[BUG]: Python testing time doubled by test_equals_only_binpreds.py

[harrism]

Version

23.04

On which installation method(s) does this occur?

Docker, Source, Rapids-Compose

Describe the issue

equals-only binpreds pytests doubled cuSpatial's Python testing time (on my machine). Without test_equals_only_binpreds.py: 16.24s. With that file: 33.10s.

It would be nice to keep testing as fast as possible, so if we don't need such large test cases, let's reduce them.

Minimum reproducible example

No response

Relevant log output

No response

Environment details

No response

Other/Misc.

No response

[thomcom]

The main reason that these tests are slow is because they use some length 10000 object arrays, which take a non-trivial amount of time for Shapely to initialize. We have at least two choices:

1. Just test 100 features instead of 10000
2. Write gpu-based random feature generators (we have a couple of passes of this already in libcuspatial and cuSpatial) to generate the features instead of Shapely

[harrism]

I think large numbers should be benchmarks, not tests. On the other hand 100 is not very large. My general rule of thumb is "should be large enough to require more than one thread block in the kernels" (I believe this is in our C++ developer guide).

I prefer to do those larger tests in C++ when possible, since it's much faster. Probably not yet possible in the binpreds tests...

For now, I think we should consider both options and see if there's an easy solution that doesn't compromise coverage.