FEEDBACK: PyArrow as a required dependency and PyArrow backed strings

[phofl]

This is an issue to collect feedback on the decision to make PyArrow a required dependency and to infer strings as PyArrow backed strings by default.

The background for this decision can be found here: https://pandas.pydata.org/pdeps/0010-required-pyarrow-dependency.html

If you would like to filter this warning without installing pyarrow at this time, please view this comment: #54466 (comment)

[mynewestgitaccount]

Something that hasn't received enough attention/discussion, at least in my mind, is this piece of the Drawbacks section of the PDEP (bolding added by me):

Including PyArrow would naturally increase the installation size of pandas. For example, installing pandas and PyArrow using pip from wheels, numpy and pandas requires about 70MB, and including PyArrow requires an additional 120MB. An increase of installation size would have negative implication using pandas in space-constrained development or deployment environments such as AWS Lambda.

I honestly don't understand how mandating a 170% increase in the effective size of a pandas installation (70MB to 190MB, from the numbers in the quoted text) can be considered okay.

For that kind of increase, I would expect/want the tradeoff to be major improvements across the board. Instead, this change comes with limited benefit but massive bloat for anyone who doesn't need the features PyArrow enables, e.g. for those who don't have issues with the current functionality of pandas.

[rebecca-palmer]

Debian/Ubuntu have system packages for pandas but not pyarrow, which would no longer be possible. (System packages are not allowed to depend on non-system packages.)

I don't know whether creating a system package of pyarrow is possible with reasonable effort, or whether this would make the system pandas packages impossible to update (and eventually require their removal when old pandas was no longer compatible with current Python/numpy).

[mroeschke]

For that kind of increase, I would expect/want the tradeoff to be major improvements across the board.

Yeah unfortunately this is where the subjective tradeoff comes into effect. pytz and dateutil as required dependencies have a similar issue for users who do not need timezone or date parsing support respectively. The hope with pyarrow is that the tradeoff improves the current functionality for common "object" types in pandas such as text, binary, decimal, and nested data.

Debian/Ubuntu have system packages for pandas but not pyarrow, which would no longer be possible.

AFAIK most pydata projects don't actually publish/manage Linux system packages for their respective libraries. Do you know how these are packaged today?

[mynewestgitaccount]

pytz and dateutil as required dependencies have a similar issue for users who do not need timezone or date parsing support respectively.

The pytz and dateutil wheels are only ~500kb. Drawing a comparison between them and PyArrow seems like a stretch, to put it lightly.

[rebecca-palmer]

Do you know how these are packaged today?

By whoever offers to do it, currently me for pandas. Of the pydata projects, Debian currently has pydata-sphinx-theme, sparse, patsy, xarray and numexpr.

An old discussion thread (anyone can post there, but be warned that doing so will expose your non-spam-protected email address) suggests that there is existing work on a pyarrow Debian package, but I don't yet know whether it ever got far enough to work.

[rebecca-palmer]

I do intend to investigate this further at some point - I haven't done so yet because Debian updated numexpr to 2.8.5, breaking pandas (#54449 / #54546), and fixing that is currently more urgent.

[jjerphan]

Hi,

Thanks for welcoming feedback from the community.

While I respect you decision, I am afraid that making pyarrow a required dependency will come with costly consequences for users and downstream libraries' developers and maintainers for two reasons:

• installing pyarrow after pandas in a fresh conda environment increases its size from approximately 100MiB to approximately 500 MiB.

