saving on memory can cost both memory and performance

[behzadnouri]

xref: #8676 (comment)

memory cost:

with int16 labels:

$ python -m memory_profiler mem-profile.py 
Filename: mem-profile.py

Line #    Mem usage    Increment   Line Contents
================================================
     4   80.156 MiB    0.000 MiB   @profile
     5                             def ix(obj):
     6   87.809 MiB    7.652 MiB       obj.ix[999]

with int64 labels:

$ python -m memory_profiler mem-profile.py 
Filename: mem-profile.py

Line #    Mem usage    Increment   Line Contents
================================================
     4   79.387 MiB    0.000 MiB   @profile
     5                             def ix(obj):
     6   79.387 MiB    0.000 MiB       obj.ix[999]

performance cost:

-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
frame_xs_mi_ix                               |   8.5303 |   0.6206 |  13.7452 |
series_xs_mi_ix                              |   8.0659 |   0.5600 |  14.4023 |
-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------

Ratio < 1.0 means the target commit is faster then the baseline.
Seed used: 1234

Target [c11e75c] : PERF: set multiindex labels with coerced dtype (GH8456)
Base   [6bbb39e] : Merge pull request #8675 from pydata/setitem

the mem-profile.py used for memory profiling:

@profile
def ix(obj):
    obj.ix[999]

if __name__ == '__main__':
    import numpy as np
    from pandas import MultiIndex, Series
    mi = MultiIndex.from_tuples([(x,y) for x in range(1000) for y in range(1000)])
    ts =  Series(np.random.randn(1000000), index=mi)
    ix(ts)

[jreback]

pls include other performance for actual 'real' indexing. getting a single scalar is useful, but other benchmarks are as well.

[mroeschke]

This looks related to ix which is deprecated.