so this change would
> not affect them. Thus such a patch could regress Derby by making it more
> likely existing bugs would be exposed.
> Given the performance gains, I think we need to start re-using
> ResultSets from Activation, and devise a way to ensure the testing
> covers the re-use. The main issue is there is a large number of
> ResultSet implementations to cover.

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


Mime
View raw message