PERF: Why does specifying the index column in pandas significantly increases the read time of a csv?

[nitinmnsn]

• 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 master branch of pandas.

Reproducible Example

I am seeing a significantly increased read time for a CSV by pandas when I specify the index_col. I do not understand the reason behind it. Can you help me understand why that is happening and if that is actually the expected behaviour? Below is the code I am using:

import pandas as pd
#save the CSV to be used
pd.DataFrame({'id':np.arange(100000000),'b':np.random.choice(['a','b','c','d'],size=(100000000,),p=[0.25,0.25,0.25,0.25])}).to_csv('df_sp.csv',index=None)
dfpd = pd.read_csv('df_sp.csv')
#read time - 10.3 seconds
dfpd = pd.read_csv('df_sp.csv',index_col='id')
#read time - 1 minute 38.6 seconds

In fact, I am seeing significant improvement if I read the dataset without specifying index_col and then set the index by dfpd = dfpd.set_index('id'). This takes just 1.6 more seconds. Why does pandas not default to always reading the dataframe with index_col as a column and then setting it as the index internally with set_index(index_col) when index_col is specified?

Installed Versions

INSTALLED VERSIONS

commit : 73c6825
python : 3.9.7.final.0
python-bits : 64
OS : Linux
OS-release : 5.4.0-89-generic
Version : #100-Ubuntu SMP Fri Sep 24 14:50:10 UTC 2021
machine : x86_64
processor : x86_64
byteorder : little
LC_ALL : None
LANG : en_IN
LOCALE : en_IN.ISO8859-1

pandas : 1.3.3
numpy : 1.21.2
pytz : 2021.1
dateutil : 2.8.2
pip : 21.2.4
setuptools : 58.0.4
Cython : None
pytest : None
hypothesis : None
sphinx : None
blosc : None
feather : None
xlsxwriter : None
lxml.etree : None
html5lib : None
pymysql : None
psycopg2 : None
jinja2 : 3.0.1
IPython : 7.25.0
pandas_datareader: None
bs4 : None
bottleneck : None
fsspec : 2021.09.0
fastparquet : None
gcsfs : None
matplotlib : 3.4.3
numexpr : None
odfpy : None
openpyxl : 3.0.7
pandas_gbq : None
pyarrow : 5.0.0
pyxlsb : None
s3fs : None
scipy : 1.7.1
sqlalchemy : 1.4.23
tables : None
tabulate : 0.8.9
xarray : 0.19.0
xlrd : 2.0.1
xlwt : None
numba : 0.53.1

Prior Performance

No response

[JonathanHourany]

I'm not a dev on this project, but I did notice that there's a recent regression fix to the speed of read_csv (see: #44192) that just got merged in. It might fix this problem too.

[jreback]

yep very likely

i think we just need a specific asv benchmark for this to close

[nitinmnsn]

I think this is an issue of not just pandas but how pandas collaborate with OS. I am not seeing this performance issue on mac os and on google colab that I set up to test this. This performance issue is there on ubuntu 20.04 and 18.04. Checked on both pandas 1.3.4 and 1.3.3

https://stackoverflow.com/questions/69692122/why-does-specifying-the-index-column-in-pandas-significantly-increases-the-read

[asishm]

#44192 doesn't fix this issue. this is on a smaller csv with just 2 columns

tested on WSL (Ubuntu 20.04)

In [11]: %timeit df = pd.read_csv("~/df_sp.csv")
789 ms ± 14.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [12]: %timeit df = pd.read_csv("~/df_sp.csv", index_col='id')
4.13 s ± 24.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [13]: pd.__version__
Out[13]: '1.4.0.dev0+1006.g7bf75b0e28'

[simonjayhawkins]

#44192 doesn't fix this issue.

It appears that fix and the regression reported in #44106 are independant of this performance issue and that this one is not a regression or improved (changed) by the change in #44192. Will remove the milestone.

%timeit dfpd = pd.read_csv('df_sp.csv')
# 10.5 s ± 265 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- master
# 11.8 s ± 597 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- 1.3.x (1.3.4+10.gc803b7eaac)
# 11.3 s ± 457 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- 1.3.4
# 12.4 s ± 351 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- 1.2.5
# 11.9 s ± 386 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- 1.1.5

%timeit dfpd = pd.read_csv('df_sp.csv',index_col='id')
# 57.7 s ± 3.06 s per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- master
# 1min 2s ± 3.55 s per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- 1.3.x (1.3.4+10.gc803b7eaac)
# 1min ± 1.87 s per loop (mean ± std. dev. of 7 runs, 1 loop each) <-- 1.3.4
# 57.8 s ± 762 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- 1.2.5
# 1min ± 1.8 s per loop (mean ± std. dev. of 7 runs, 1 loop each)  <-- 1.1.5

Further investigation and contributions welcome.