PERF/TYP: typing cast in __getitem__ gives quite some overhead

[jorisvandenbossche]

While doing some unrelated profiling, I noticed typing overhead in DatetimeArray.__getitem__. Isolating that here, there is a typing cast(..) call in that method, and removing that as done in the PR, I see the following impact:

arr = pd.date_range("2012", periods=3).array

In [4]: %timeit -r 20 arr[0]
4.76 µs ± 36 ns per loop (mean ± std. dev. of 20 runs, 100000 loops each)  # <-- master
3.44 µs ± 34.4 ns per loop (mean ± std. dev. of 20 runs, 100000 loops each)  # <-- PR

(I did this timing several times back and forth on a computer with no other activity at that moment, and the difference was stable)

Personally, I wouldn't have expected such a relative big impact of a typing cast. Because if that's the case, we should be very careful with using that in "hot paths".

cc @jbrockmendel @simonjayhawkins @Dr-Irv

[jbrockmendel]

wow. ive been assuming that the casts are negligible. if that wrong, then there are probably a lot of small improvements to be made by avoiding them across the codebase.

On the other hand, most of this is in the Union:

In [1]: from pandas._typing import *

In [2]: from pandas.core.arrays.datetimelike import *

In [3]: %timeit Union[DatetimeLikeArrayT, DTScalarOrNaT]
879 ns ± 19.5 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

In [4]:  %timeit cast(Union[DatetimeLikeArrayT, DTScalarOrNaT], NaT)
920 ns ± 8.15 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

So at least in this case we can get most of the benefit by just doing the Union once at the module level

[Dr-Irv]

So at least in this case we can get most of the benefit by just doing the Union once at the module level

Another (maybe easier) solution is to just put the type in double quotes:

        result = cast(
            "Union[DatetimeLikeArrayT, DTScalarOrNaT]", super().__getitem__(key)
        )

I did a grep exercise, and that is what is done elsewhere in the code. The only other place I could find (via grep) that we need to fix this is in pandas/core/arrays/categorical.py, line 517

[jbrockmendel]

@jorisvandenbossche want to change the annotation to a string and call this a win?

[Dr-Irv]

@jorisvandenbossche can you also make a similar change in pandas/core/arrays/categorical.py:astype() so we are consistent in terms of doing casts with Union

[jorisvandenbossche]

Yeah, I had actually searched for other casts in the /arrays submodule, and this categorical one is the only other cast with a composite type (or that is not predefined). Since that's not typically a hot path, I didn't change it, but of course also can't harm, so pushed that change now anyway.