BUG #2196: Useless RECHECK on RTREE index

Started by A Gattikerabout 20 years ago2 messagesbugs
Jump to latest
#1A Gattiker
agattik@gmail.com

The following bug has been logged online:

Bug reference: 2196
Logged by:
Email address: agattik@gmail.com
PostgreSQL version: 8.1.2
Operating system: alphaev68-dec-osf5.1a
Description: Useless RECHECK on RTREE index
Details:

create table tpoints (x int, y int);
CREATE INDEX i_tpoints ON tpoints USING RTREE (box(point(x,y),
point(x,y)));
explain select * from tpoints where box '((0,0),(1,1))' && box(point(x,y),
point(x,y));

QUERY PLAN

----------------------------------------------------------------------------
----------------------------------------------------------------------------
---
Bitmap Heap Scan on tpoints (cost=1.05..10.46 rows=10 width=8)
Recheck Cond: ('(1,1),(0,0)'::box && box(point((x)::double precision,
(y)::double precision), point((x)::double precision, (y)::double
precision)))
-> Bitmap Index Scan on i_tpoints (cost=0.00..1.05 rows=10 width=0)
Index Cond: ('(1,1),(0,0)'::box && box(point((x)::double precision,
(y)::double precision), point((x)::double precision, (y)::double
precision)))
(4 rows)

The RECHECK is necessary in case of polygons, but useless in case of box
overlap.

pg_amop.amopreqcheck is correctly set to false, but the planner seems to
ignore that.

select pg_amop.* from pg_amop join pg_opclass c ON (amopclaid=c.oid) join
pg_operator op ON (amopopr=op.oid) where opcname='box_ops' and
oprname='&&';
amopclaid | amopsubtype | amopstrategy | amopreqcheck | amopopr
-----------+-------------+--------------+--------------+---------
425 | 0 | 3 | f | 500
2593 | 0 | 3 | f | 500
(2 rows)

#2Tom Lane
tgl@sss.pgh.pa.us
In reply to: A Gattiker (#1)
Re: BUG #2196: Useless RECHECK on RTREE index

"" <agattik@gmail.com> writes:

Bitmap Heap Scan on tpoints (cost=1.05..10.46 rows=10 width=8)
Recheck Cond: ('(1,1),(0,0)'::box && box(point((x)::double precision,
(y)::double precision), point((x)::double precision, (y)::double
precision)))
-> Bitmap Index Scan on i_tpoints (cost=0.00..1.05 rows=10 width=0)
Index Cond: ('(1,1),(0,0)'::box && box(point((x)::double precision,
(y)::double precision), point((x)::double precision, (y)::double
precision)))
(4 rows)

The RECHECK is necessary in case of polygons, but useless in case of box
overlap.

You don't understand what a bitmap scan's recheck condition is for.
The above plan is correct.

regards, tom lane