Hello, hackers.
I've fixed split algorithm that was implemented in cube extension. I've changed it according to the original Guttman
paper(old version was more simple algorithm) and also ported Alexander Korotkov's algorithm from box datatype indexing
thatwork faster and better on low dimensions.
On my test dataset (1M records, 7 dimensions, real world database of goods) numbers was following:
Building index over table (on expression):
old: 67.296058 seconds
new: 48.842391 seconds
Cube point search, mean, 100 queries
old: 0.001025 seconds
new: 0.000427 seconds
Index on field size:
old: 562 MB
new: 283 MB
Stas.