Possible performance regression in Series.xs() and DataFrame.xs() with MultiIndex.

[taozuoqiao]

Performance of Series.xs() when getting values at specified index and the first level of a large dataset seems to have regressed significantly, after pandas was upgraded to v1.0.5 from v0.25.3. So is DataFrame.xs().

code:

import numpy as np
import pandas as pd
n=1e4
data = pd.Series(np.arange(n**2), index=pd.MultiIndex.from_product([np.arange(n), -np.arange(n)]))
%timeit data.xs(100, level=0) #first level
%timeit data.xs(-100, level=1) #second level

outcome

## 0.25.3
%timeit data.xs(100, level=0)
571 µs ± 1.79 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

%timeit data.xs(-100, level=1)
193 ms ± 947 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)


## 1.0.5
%timeit data.xs(100, level=0)
683 ms ± 3.35 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)  

%timeit data.xs(-100, level=1)
197 ms ± 2.42 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) 

[simonjayhawkins]

it appears #29840 is responsible for the slowdown here

9dd1b50 is the first bad commit
commit 9dd1b50
Author: jbrockmendel jbrockmendel@gmail.com
Date: Fri Nov 29 15:10:02 2019 -0800

DEPR: remove FrozenNDarray (#29840)

cc @jbrockmendel

[lukemanley]

I think this can be closed.

## main

%timeit data.xs(100, level=0)
198 µs ± 1.79 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

%timeit data.xs(-100, level=1)
82.3 ms ± 3.22 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

[jreback]

yep
do we have covering asvs for these cases?

[lukemanley]

do we have covering asvs for these cases?

ASVs added in linked PR