database/postgres
1
0
Fork 0
mirror of https://github.com/postgres/postgres.git synced 2025-07-27 12:41:57 +03:00
Code Activity

Cube extension kNN support

Browse Source 
Introduce distance operators over cubes:
<#> taxicab distance
<->  euclidean distance
<=> chebyshev distance

Also add kNN support of those distances in GiST opclass.

Author: Stas Kelvich
This commit is contained in:
Teodor Sigaev
2015-12-18 14:38:27 +03:00
parent 3d0c50ffa0
commit 33bd250f6c
12 changed files with 1684 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

96
doc/src/sgml/cube.sgml
View File

@ -75,6 +75,8 @@
entered in. The <type>cube</> functions
automatically swap values if needed to create a uniform
<quote>lower left &mdash; upper right</> internal representation.
When corners coincide cube stores only one corner along with a
special flag in order to reduce size wasted.
</para>
<para>
@ -131,6 +133,19 @@
<entry><literal>a &lt;@ b</></entry>
<entry>The cube a is contained in the cube b.</entry>
</row>
<row>
<entry><literal>a -&gt; n</></entry>
<entry>Get n-th coordinate of cube.</entry>
</row>
<row>
<entry><literal>a ~&gt; n</></entry>
<entry>
Get n-th coordinate in 'normalized' cube representation. Noramlization
means coordinate rearrangement to form (lower left, upper right).
</entry>
</row>
</tbody>
</tgroup>
</table>
@ -143,6 +158,87 @@
data types!)
</para>
<para>
GiST index can be used to retrieve nearest neighbours via several metric
operators. As always any of them can be used as ordinary function.
</para>
<table id="cube-gistknn-operators">
<title>Cube GiST-kNN Operators</title>
<tgroup cols="2">
<thead>
<row>
<entry>Operator</entry>
<entry>Description</entry>
</row>
</thead>
<tbody>
<row>
<entry><literal>a &lt;-&gt; b</></entry>
<entry>Euclidean distance between a and b</entry>
</row>
<row>
<entry><literal>a &lt;#&gt; b</></entry>
<entry>Taxicab (L-1 metric) distance between a and b</entry>
</row>
<row>
<entry><literal>a &lt;=&gt; b</></entry>
<entry>Chebyshev (L-inf metric) distance between a and b</entry>
</row>
</tbody>
</tgroup>
</table>
<para>
Selection of nearing neigbours can be done in the following way:
</para>
<programlisting>
SELECT c FROM test
ORDER BY cube(array[0.5,0.5,0.5])<->c
LIMIT 1;
</programlisting>
<para>
Also kNN framework allows us to cheat with metrics in order to get results
sorted by selected coodinate directly from the index without extra sorting
step. That technique significantly faster on small values of LIMIT, however
with bigger values of LIMIT planner will switch automatically to standart
index scan and sort.
That behavior can be achieved using coordinate operator
(cube c)~&gt;(int offset).
</para>
<programlisting>
=> select cube(array[0.41,0.42,0.43])~>2 as coord;
coord
-------
0.42
(1 row)
</programlisting>
<para>
So using that operator as kNN metric we can obtain cubes sorted by it's
coordinate.
</para>
<para>
To get cubes ordered by first coordinate of lower left corner ascending
one can use the following query:
</para>
<programlisting>
SELECT c FROM test ORDER BY c~>1 LIMIT 5;
</programlisting>
<para>
And to get cubes descending by first coordinate of upper right corner
of 2d-cube:
</para>
<programlisting>
SELECT c FROM test ORDER BY c~>3 DESC LIMIT 5;
</programlisting>
<para>
The standard B-tree operators are also provided, for example