PERF: selection inside a row using a list of labels 30 slower than selection by the same individual labels

[xstreck1]

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

The code below accesses two columns either together or separately:

import pandas as pd
import numpy as np
import timeit

# Create large DataFrame
n_rows = 100_000
df = pd.DataFrame({
    'a': np.random.random(n_rows),
    'b': np.random.random(n_rows)
})

# Method 1: Access columns inside function
def sum_inside(row):
    return row[0] + row[1]

# Method 2: Pass values directly
def sum_outside(a, b):
    return a + b

# Time Method 1
t1 = timeit.timeit(lambda: df.apply(lambda row: sum_inside(row[['a','b']].values), axis=1), number=1)

# Time Method 2  
t2 = timeit.timeit(lambda: df.apply(lambda row: sum_outside(row['a'], row['b']), axis=1), number=1)

print(f"Method 1 (access inside): {t1:.4f} seconds")
print(f"Method 2 (pass values): {t2:.4f} seconds")

Output:

Method 1 (access inside): 15.6217 seconds
Method 2 (pass values): 0.5135 seconds

Using iloc does not suffer the same issue. Replace:

t1 = timeit.timeit(lambda: df.apply(lambda row: sum_inside(row.iloc[0:2].values), axis=1), number=1)
t2 = timeit.timeit(lambda: df.apply(lambda row: sum_outside(row.iloc[0], row.iloc[1]), axis=1), number=1)

Output:

Method 1 (access inside): 1.5293 seconds
Method 2 (pass values): 0.8811 seconds

Installed Versions

INSTALLED VERSIONS

commit : 0691c5c
python : 3.11.7
python-bits : 64
OS : Windows
OS-release : 10
Version : 10.0.26100
machine : AMD64
processor : AMD64 Family 23 Model 49 Stepping 0, AuthenticAMD
byteorder : little
LC_ALL : None
LANG : en_US.UTF-8
LOCALE : English_Germany.1252

pandas : 2.2.3
numpy : 1.26.3
pytz : 2024.1
dateutil : 2.8.2
pip : 24.0
Cython : None
sphinx : 8.1.3
IPython : 8.22.1
adbc-driver-postgresql: None
adbc-driver-sqlite : None
bs4 : 4.12.3
blosc : None
bottleneck : None
dataframe-api-compat : None
fastparquet : None
fsspec : 2024.2.0
html5lib : None
hypothesis : None
gcsfs : None
jinja2 : 3.1.2
lxml.etree : 4.9.4
matplotlib : 3.8.3
numba : 0.59.0
numexpr : None
odfpy : None
openpyxl : None
pandas_gbq : None
psycopg2 : None
pymysql : None
pyarrow : 14.0.2
pyreadstat : None
pytest : 8.2.0
python-calamine : None
pyxlsb : None
s3fs : None
scipy : 1.12.0
sqlalchemy : None
tables : None
tabulate : None
xarray : None
xlrd : None
xlsxwriter : None
zstandard : None
tzdata : 2024.1
qtpy : None
pyqt5 : None

Prior Performance

No response

[rhshadrach]

Thanks for the report. This can be reproduced on just a Series.

ser = pd.Series({"a": 1, "b": 2})

%timeit row[["a", "b"]]
# 103 μs ± 273 ns per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

%timeit row["a"], row["b"]
# 1.76 μs ± 12.4 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

I haven't checked, but likely looking up a single label has a fastpath, and requires less validation. Certainly any perf improvement is always welcome, but I'm not sure there is an issue here.

I presume the example in the OP is just a simplified version of the operation you're trying to perform. In any case, avoiding apply at all costs is the best way to get good performance out of pandas.

[jorisvandenbossche]

I haven't checked, but likely looking up a single label has a fastpath, and requires less validation.

Doing a quick profile of both, the single label lookup is indeed a simpler index looking (though the index get_loc), while the list-like case is a lot more complex: it first converts the lists of labels to an Index (involving some validation), then determines an indexer array, and then reindexes (takes) the original series with that array.
All this involves some overhead that is significant for such a tiny series, but for larger series becomes less relevant.

[rhshadrach]

Thanks @jorisvandenbossche, closing.