We deliberately use floats for the index (and therefore the boxes)
because it makes the indexes 50% smaller (and therefore more cachable,
and therefore faster for large databases). When the boxes are computed
against the double inputs, they are ajusted outwards, to they are always
slightly overdetermined... the index may pull slightly more values than
it "needs" to, but it will never accidentally miss a record due to our
type-truncation.
We feel that being a little overdetermined is well worth it for the 50%
size reduction, particularly since the r-tree is almost always
overdetermined anyways, and requires a second pass with a "real" spatial
function like st_contains() or st_distance().
p
Bernhard Reimar Hoefle wrote:
Hi!
I do want to use the GiST index for Points and Multipoints. As it works on
BBoxes, the BBoxes should be as small as possible.
If I'm right BBox coordinates are stored as float4. For large coordinates, such
as 3212345611.34 the box coordinates are rounded/cut? to 3212345600.00. I guess
that this behaviour cannot be changed because my coordinates have always more
than 8 digits and therefore BBoxes in the spatial index overlap?
But why do I get different results, already using small/reduced coordinates, for
POINT and MULTIPOINT objects?
Do I have to change the BBOXCACHE_BEHAVIOURS, if I want exact BBoxes, also for
MULTIPOINTs?
Example:
POINT(-2.730 -5.679 536.27) | BOX(-2.73000001907349 -5.68000030517578,-2.7299997
8065491 -5.67999982833862)
MULTIPOINT(-4.5599999986589 -6.08999999985099 552.17,-1.92000000178814
-4.12000000011176 536.28) | BOX(-6 -6.5,0 -4)
Thanks,
Bernhard
_______________________________________________
postgis-users mailing list
[email protected]
http://postgis.refractions.net/mailman/listinfo/postgis-users
--
Paul Ramsey
Refractions Research
http://www.refractions.net
[EMAIL PROTECTED]
Phone: 250-383-3022
Cell: 250-885-0632
_______________________________________________
postgis-users mailing list
[email protected]
http://postgis.refractions.net/mailman/listinfo/postgis-users
