20111223

Push down query optimization in DataCleaner

As a follow-up to my previous post about how we make DataCleaner super-fast by applying some nice multi-threading tricks, The DataCleaner engine explained, I would now like to touch upon another performance booster: Push down query optimization.

To my knowledge "push down query optimization" is a trick that only very few tools support, since it requires a flow model that was actually built for it. The idea is that by inspecting an execution flow the tool might be able to identify steps in the beginning or in the end of the flow that can be replaced by query modifications.

For example, if your data flow begins with a filtering action that removes all records of a given type or restricts the further processing to only be the first 1000 records or something like that. Most tools simply require you to write some SQL yourself, which is also doable, but as I've said before on this blog, I think writing SQL is a barrier to productivity, creativity and good data quality results. So in DataCleaner we do not offer this option, because we have something that is much, much nicer. That solution is push down query optimization!

Let me illustrate. I will be using the Sakila example database for MySQL:

Say you want to do a simple pattern finding of film titles in the Sakila database, you would select the title column and you would get a result like this:

In the DataCleaner logs we can see what queries are actually fired to the database. Open up the log file (in the logs folder) and inspect datacleaner.log. You will find a line like this:

Executing query: SELECT `nicer_but_slower_film_list`.`title` FROM sakila.`nicer_but_slower_film_list`

That's fine. You can inspect the results closer, but that's not what this topic is about, so I'll carry on... Now let's say you want to refine your job. Let's instead see how the pattern distribution is if we want to only look at a few categories of films. So I add a 'Equals' filter to only select horror, sports and action movies and apply it to my pattern finder:

If we run the job, and inspect the log file again, we see now this entry:

Executing query: SELECT `nicer_but_slower_film_list`.`title`, `nicer_but_slower_film_list`.`category` FROM sakila.`nicer_but_slower_film_list` WHERE (`nicer_but_slower_film_list`.`category` = 'Horror' OR `nicer_but_slower_film_list`.`category` = 'Action' OR `nicer_but_slower_film_list`.`category` = 'Sports')

What's surprising here is that the filter actually got query optimized. Not all filters have this ability, since some of them have richer functionality than can be expressed as a query modification. But some of them do, and typically these are the small functions that make a big difference.

Let's also apply a Max rows filter that limits the analysis for only 20 records and chain it so that it depends on the Equals filter:

If we now run the job, both filters will have been applied to the query:

Executing query: SELECT `nicer_but_slower_film_list`.`title`, `nicer_but_slower_film_list`.`category` FROM sakila.`nicer_but_slower_film_list` WHERE (`nicer_but_slower_film_list`.`category` = 'Horror' OR `nicer_but_slower_film_list`.`category` = 'Action' OR `nicer_but_slower_film_list`.`category` = 'Sports') LIMIT 20

That means that we do as much as we can to optimize the query, without ever having to ask the user to help us. So if you modify the logical job, the physical queries are automatically adapted! This is why push down query optimization is a superior optimization technique to raw SQL. Happy data cleaning!

Additional information for developers: If you're developing plugins to DataCleaner and want to make a query optimized filter, then simply make sure you implement the QueryOptimizedFilter interface! Happy coding!

2 comments:

Matt Casters said...

Nice! For filters it makes sense in most but not all cases to filter in the database. However, making this optional seems like the thing to do since there are a lot of RDBMS situations out there, indexes, nework speed and so on need to be considered.
The thing is, I have unfortunately seen a lot of cases where the database filtered, summed, grouped and counted slower than the data could be read over JDBC. That in turn allows tools like DataCleaner and Kettle to work faster on the raw data in those cases... ;-)

Kasper Sørensen said...

Glad you like the approach Matt :) Actually another point that I wanted to add is that this optimization will also be pushed to non-RDBMS sources, eg. CSV files, since the DC query layer is programmatic, not SQL-based. So the materialization of the actual SQL query or the instructions to the CSV reader happens at a later stage (using MetaModel).

Obviously not all filters can use this approach. We also have some filters like the JavaScript filter which are completely scriptable by the end user and that is not translatable to a query. In those cases we simply process the filter in the engine.

You're right in that often time it's not even optimal to push transformations to the query. My general rule of thumb is: If it decreases the amount of data being transmitted (I/O), then do the optimization, if not, then do it in the row processing engine instead. I think you have the same strategy in Kettle? That's why we so far we don't have a "push down" optimization technique for transformers in DC, since these will grow the size (add columns) of the dataset and that will increase I/O between DB and client. And besides, DC's (and Kettle's) engine is multi-threaded, while most RDBMS'es dispatch every query in a single thread.