Thanks for your review!
On Sun, Jan 30, 2022 at 7:24 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Actually, that code has got worse problems than that. I tried to improve
> our regression tests to exercise that code path, as attached. What I got
> was
>
> +SELECT point(x,x), (SELECT circle_center(f1) FROM gcircle_tbl ORDER BY f1 <-> p
> oint(x,x) LIMIT 1) as c FROM generate_series(0,1000,1) x;
> +ERROR: index returned tuples in wrong order
I tried to figure out what is the problem with this query. This error
happens when actual distance is less than estimated distance.
For this specific query it happened while comparing these values:
50.263279680219099532223481219262 (actual distance returned by
dist_cpoint in geo_ops.c) and 50.263279680219113743078196421266
(bounding box distance returned by computeDistance in gistproc.c).
So for me it looks like this error is not really related to KNN scan
code but to some floating-arithmetic issue in distance calculation
functions for geometry type.Would be great to figure out a fix but
for now I didn’t manage to find a better way than comparing the
difference of distance with FLT_EPSILON which definitely doesn't seem
like the way to fix :(