Packages size

libgoogle-cloud-2.12.0-h840a212_1 :                 46106632 bytes,
python-3.11.4-hab00c5b_0_cpython :                  30679695 bytes,
libarrow-12.0.1-h10ac928_8_cpu :                    27696900 bytes,
ucx-1.14.1-h4a2ce2d_3 :                             15692979 bytes,
pandas-2.0.3-py311h320fe9a_1 :                      14711359 bytes,
numpy-1.25.2-py311h64a7726_0 :                      8139293 bytes,
libgrpc-1.56.2-h3905398_1 :                         6331805 bytes,
libopenblas-0.3.23-pthreads_h80387f5_0 :            5406072 bytes,
aws-sdk-cpp-1.10.57-h85b1a90_19 :                   4055495 bytes,
pyarrow-12.0.1-py311h39c9aba_8_cpu :                3989550 bytes,
libstdcxx-ng-13.1.0-hfd8a6a1_0 :                    3847887 bytes,
rdma-core-28.9-h59595ed_1 :                         3735644 bytes,
libthrift-0.18.1-h8fd135c_2 :                       3584078 bytes,
tk-8.6.12-h27826a3_0 :                              3456292 bytes,
openssl-3.1.2-hd590300_0 :                          2646546 bytes,
libprotobuf-4.23.3-hd1fb520_0 :                     2506133 bytes,
libgfortran5-13.1.0-h15d22d2_0 :                    1437388 bytes,
pip-23.2.1-pyhd8ed1ab_0 :                           1386212 bytes,
krb5-1.21.2-h659d440_0 :                            1371181 bytes,
libabseil-20230125.3-cxx17_h59595ed_0 :             1240376 bytes,
orc-1.9.0-h385abfd_1 :                              1020883 bytes,
ncurses-6.4-hcb278e6_0 :                            880967 bytes,
pygments-2.16.1-pyhd8ed1ab_0 :                      853439 bytes,
jedi-0.19.0-pyhd8ed1ab_0 :                          844518 bytes,
libsqlite-3.42.0-h2797004_0 :                       828910 bytes,
libgcc-ng-13.1.0-he5830b7_0 :                       776294 bytes,
ld_impl_linux-64-2.40-h41732ed_0 :                  704696 bytes,
libnghttp2-1.52.0-h61bc06f_0 :                      622366 bytes,
ipython-8.14.0-pyh41d4057_0 :                       583448 bytes,
bzip2-1.0.8-h7f98852_4 :                            495686 bytes,
setuptools-68.1.2-pyhd8ed1ab_0 :                    462324 bytes,
zstd-1.5.2-hfc55251_7 :                             431126 bytes,
libevent-2.1.12-hf998b51_1 :                        427426 bytes,
libgomp-13.1.0-he5830b7_0 :                         419184 bytes,
xz-5.2.6-h166bdaf_0 :                               418368 bytes,
libcurl-8.2.1-hca28451_0 :                          372511 bytes,
s2n-1.3.48-h06160fa_0 :                             369441 bytes,
aws-crt-cpp-0.21.0-hb942446_5 :                     320415 bytes,
readline-8.2-h8228510_1 :                           281456 bytes,
libssh2-1.11.0-h0841786_0 :                         271133 bytes,
prompt-toolkit-3.0.39-pyha770c72_0 :                269068 bytes,
libbrotlienc-1.0.9-h166bdaf_9 :                     265202 bytes,
python-dateutil-2.8.2-pyhd8ed1ab_0 :                245987 bytes,
re2-2023.03.02-h8c504da_0 :                         201211 bytes,
aws-c-common-0.9.0-hd590300_0 :                     197608 bytes,
aws-c-http-0.7.11-h00aa349_4 :                      194366 bytes,
pytz-2023.3-pyhd8ed1ab_0 :                          186506 bytes,
aws-c-mqtt-0.9.3-hb447be9_1 :                       162493 bytes,
aws-c-io-0.13.32-h4a1a131_0 :                       154523 bytes,
ca-certificates-2023.7.22-hbcca054_0 :              149515 bytes,
lz4-c-1.9.4-hcb278e6_0 :                            143402 bytes,
python-tzdata-2023.3-pyhd8ed1ab_0 :                 143131 bytes,
libedit-3.1.20191231-he28a2e2_2 :                   123878 bytes,
keyutils-1.6.1-h166bdaf_0 :                         117831 bytes,
tzdata-2023c-h71feb2d_0 :                           117580 bytes,
gflags-2.2.2-he1b5a44_1004 :                        116549 bytes,
glog-0.6.0-h6f12383_0 :                             114321 bytes,
c-ares-1.19.1-hd590300_0 :                          113362 bytes,
libev-4.33-h516909a_1 :                             106190 bytes,
aws-c-auth-0.7.3-h28f7589_1 :                       101677 bytes,
libutf8proc-2.8.0-h166bdaf_0 :                      101070 bytes,
traitlets-5.9.0-pyhd8ed1ab_0 :                      98443 bytes,
aws-c-s3-0.3.14-hf3aad02_1 :                        86553 bytes,
libexpat-2.5.0-hcb278e6_1 :                         77980 bytes,
libbrotlicommon-1.0.9-h166bdaf_9 :                  71065 bytes,
parso-0.8.3-pyhd8ed1ab_0 :                          71048 bytes,
libzlib-1.2.13-hd590300_5 :                         61588 bytes,
libffi-3.4.2-h7f98852_5 :                           58292 bytes,
wheel-0.41.1-pyhd8ed1ab_0 :                         57374 bytes,
aws-c-event-stream-0.3.1-h2e3709c_4 :               54050 bytes,
aws-c-sdkutils-0.1.12-h4d4d85c_1 :                  53123 bytes,
aws-c-cal-0.6.1-hc309b26_1 :                        50923 bytes,
aws-checksums-0.1.17-h4d4d85c_1 :                   50001 bytes,
pexpect-4.8.0-pyh1a96a4e_2 :                        48780 bytes,
libnuma-2.0.16-h0b41bf4_1 :                         41107 bytes,
snappy-1.1.10-h9fff704_0 :                          38865 bytes,
typing_extensions-4.7.1-pyha770c72_0 :              36321 bytes,
libuuid-2.38.1-h0b41bf4_0 :                         33601 bytes,
libbrotlidec-1.0.9-h166bdaf_9 :                     32567 bytes,
libnsl-2.0.0-h7f98852_0 :                           31236 bytes,
wcwidth-0.2.6-pyhd8ed1ab_0 :                        29133 bytes,
asttokens-2.2.1-pyhd8ed1ab_0 :                      27831 bytes,
stack_data-0.6.2-pyhd8ed1ab_0 :                     26205 bytes,
executing-1.2.0-pyhd8ed1ab_0 :                      25013 bytes,
_openmp_mutex-4.5-2_gnu :                           23621 bytes,
libgfortran-ng-13.1.0-h69a702a_0 :                  23182 bytes,
libcrc32c-1.1.2-h9c3ff4c_0 :                        20440 bytes,
aws-c-compression-0.2.17-h4d4d85c_2 :               19105 bytes,
ptyprocess-0.7.0-pyhd3deb0d_0 :                     16546 bytes,
pure_eval-0.2.2-pyhd8ed1ab_0 :                      14551 bytes,
libblas-3.9.0-17_linux64_openblas :                 14473 bytes,
liblapack-3.9.0-17_linux64_openblas :               14408 bytes,
libcblas-3.9.0-17_linux64_openblas :                14401 bytes,
six-1.16.0-pyh6c4a22f_0 :                           14259 bytes,
backcall-0.2.0-pyh9f0ad1d_0 :                       13705 bytes,
matplotlib-inline-0.1.6-pyhd8ed1ab_0 :              12273 bytes,
decorator-5.1.1-pyhd8ed1ab_0 :                      12072 bytes,
backports.functools_lru_cache-1.6.5-pyhd8ed1ab_0 :  11519 bytes,
pickleshare-0.7.5-py_1003 :                         9332 bytes,
prompt_toolkit-3.0.39-hd8ed1ab_0 :                  6731 bytes,
backports-1.0-pyhd8ed1ab_3 :                        5950 bytes,
python_abi-3.11-3_cp311 :                           5682 bytes,
_libgcc_mutex-0.1-conda_forge :                     2562 bytes,

