Sorry to day is a really busy day…

 

I meant to say you should reflect on what you want to accomplish in your query.



 


From: Matt Doran [mailto:matt.doran@papercut.com]
Sent: Monday, December 03, 2007 4:34 PM
To: derby-user@db.apache.org
Cc: msegel@segel.com
Subject: Re: FW: Advice on *very* badly performing query

 

Hi Michael,

Michael Segel wrote:

The short simple answer… You get what you pay for.

 

The longer answer… Query optimization is a black art. Cloudscape was designed as a lightweight no frills embeddable DB.

Now you Cloudscape morphed in to Derby and JavaDB. But you lose the input from the folks at IBM who handle Query Optimization.

But I'm more than happy to work with the constraints of Derby, if only I could understand them.   And that's the help I was looking for here.

 

I have to run to a customer site, but using one of your examples… you noticed that the query performance changed when you had the field in the select columns as well as the where clause, but you didn’t when you had the field just in the where clause. So keep it in the selected fields. You could also try and change the order of the tables you’re joining.

I didn't change the where clause, just changing the select fields causes the dramatic query plan change.   I have a feeling it might be the fact that I'm selecting an attribute from the 5th join table ... but I'd like a better understanding of what's triggering the change so I can avoid it if possible.


And you may want to reflect your 5 table join.  Depending on the database and its tuning. Joining more than 3-4 tables can have a drastic negative impact on its performance.

I'm not sure what you mean to "reflect your 5 table join"?  

The fundamental issue here is that in this poor performing case, derby is not looking at the index on the very large table that would immediately reduce the dataset.   For whatever reason the optimizer is making a the worst possible case decision.

 And if speed really is important look at Informix (IDS 11) now offered by IBM.

Unfortunately as an off the shelf Java application that runs on Windows, Mac and Linux ... we really need a simple embedded DB that we can ship as the default.   Unfortunately Derby's query optimizer let's it down badly sometimes.

 

HTH

 

 


From: Matt Doran [mailto:matt.doran@papercut.com]
Sent: Sunday, December 02, 2007 11:06 PM
To: derby-user@db.apache.org
Subject: Advice on *very* badly performing query (with reproduction recipe)

 

Hi there,

We use Apache Derby in our commercial application, PaperCut NG.  It's proven to be very reliable, however we occasionally get reports of very bad performance in some areas.  We haven't had the time to investigate them fully previously (usually upgrading to an external DB like Postgres or SQL Server fixes the issue).  This time we had a look in more detail with a recent report, and we've found some very strange performance characteristics ... and would love some advice and assistance.

We have a query that is doing inner joins to 5 tables.  It's quite a simple query, but the core table has about 300,000 rows, and where limiting the results based on a date in that table that is indexed.  Here's a summary of my situation/findings:

  • Using the latest Derby release 10.3.1.4, with a Java 1.5 VM on Windows.
  • We only have a single WHERE clause, which is on the indexed date field is restricting the data such that no data is returned.  e.g. log_date > (latest log date).  So derby should quickly detect there is litte/no data to return.
  • Running the original query takes 22 minutes running 100% CPU.
  • Running a count(*) for the same query is quick (< 1 sec).
  • Removing the ORDER BY and changing the select list to just include a single field from each table and it still takes 22 minutes.
  • Changing the select list to retrieve only a single field from 2 of the table and it still takes 22 minutes (I have a log of the query and the runtime stats for this attached "derby-slow.log").
  • Changing the select list to a single field from 1 of the tables makes the query run fast - less than a second. (I have a log of the query and the runtime stats for this attached "derby-fast.log").
  • Running the original query on the same dataset in PostgreSQL or SQL Server is very fast (less than a second).  This is why we often recommend customers upsize to Postgres or SQL Server.
  • Also the SQL is generated via Hibernate ORM, so we have some limitations in how we can modify the SQL.


>From the query plan it seems that seems that it stops using the date index on the "tbl_printer_usage_log" log table, and changes from Hash joins to Nested Loop joins.  On a large table like this when providing a where clause that on a field that is indexed .... we have to ensure that derby uses the index.

If I increase the pageCacheSize to 100,000 pages, it reduces the time of the query to about 2-3 minutes, but it's still very slow compared to when the correct index is used.

Can anyone please help me understand the following:

  • Why does the query plan change dramatically, just by changing the fields that are retrieved?
  • Why is derby avoiding the most obvious index?  The date field in the 300,000 row table (the date field is the only field in our where clause).
  • Is there anyway to avoid this behavior?

If we can understand what's causing this, we'll be able to make a much more effective use of Derby.  At the moment, on customers with large datasets, we currently just recommend they "upsize" to Postgres or SQL Server and the problem goes away.  However, we'd much prefer to fix this and have our Derby database behave better.

I'd be happy to provide the derby database that exhibits these problems if someone would like to see what's going on.  The database is from a customer, so I don't want to post it publicly, but if you send me an email off-list I'd be happy to provide it.

Regards.


-- 
Matt Doran
PaperCut Software International Pty. Ltd.
Phone:   +61 (3) 9807 5767
E-mail:  matt.doran@papercut.com
Profile: http://www.papercut.com/about/#matt
Blog:    http://www.papercut.com/blog/