[ENH] investigating & improving runtime performance of data type checks (input, output)

[danbartl]

Investigation to see why checking for datatypes can take very long.
A description of all checks can be found here.
https://docs.google.com/spreadsheets/d/1cDx_yg2HhFcPfNR1Trc1Vr1FDuSOTkXwid5GTKIS2aE/edit?usp=sharing

After some investigation, most computing time was consumed when having _index_equally_spaced for periodindex.

Some investigations revealed that groupby can optimize / reduce time needed slightly, but periodindex should be avoided.

Findings are included here:
https://github.com/danbartl/sktime/blob/check_test/perftest.ipynb

Conclusion:

• most checks are very fast, only _index_equally_spaced can take very long
• Will create a pull request to change method to groupby, increasing performance
• periodindex should be avoided for large datasets
• unfortunately, only periodindex supports freq argument for multiindex, further supporting the idea that freq should be outside argument
• polars library performs 5 times as fast as fastest pandas method

[fkiraly]

Amazing and impressive!

Based on this, I wonder whether we can integrate a slightly more polished version of the performance testing into sktime itself, as a developer utility.

[fkiraly]

Re some of these checks:

some changes can be made by accident by transformers, but this can be certified in the test suite (as opposed to in an output check).

Regarding whether transformers can change:

• in-principle, transformers could return an empty data frame, e.g., when subsetting or selecting
• transformers can also change the number of time points, number of samples in a panel, number of variables