Re: WIP: Hash Join-Filter Pruning using Bloom Filters

On Sun, Nov 2, 2008 at 5:36 PM, Hannes Eder <hannes@hanneseder.net> wrote:
> On Sun, Nov 2, 2008 at 10:49 PM, Jonah H. Harris <jonah.harris@gmail.com> wrote:
>> Similarly, I
>> created a GUC to enable pruning, named bloom_pruning.
>
> I guess calls to bloom_filter_XXX should be surrounded by "if
> (bloom_pruning) ..." or a similar construct, i.e. make use of the GUC
> variable bloom_pruning in the rest of the code.

It's effective as-is for a preliminary patch.  The GUC code is the
least of my worries.

> Can you provide some figures on the performance impact of the bloom filter?

It depends on the queries.  I've been trying to find a good suite of
hash join tests... but not much luck.

CREATE TABLE t1 (id INTEGER PRIMARY KEY, x INTEGER);
CREATE TABLE t2 (id INTEGER PRIMARY KEY, x INTEGER);
INSERT INTO t1 (SELECT ge, ge % 100 FROM generate_series(1, 1000000) ge);
INSERT INTO t2 (SELECT * FROM t1);
VACUUM ANALYZE;
SELECT COUNT(*) FROM t1, t2WHERE t1.id = t2.id      AND t1.x < 30      AND t2.x > 10;
SET bloom_pruning TO off;
EXPLAIN
SELECT COUNT(*) FROM t1, t2WHERE t1.id = t2.id      AND t1.x < 30      AND t2.x > 10;
\timing
SELECT COUNT(*) FROM t1, t2WHERE t1.id = t2.id      AND t1.x < 30      AND t2.x > 10;
\timing
EXPLAIN
SELECT * FROM t1, t2WHERE t1.id = t2.id      AND t1.x < 30      AND t2.x > 10;
\timing
SELECT * FROM t1, t2WHERE t1.id = t2.id      AND t1.x < 30      AND t2.x > 10;
\timing
SET bloom_pruning TO on;
\timing
SELECT COUNT(*) FROM t1, t2WHERE t1.id = t2.id      AND t1.x < 30      AND t2.x > 10;
\timing
EXPLAIN
SELECT * FROM t1, t2WHERE t1.id = t2.id      AND t1.x < 30      AND t2.x > 10;
\timing
SELECT * FROM t1, t2WHERE t1.id = t2.id      AND t1.x < 30      AND t2.x > 10;
\timing

-- Without Pruning
Time: 1142.843 ms
Time: 1567.355 ms

-- With Pruning
Time: 891.557 ms
Time: 1269.634 ms



-- 
Jonah H. Harris, Senior DBA
myYearbook.com