• pyarrow also depends on libarrow which itself depends on several notable C and C++ libraries. This constraints the installation of other packages whose dependencies might be incompatible with libarrow's, making pandas potentially unusable in some context.

Have you considered those two observations as drawbacks before taking the decision?

[lithomas1]

Hi,

Thanks for welcoming feedback from the community.

While I respect you decision, I am afraid that making pyarrow a required dependency will come with costly consequences for users and downstream libraries' developers and maintainers for two reasons:

• installing pyarrow after pandas in a fresh conda environment increases its size from approximately 100MiB to approximately 500 MiB.

Packages size

• pyarrow also depends on libarrow which itself depends on several notable C and C++ libraries. This constraints the installation of other packages whose dependencies might be incompatible with libarrow's, making pandas potentially unusable in some context.

Have you considered those two observations as drawbacks before taking the decision?

This is discussed a bit in https://github.com/pandas-dev/pandas/pull/52711/files#diff-3fc3ce7b7d119c90be473d5d03d08d221571c67b4f3a9473c2363342328535b2R179-R193
(for pip only I guess).

While currently the build size for pyarrow is pretty large, it doesn't "have" to be that big. I think by pandas 3.0
(when pyarrow will actually become required), at least some components will be spun out/made optional/something like that (I heard that the arrow people were talking about this).

(cc @jorisvandenbossche for more info on this)

I'm not an Arrow dev myself, but if is something that just needs someone to look at, I'm happy to put some time in help give Arrow a nudge in the right direction.

Finally, for clarity purposes, is the reason for concern also AWS lambda/pyodide/Alpine, or something else?

(IMO, outside of stuff like lambda funcs, pyarrow isn't too egregious in terms of package size compared to pytorch/tensorflow but it's definetely something that can be improved)

[jjerphan]

If libarrow is slimmed down by having non-essential Arrow features be extracted into other libraries which could be optional dependencies, I think most people's concerns would be addressed.

