Hi,

In 2016 I noticed that cube's Guttman polynomial split can produce
very unbalanced GiST trees [0], but never got around to fixing it.  The
problem is a bit dramatic for sorted input: repeated splits make
insertion quadratic and produce an enormous index.

For example, I inserted sorted two-dimensional points with an O2 build
without assertions.  These are medians of three runs:

 rows | master time | patched time | master pages | patched pages
------+-------------+--------------+--------------+--------------
 1000 |      5.91 s |      0.0119 s |         1757 |             9
 2000 |     24.74 s |      0.0150 s |         7929 |            18
 4000 |    110.63 s |      0.0219 s |        33712 |            36

The workload was:

CREATE TABLE t (id int, c cube);
CREATE INDEX ON t USING gist (c);
INSERT INTO t
SELECT g, cube(ARRAY[g::float8 / 4000, g::float8 / 4000 + 1])
FROM generate_series(1, 4000) g;

The patch generalizes the double-sorting split used by the built-in box
and point GiST operator classes to arbitrary cube dimensionality.  The
first patch adds regression coverage. The second implements the new
split.

PFA the patch set.


Best regards, Andrey Borodin

[0] 
https://www.postgresql.org/message-id/CAJEAwVEVq9Ry7KxApHbFTB67E0ntn2EBTQiUXBG%3Dq5E4P7w1VQ%40mail.gmail.com

Attachment: 0001-Add-regression-coverage-for-cube-GiST-page-splits.patch
Description: Binary data

Attachment: 0002-Use-double-sorting-for-cube-GiST-page-splits.patch
Description: Binary data


Reply via email to