[postgis-users] ST_DWithin slow if query is joined to another table, but fast if ST_DWithin cast to an integer

Mon Nov 27 09:15:46 PST 2023

I’m guessing it has to do with the costing we have on ST_DWithin.

I recall we increased that back recently because of some complaints but can’t recall the version.

What does your costing read from ST_DWithin(geography, geography, …) read?

Run this query to see the code:

SELECT pg_get_functiondef(oid)

FROM pg_proc

WHERE proname = 'st_dwithin' AND prosrc = 'geography_dwithin';

Mine reads

    COST 5000

    IMMUTABLE STRICT PARALLEL SAFE     SUPPORT postgis_index_supportf

Also what happens if you switch your query to put the ST_Dwithin in the JOIN clause instead of the WHERE?

from building b join customer c ON (b.id = c.building_id  AND  st_dwithin(st_makepoint(b.longitude, b.latitude)::geography, st_makepoint(-96.7804060, 33.2471770)::geography, 50000) );

From: postgis-users <postgis-users-bounces at lists.osgeo.org> On Behalf Of Cameron McCloud via postgis-users
Sent: Monday, November 27, 2023 8:32 AM
To: PostGIS Users Discussion <postgis-users at lists.osgeo.org>
Cc: Cameron McCloud <cameron.mccloud at gmail.com>
Subject: Re: [postgis-users] ST_DWithin slow if query is joined to another table, but fast if ST_DWithin cast to an integer

Hi,

Just pinging this again to see if anyone has an idea of what's going on. >From the EXPLAIN output of the second query in the DBFiddle example it does seem that there is a cross-join going on:

 Rows Removed by Join Filter: 101705643

, even though there is an "ON" clause in the join.

Cam.

On Thu, Nov 9, 2023 at 10:27 AM Cameron McCloud <cameron.mccloud at gmail.com> wrote:

https://dbfiddle.uk/qJDW-DjP

On Thu, Nov 9, 2023 at 9:11 AM Cameron McCloud <cameron.mccloud at gmail.com <mailto:cameron.mccloud at gmail.com> > wrote:

Here's the EXPLAIN for the second query with the join and WHERE on ST_DWithin:

 Nested Loop  (cost=0.00..403093.00 rows=2 width=20)
   Join Filter: (b.id <http://b.id>  = c.building_id)
   ->  Seq Scan on customer c  (cost=0.00..231.00 rows=16000 width=4)
   ->  Materialize  (cost=0.00..402382.01 rows=2 width=20)
         ->  Seq Scan on building b  (cost=0.00..402382.00 rows=2 width=20)
               Filter: st_dwithin((st_makepoint((longitude)::double precision, (latitude)::double precision))::geography, '0101000020E61000008AE6012CF23158C065A9F57EA39F4040'::geography, '50000'::double precision, true)
 JIT:
   Functions: 7
   Options: Inlining false, Optimization false, Expressions true, Deforming true

Here's the EXPLAIN ANALYZE:

 Nested Loop  (cost=0.00..403093.00 rows=2 width=20) (actual time=29.391..5196.713 rows=6453 loops=1)
   Join Filter: (b.id <http://b.id>  = c.building_id)
   Rows Removed by Join Filter: 82424169
   ->  Seq Scan on customer c  (cost=0.00..231.00 rows=16000 width=4) (actual time=0.004..2.484 rows=16000 loops=1)
   ->  Materialize  (cost=0.00..402382.01 rows=2 width=20) (actual time=0.000..0.149 rows=5152 loops=16000)
         ->  Seq Scan on building b  (cost=0.00..402382.00 rows=2 width=20) (actual time=3.292..28.549 rows=6453 loops=1)
               Filter: st_dwithin((st_makepoint((longitude)::double precision, (latitude)::double precision))::geography, '0101000020E61000008AE6012CF23158C065A9F57EA39F4040'::geography, '50000'::double precision, true)
               Rows Removed by Filter: 9547
 Planning Time: 0.182 ms
 JIT:
   Functions: 7
   Options: Inlining false, Optimization false, Expressions true, Deforming true
   Timing: Generation 0.256 ms, Inlining 0.000 ms, Optimization 0.158 ms, Emission 3.133 ms, Total 3.547 ms
 Execution Time: 5197.586 ms

Here's the EXPLAIN for the 3rd query (JOIN but convert ST_DWithin to integer)

 Hash Join  (cost=431.00..402854.10 rows=80 width=20)
   Hash Cond: (b.id <http://b.id>  = c.building_id)
   ->  Seq Scan on building b  (cost=0.00..402422.00 rows=80 width=20)
         Filter: (CASE st_dwithin((st_makepoint((longitude)::double precision, (latitude)::double precision))::geography, '0101000020E61000008AE6012CF23158C065A9F57EA39F4040'::geography, '50000'::double precision, true) WHEN CASE_TEST_EXPR THEN 1 ELSE 0 END = 1)
   ->  Hash  (cost=231.00..231.00 rows=16000 width=4)
         ->  Seq Scan on customer c  (cost=0.00..231.00 rows=16000 width=4)
 JIT:
   Functions: 12
   Options: Inlining false, Optimization false, Expressions true, Deforming true

