BENCH: Index.take benchmark is measuring wrong thing

[jorisvandenbossche]

The take benchmarks indexing.IndexingMethods.time_take_intindex is benchmarking take on a boolean list, which take interprets as a [0, 1, 0, 1, ...] values, which is a bit silly to benchmark.
We should add some actual benchmark on integers instead.

[sangramga]

@jorisvandenbossche I would like to make this issue to be my first contribution. I have looked at the source code and ready to take it up.
Although I have unable to see as to why and how a Boolean list self.indexer is interpreted as integer [0,1,0,1,...] values?
Is it possible to resolve this issue by creating a new indexer list like self.int_indexer for benchmarking? What do you have in mind?

[jorisvandenbossche]

Is it possible to resolve this issue by creating a new indexer list like self.int_indexer for benchmarking? What do you have in mind?

Yes, although I think you can just replace the existing self.indexer with a new list of ints instead of bools (we don't need to keep the old bench of using booleans).

Although I have unable to see as to why and how a Boolean list self.indexer is interpreted as integer [0,1,0,1,...] values?

If you try to run the code (or similar), you will see:

In [84]: s = pd.Series(range(5))

In [85]: s.take([True, False, True, False, True])
Out[85]: 
1    1
0    0
1    1
0    0
1    1
dtype: int64

In [86]: s.loc[[True, False, True, False, True]]
Out[86]: 
0    0
2    2
4    4
dtype: int64

so take interprets the True, False as 0, 1 and not as a mask (like is done in loc)

[sangramga]

@jorisvandenbossche I made the following changes in indexing.IndexingMethods.time_take_intindex

Replaced the boolean list indexer with randomized integer list
self.indexer = list(np.random.randint(0, 100000, size=100000))
Does this resolve the issue?

Also where should I add Tests? Should I add/modify any tests?

[hardikpnsp]

It seems like the issue has been open for quite a while. I looked into the take benchmark, It is still using boolean values as indexer. Does not seem like anyone is working on the issue, I will quickly fix that.

[hardikpnsp]

take