Performance regression in DataFrame reduction ops

[jorisvandenbossche]

The stat_ops.FrameOps are generally showing a slowdown since a few days, see eg https://pandas.pydata.org/speed/pandas/#stat_ops.FrameOps.time_op?python=3.8&Cython=0.29.21&p-op='mean'&p-dtype='int'&p-axis=0&commits=8dbb593d-7043f8fa

The indicated range is 8dbb593...7043f8f, but don't directly see which commit might be the cause cc @jbrockmendel

[IngErnestoAlvarez]

I think it's in this commit: 7043f8f.
Where astype function was deleted and replaced with astype_nansafe on "pandas/core/indexes/base.py::Index::astype" method.
Before, it only calls astype_nansafe if this condition was true is_integer_dtype(dtype) and not is_extension_array_dtype(dtype).
Now it calls it every time.
That can be the issue. "pandas/core/indexes/numeric.py::NumericIndex::astype" method did manage to know when to call astype_nansafe and when to call directly super().astype. Now that it's deleted, it always call astype_nansafe.

[jorisvandenbossche]

That commit has been reverted, but that didn't seem to have solved the performance regression.

@jbrockmendel can you look into this?

[jbrockmendel]

That commit has been reverted, but that didn't seem to have solved the performance regression.

does this mean we can un-revert? its a nice little cleanup

can you look into this?

cant make any promises at the moment, but ill keep a tab open for this so it doesnt fall off the radar screen entirely.

[jorisvandenbossche]

does this mean we can un-revert?

No, it was reverted because of a behaviour regression -> #38607

cant make any promises at the moment, but ill keep a tab open for this so it doesnt fall off the radar screen entirely.

If you can't look at it on the short term, I would propose to revert the change for now, until you have time (otherwise later it might become much more difficult to do that, if it turns out to be needed).
From quickly testing the different potential commits in the range, it seems #38507 is the cause of the slowdown.

[jorisvandenbossche]

Ah, I think the reason is that it is now doing a non-copy ravel, while before the copying ravel ensured we would have good layout for a performant reduction (while this is not the case when the DataFrame is created from a 2D numpy array).

[jbrockmendel]

Huh, and to think I made a non-copying ravel specifically with perf in mind. you learn something every day

[simonjayhawkins]

@jbrockmendel @jorisvandenbossche Any idea what the status is here?

[jbrockmendel]

Fixed by #41911/#41924

[simonjayhawkins]

Fixed by #41911/#41924

These are for the recent perf regression? This issue is quite old.

[jbrockmendel]

my mistake, i thought this was the one discussed on the call on wednesday. no idea off the top of my head what the status is here.