And the EXPLAIN ANALYZE:

 Hash Join  (cost=431.00..402854.10 rows=80 width=20) (actual time=6.470..33.691 rows=6453 loops=1)
   Hash Cond: (b.id <http://b.id>  = c.building_id)
   ->  Seq Scan on building b  (cost=0.00..402422.00 rows=80 width=20) (actual time=4.280..30.210 rows=6453 loops=1)
         Filter: (CASE st_dwithin((st_makepoint((longitude)::double precision, (latitude)::double precision))::geography, '0101000020E61000008AE6012CF23158C065A9F57EA39F4040'::geography, '50000'::double precision, true) WHEN CASE_TEST_EXPR THEN 1 ELSE 0 END = 1)
         Rows Removed by Filter: 9547
   ->  Hash  (cost=231.00..231.00 rows=16000 width=4) (actual time=2.173..2.174 rows=16000 loops=1)
         Buckets: 16384  Batches: 1  Memory Usage: 691kB
         ->  Seq Scan on customer c  (cost=0.00..231.00 rows=16000 width=4) (actual time=0.007..0.931 rows=16000 loops=1)
 Planning Time: 0.132 ms
 JIT:
   Functions: 12
   Options: Inlining false, Optimization false, Expressions true, Deforming true
   Timing: Generation 0.311 ms, Inlining 0.000 ms, Optimization 0.184 ms, Emission 4.085 ms, Total 4.580 ms
 Execution Time: 34.239 ms

On Wed, Nov 8, 2023 at 6:10 PM Regina Obe <lr at pcorp.us <mailto:lr at pcorp.us> > wrote:

First of all is that really your join clause?  You seem to be missing an ON

from building b

join customer c

Also please output

EXPLAIN 

And the 

EXPLAIN ANALYZE of each query

From: postgis-users <postgis-users-bounces at lists.osgeo.org <mailto:postgis-users-bounces at lists.osgeo.org> > On Behalf Of Cameron McCloud via postgis-users
Sent: Wednesday, November 8, 2023 5:38 AM
To: postgis-users at lists.osgeo.org <mailto:postgis-users at lists.osgeo.org> 
Cc: Cameron McCloud <cameron.mccloud at gmail.com <mailto:cameron.mccloud at gmail.com> >
Subject: Re: [postgis-users] ST_DWithin slow if query is joined to another table, but fast if ST_DWithin cast to an integer

Test code is here: https://github.com/cameronmccloud/postgis_slow_st_dwithin/blob/main/test.sql

On Wed, Nov 8, 2023 at 9:25 AM Cameron McCloud <cameron.mccloud at gmail.com <mailto:cameron.mccloud at gmail.com> > wrote:

Hi,

Postgis Version: 3.3.4

Postgres Version: 14.9

We have 2 tables, "building" with lat/long and "customer" with a FK to building. There's a 1:1 relationship between the two.

The test tables we're using have 16K rows each. Our production data has a lot more, but we could reproduce this on a smaller dataset.

We found some odd behaviour when using ST_DWITHIN in a WHERE clause but only when the "building" table is joined to the "customer" table.

We also found that converting the result of ST_DWITHIN to an integer (1/0 for true/false) and using the integer in the WHERE is fast.

-- This query works as expected and takes 60ms returning 6K rows.

select b.*

from building b

where

  st_dwithin(st_makepoint(b.longitude, b.latitude)::geography, st_makepoint(-96.7804060, 33.2471770)::geography, 50000);

-- This query is orders of magnitude slower - 3000ms, even though joining the two tables without the WHERE takes 30ms

select b.*

from building b

join customer c

where

  st_dwithin(st_makepoint(b.longitude, b.latitude)::geography, st_makepoint(-96.7804060, 33.2471770)::geography, 50000);

-- This query converts the result of ST_DWITHIN to an integer. It's fast and takes 80ms

select b.*

from building b

join customer c

where

  case st_dwithin(st_makepoint(b.longitude, b.latitude)::geography, st_makepoint(-96.7804060, 33.2471770)::geography, 50000)

    when true then 1

    else 0

  end = 1;

We have no idea why this is the case, but we're curious. In our production scenario using the "case...when" brought a query down from 6 minutes to 6 seconds. We'd love to know why this might be and if there are other scenarios like this that could increase the performance of our Postgis queries.

Thanks,

Cam.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/postgis-users/attachments/20231127/3e6525f7/attachment.htm>