Cake slice

[Algunenano]

Trac: https://trac.osgeo.org/postgis/ticket/4676

Some performance improvements, mostly based around the PG12+ ability of retrieve only a slice of a toasted object.

Improvements for big geometries:

• ST_SRID

-- explain analyze Select ST_SRID(the_geom) from benchmark_4c7214d90a79aa6760367a084a4d4a2f61fbe1c6cc4f7f9e76020;
-- Before latency average = 16.983 ms
-- After latency average = 0.145 ms
-- Perf: 117 x

• ST_NDims:

EXPLAIN ANALYZE SELECT ST_NDims(the_geom) from benchmark_4c7214d90a79aa6760367a084a4d4a2f61fbe1c6cc4f7f9e76020;
-- Before latency average = 16.635 ms
-- After latency average = 0.143 ms 
-- Perf: 116.32x

• Equivalent improvements for ST_zmflag, postgis_hasbbox, SE_Is3D, SE_IsMeasured

Improvements for point (cast from postgis point to postgresql point):

-- EXPLAIN ANALYZE Select point(the_geom) from benchmark_7773a711c8441d4b494a51fd9feebeac7a9b9c734619398620293;
-- Before latency average = 215.468 ms
-- After latency average = 164.299 ms
-- Perf: 1.31x

Improvements for big geometry comparison

This affects any function using gserialized_datum_get_box2df_p, which are all the gist ones.

huge_polygon && huge_polygon

-- explain analyze Select a.the_geom && a.the_geom from benchmark_4c7214d90a79aa6760367a084a4d4a2f61fbe1c6cc4f7f9e76020 a, benchmark_4c7214d90a79aa6760367a084a4d4a2f61fbe1c6cc4f7f9e76020 b;
-- Before latency average = 423.047 ms
-- After latency average = 0.886 ms
-- Perf: 477.47x

point && point maintains performance

-- explain analyze with points as (Select the_geom from benchmark_7773a711c8441d4b494a51fd9feebeac7a9b9c734619398620293 limit 1000) Select a.the_geom && b.the_geom from points a, points b;
-- Before latency average = 321.904 ms
-- After commit to avoid re-reads = 323.928 ms
-- Perf: 0.993x

Example with a query to read a tile

Table with index created but not selected:

Indexes:
    "lpr_000b16a_e_1_pkey" PRIMARY KEY, btree (cartodb_id)
    "lpr_000b16a_e_1_the_geom_idx" gist (the_geom)
    "lpr_000b16a_e_1_the_geom_webmercator_idx" gist (the_geom_webmercator)

Query:

explain analyze Select count(*) from benchmark_4c7214d90a79aa6760367a084a4d4a2f61fbe1c6cc4f7f9e76020 where the_geom_webmercator && ST_TileEnvelope(3,1,1);

• Pgbench:

explain analyze Select count(*) from benchmark_4c7214d90a79aa6760367a084a4d4a2f61fbe1c6cc4f7f9e76020 where the_geom_webmercator && ST_TileEnvelope(3,1,1);
-- Before latency average = 16.845 ms
-- After latency average = 0.209 ms
-- Perf: 80.59x

• Plan before:

 Aggregate  (cost=4.39..4.40 rows=1 width=8) (actual time=34.918..34.918 rows=1 loops=1)
   ->  Seq Scan on benchmark_4c7214d90a79aa6760367a084a4d4a2f61fbe1c6cc4f7f9e76020  (cost=0.00..4.38 rows=3 width=0) (actual 
time=20.303..34.907 rows=3 loops=1)
         Filter: (the_geom_webmercator && '0103000020110F00000100000005000000DC183A68F4A96CC193107C45F81B6341DC183A68F4A96CC1
DC183A68F4A96C4193107C45F81B63C1DC183A68F4A96C4193107C45F81B63C193107C45F81B6341DC183A68F4A96CC193107C45F81B6341'::geometry)
         Rows Removed by Filter: 10
 Planning Time: 0.170 ms
 Execution Time: 34.958 ms

• Plan afte is the same, just faster:

 Aggregate  (cost=4.39..4.40 rows=1 width=8) (actual time=0.159..0.159 rows=1 loops=1)
   ->  Seq Scan on benchmark_4c7214d90a79aa6760367a084a4d4a2f61fbe1c6cc4f7f9e76020  (cost=0.00..4.38 rows=3 width=0) (actual 
time=0.134..0.153 rows=3 loops=1)
         Filter: (the_geom_webmercator && '0103000020110F00000100000005000000DC183A68F4A96CC193107C45F81B6341DC183A68F4A96CC1
DC183A68F4A96C4193107C45F81B63C1DC183A68F4A96C4193107C45F81B63C193107C45F81B6341DC183A68F4A96CC193107C45F81B6341'::geometry)
         Rows Removed by Filter: 10
 Planning Time: 0.150 ms
 Execution Time: 0.195 ms

[pramsey]

👍