PERF: df.unstack() is 500 times slower since pandas>=2.1

[sbonz]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this issue exists on the latest version of pandas.

• I have confirmed this issue exists on the main branch of pandas.

Reproducible Example

import pandas as pd
import numpy as np
import time
df = pd.DataFrame(np.random.random(size=(10000, 100)))
st = time.time()
df.unstack() # this operation takes 500x more in pandas>=2.1
print(f"time {time.time() -st}")

Installed Versions

INSTALLED VERSIONS ------------------ commit : bdc79c1 python : 3.11.9.final.0 python-bits : 64 OS : Windows OS-release : 10 Version : 10.0.19045 machine : AMD64 processor : Intel64 Family 6 Model 158 Stepping 9, GenuineIntel byteorder : little LC_ALL : None LANG : None LOCALE : English_United Kingdom.1252

pandas : 2.2.1
numpy : 1.26.4
pytz : 2023.3.post1
dateutil : 2.8.2
setuptools : 68.2.2
pip : 23.3.1
Cython : None
pytest : None
hypothesis : None
sphinx : None
blosc : None
feather : None
xlsxwriter : None
lxml.etree : None
html5lib : None
pymysql : None
psycopg2 : None
jinja2 : None
IPython : None
pandas_datareader : None
adbc-driver-postgresql: None
adbc-driver-sqlite : None
bs4 : None
bottleneck : 1.3.7
dataframe-api-compat : None
fastparquet : None
fsspec : None
gcsfs : None
matplotlib : None
numba : None
numexpr : 2.8.7
odfpy : None
openpyxl : None
pandas_gbq : None
pyarrow : None
pyreadstat : None
python-calamine : None
pyxlsb : None
s3fs : None
scipy : None
sqlalchemy : None
tables : None
tabulate : None
xarray : None
xlrd : None
zstandard : None
tzdata : 2023.3
qtpy : None
pyqt5 : None

Prior Performance

same code as above is 500x faster for pandas<=2.0.3.
Issue happens on Windows and Linux, with Python 3.10 and 3.12, with backend numpy and pyarrow.
The slow down seems to be in the stack_v3 function in the initial loop.

[jbrockmendel]

Cc @rhshadrach

[asishm]

on main it's about 5x faster than on 2.2.2 but still extremely slow compared to 2.0.3

on 2.0.3 -> 17ms
2.2.2 -> 5.4 s
main -> 1.08 s

[sam-baumann]

take

[sam-baumann]

Looked into this. In the sample code from the original issue, the df being used for testing is just random values, rather than the result of a stack(). The following code actually runs 2-3x faster on main than 2.0.3 on my machine.

Seems like the performance issue only comes up when the df is not in the form expected by unstack(). @sbonz did you see this on real data?

import pandas as pd
import numpy as np
import time
data = np.random.randint(0, 100,size=(100000, 1000))
df = pd.DataFrame(data=data).stack()

st = time.time()
df.unstack() 
print(f"time {time.time() -st}")

[sbonz]

@sam-baumann yes, I noticed the slowdown because some tests (with real data) in our pipeline started timing out.

[sbonz]

@sam-baumann I was wondering if you have had any chance to look at this?

[sam-baumann]

Hi @sbonz - I did look a bit further into this - I think I may have to remove myself from this issue because I don't think I'm familiar enough with this part of the codebase to be of much more help here. Sorry!

[mroeschke]

@rhshadrach would it make sense to carve out a fastpath in stack_v3 for a homogenously typed DataFrame with unique columns to just do frame._values.ravel()?

[rhshadrach]

Yea - I think adding a fastpath makes sense. I'm going to make an attempt shortly.

[jbrockmendel]

just do frame._values.ravel()?

Just a note on this: consider arr.reshape(-1) since ravel can make a copy in some cases.