PERF: high memory consumption for unstack

[hkad98]

Pandas version checks

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

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

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

Reproducible Example

I use memory_profiler. Run mprof run file_name to measure memory consumption on each line of code.

import pandas as pd
import random
import string
from memory_profiler import profile

def random_string():
    return ''.join(random.choices(string.ascii_letters, k=7))

@profile
def main():
    records_count = 63531
    df = pd.DataFrame(
        {
            "A": random.choices([random_string() for _ in range(24)], k=records_count),
            "B": random.choices([random_string() for _ in range(14580)], k=records_count),
            "C": random.choices([random_string() for _ in range(9)], k=records_count),
            "D": random.choices([random_string() for _ in range(2311)], k=records_count),
            "E": random.choices([random_string() for _ in range(2)], k=records_count),
            "F": random.choices([random_string() for _ in range(280)], k=records_count),
            "M": random.sample(range(0, records_count), records_count)
        }
    )

    grouped_df = df.groupby(["A", "B", "C", "D", "E", "F"], dropna=False)[["M"]].sum(min_count=1, numeric_only=False)
    grouped_df.unstack("F")

if __name__ == "__main__":
    main()

Memory usage for unstack:

    27    264.1 MiB    171.8 MiB           1       grouped_df.unstack("F")

I tried to improve memory consumption with the following changes:

• use category instead of string
• use pivot_table instead of groupby + unstack
• use reset_index + pivot instead of unstack

None of the things above worked.

I tried to turn on CoW, but it made it worse.

  28    390.6 MiB    306.1 MiB           1       grouped_df.unstack("F")

I am aware that CoW does not support unstack yet (https://github.com/pandas-dev/pandas/issues/49473), but I would not expect that turning it on will make unstack worse.

Installed Versions

INSTALLED VERSIONS ------------------ commit : 0f437949513225922d851e9581723d82120684a6 python : 3.10.8.final.0 python-bits : 64 OS : Darwin OS-release : 22.5.0 Version : Darwin Kernel Version 22.5.0: Thu Jun 8 22:22:20 PDT 2023; root:xnu-8796.121.3~7/RELEASE_ARM64_T6000 machine : arm64 processor : arm byteorder : little LC_ALL : None LANG : en_US.UTF-8 LOCALE : None.UTF-8 pandas : 2.0.3 numpy : 1.25.2 pytz : 2023.3 dateutil : 2.8.2 setuptools : 63.2.0 pip : 22.2.2 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 bs4 : None bottleneck : None brotli : None fastparquet : None fsspec : None gcsfs : None matplotlib : None numba : None numexpr : None odfpy : None openpyxl : None pandas_gbq : None pyarrow : None pyreadstat : None pyxlsb : None s3fs : None scipy : None snappy : None sqlalchemy : None tables : None tabulate : None xarray : None xlrd : None zstandard : None tzdata : 2023.3 qtpy : None pyqt5 : None

Prior Performance

No response

[phofl]

Hm I can't reproduce this. Also on an ARM Mac

 29    141.8 MiB     34.1 MiB           1       grouped_df.unstack("F")

Can you provide a simple reproducer? Less columns and without groupby

[hkad98]

@phofl I will try to find a simpler reproducer. May I ask how grouped_df.unstack("F") in your reproducer was executed on line 29? Did you add any lines? What version of pandas did you use? What is wrong with groupby?

[phofl]

I added one line to activate/deactivate CoW. Otherwise this was copied as is. Examples should always be as simple as possible, e.g. no unnecessary operations. If your problem occurs in unstack, then groupby is unnecessary

[hkad98]

@phofl what about the following code?

import pandas as pd
from memory_profiler import profile

@profile
def main():
    df = pd.read_parquet("reproducer.parquet")
    df.unstack("F")

if __name__ == "__main__":
    main()

I put reproducer.parquet in archiv.zip.

Line #    Mem usage    Increment  Occurrences   Line Contents
=============================================================
    11     77.0 MiB     77.0 MiB           1   @profile
    12                                         def main():
    13    115.3 MiB     38.3 MiB           1       df = pd.read_parquet("reproducer.parquet")
    14    290.0 MiB    174.8 MiB           1       df.unstack("F")

Still getting huge memory consumption for unstack.

[phofl]

Nope, still can't reproduce

Line #    Mem usage    Increment  Occurrences   Line Contents
=============================================================
     4    146.7 MiB    146.7 MiB           1   @profile
     5                                         def main():
     6    178.8 MiB     32.1 MiB           1       df = pd.read_parquet("reproducer.parquet")
     7    215.9 MiB     37.0 MiB           1       df.unstack("F")

[hkad98]

That is strange. What version of Pandas do you use?

[phofl]

tried on main and 2.0.3

[hkad98]

Hi @phofl, I tried to reproduce the issue independently, and I think I succeeded. See the following runs in my public repo.

Ubuntu: https://github.com/hkad98/pandas-reproducer/actions/runs/5784679774/job/15675843719 MacOS: https://github.com/hkad98/pandas-reproducer/actions/runs/5784759695/job/15676084594

Note that both runners use x86

Unstack increment for:

• Ubuntu 5.2 MiB
• MacOS 29.2 MiB (~6 times more)

Unfortunately, GitHub does not provide runners with ARM architecture. I tried locally running the same script with Ubuntu + ARM, and the results were the same as Ubuntu with x86. I think that issue is in the combination of macOS13 with ARM.

[mroeschke]

Seems like the original issue wasn't directly reproducible so closing.