Switch pytest traceback to native

[galipremsagar]

Description

In cudf & cuml we have observed a ~10% to ~20% respectively speed up of pytest suite execution by switching pytest traceback to --native:

currently:

102474 passed, 2117 skipped, 902 xfailed in 892.16s (0:14:52)

--tb=short:

102474 passed, 2117 skipped, 902 xfailed in 898.99s (0:14:58)

--tb=no:

102474 passed, 2117 skipped, 902 xfailed in 815.98s (0:13:35)

--tb=native:

102474 passed, 2117 skipped, 902 xfailed in 820.92s (0:13:40)

This PR makes similar change to dask-cuda repo.

xref: rapidsai/cudf#16851

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[pentschev]

Do you have more information exactly how does this compare to --tb=auto? From the docs:

pytest --tb=auto    # (default) 'long' tracebacks for the first and last
                     # entry, but 'short' style for the other entries
pytest --tb=long    # exhaustive, informative traceback formatting
pytest --tb=short   # shorter traceback format

It seems that with auto we get "exhaustive" traceback (not really sure what that means), is that the case with native too? Although 11.8.0 builds seem faster, 12.5.1 do not (see table below), still waiting on 12.0.1 to complete, but in general I'm not sure there's conclusive benefit here to justify this change, in particular if it happens to omit useful parts of the traceback.

CUDA version	This change	Build PR #1385	Build PR #1384
11.8.0	57:11	63:28	65:07
12.0.1	48:25	47:58	49:51
12.5.1	N/A	74:39	84:57

EDIT: I had 12.0.1/12.5.1 results swapped, I've corrected that now.

[pentschev]

Based on the uncertainty from this PR as per my comment, as well as the seeming regression in performance in UCX-Py, I'm "requesting changes" until we can determine what's the correct way forward.

Please post any other thoughts and comments in the PR in the meantime.

[pentschev]

/merge