I have a rough proof-of-concept for getting nearest-neighbor searches
working with cubes. When I say "rough", I mean "I have no idea what I'm
doing and I haven't written C for 15 years but I hear it got standardized
please don't hurt me". It seems to be about 400x faster for a 3D cube with
1 million rows, more like 10-30x for a 6D cube with 10 million rows.
The patch adds operator <-> (which is just the existing cube_distance
function) and support function 8, distance (which is just g_cube_distance, a
wrapper around cube_distance).
The code is in no way production-quality; it is in fact right around "look!
it compiles!", complete with pasted-in, commented-out code from something I
was mimicking. I thought I'd share at this early stage in the hopes I might
get some pointers, such as:
- What unintended consequences should I be looking for?
- What benchmarks should I do?
- What kind of edge cases might I consider?
- I'm just wrapping cube_distance and calling it through DirectFunctionCall;
it's probably more proper to extract out the "real" function and call it
from both cube_distance and g_cube_distance. Right?
- What else don't I know? (Besides C, funny man.)
The patch, such as it is, is at:
https://github.com/jaylevitt/postgres/commit/9cae4ea6bd4b2e582b95d7e1452de0a7aec12857
with an even-messier test at
https://github.com/jaylevitt/postgres/commit/daa33e30acaa2c99fe554d88a99dd7d78ff6c784
I initially thought this patch made inserting and indexing slower, but then
I realized the fast version was doing 1 million rows, and the slow one did
10 million rows. Which means: dinnertime.
Jay Levitt