Edit: See conda-forge/arrow-cpp-feedstock#1035

[DerThorsten]

Hi,

Thanks for welcoming feedback from the community.
For wasm builds of python / python-packages (ie pyodide / emscripten-forge) package size really matters since these packages have to be downloaded from within the browser. Once a package is too big, usability suffers drastically.

With pyarrow as a required dependency, pandas is less usable from python in the browser.

[surfaceowl]

Debian/Ubuntu have system packages for pandas but not pyarrow, which would no longer be possible. (System packages are not allowed to depend on non-system packages.)

I don't know whether creating a system package of pyarrow is possible with reasonable effort, or whether this would make the system pandas packages impossible to update (and eventually require their removal when old pandas was no longer compatible with current Python/numpy).

There is another way - use virtual environments in user space instead of system python. The Python Software Foundation recommends users create virtual environments; and Debian/Ubuntu want users to leave the system python untouched to avoid breaking system python.

Perhaps Pandas could add some warnings or error messages on install to steer people to virtualenv. This approach might avoid or at least defer work of adding pyarrow to APT as well as the risks of users breaking system python. Also which I'm building projects I might want a much later version of pandas/pyarrow than would ever ship on Debian given the release strategy/timing delay.

On the other hand, arrow backend has significant advantages and with the rise of other important packages in the data space that also use pyarrow (polars, dask, modin), perhaps there is sufficient reason to add pyarrow to APT sources.

A good summary that might be worth checking out is Externally managed environments. The original PEP 668 is found here.

[stonebig]

I think it's the rigth path for performance in WASM.

[mlkui]

This is a good idea!
But I think there are also two important features should also be implemented except strings:

1. Zero-copy for multi-index dataframe. Currently, multi-index dataframe can not be convert from arrow table with zero copy(zero_copy_only=True), which is a BIGGER problem for big dataframe. You can reset_index() the dataframe, convert it to arrow table, and convert arrow table back to dataframe with zero copy, but after all, you must use call set_index() to the dataframe to get multi-index back, then copy happens.
2. Zero-copy for pandas.concat. Arrow table concat can be zero-copy, but when concat two zero-copy dataframe(convert from arrow table), copy happens even pandas COW is turned on. Also, currently, trying to concat two arrow table and then convert the table to dataframe with zero_copy_only=True is also not allowed as the chunknum>1.

[phofl]

@mlkui

Regarding concat: This should already be zero copy:

df = pd.DataFrame({"a": [1, 2, 3]}, dtype="int64[pyarrow]")
df2 = pd.DataFrame({"a": [1, 2, 3]}, dtype="int64[pyarrow]")

x = pd.concat([df, df2])

This creates a new dataframe that has 2 pyarrow chunks.

Can you open a separate issue if this is not what you are looking for?

[mlkui]

@phofl
Thanks for your reply. But your example may be too simple. Please view the following codes(pandas 2.0.3 and pyarrow 12.0/ pandas 2.1.0 and pyarrow 13.0):

        with pa.memory_map("d:\\1.arrow", 'r') as source1, pa.memory_map("d:\\2.arrow", 'r') as source2, pa.memory_map("d:\\3.arrow", 'r') as source3, pa.memory_map("d:\\4.arrow", 'r') as source4:

            c1 = pa.ipc.RecordBatchFileReader(source1).read_all().column("p")
            c2 = pa.ipc.RecordBatchFileReader(source2).read_all().column("v")
            c3 = pa.ipc.RecordBatchFileReader(source1).read_all().column("p")
            c4 = pa.ipc.RecordBatchFileReader(source2).read_all().column("v")
            print("RSS: {}MB".format(pa.total_allocated_bytes() >> 20))

            s1 = c1.to_pandas(zero_copy_only=True)
            s2 = c2.to_pandas(zero_copy_only=True)
            s3 = c3.to_pandas(zero_copy_only=True)
            s4 = c4.to_pandas(zero_copy_only=True)
            print("RSS: {}MB".format(pa.total_allocated_bytes() >> 20))

            dfs = {"p": s1, "v": s2}
            df1 = pd.concat(dfs, axis=1, copy=False)                            #zero-copy
            print("RSS: {}MB".format(pa.total_allocated_bytes() >> 20))

            dfs2 = {"p": s3, "v": s4}
            df2 = pd.concat(dfs2, axis=1, copy=False)                           #zero-copy
            print("RSS: {}MB".format(pa.total_allocated_bytes() >> 20))

            # NOT zero-copy
            result_df = pd.concat([df1, df2], axis=0, copy=False)

        with pa.memory_map("z1.arrow", 'r') as source1, pa.memory_map("z2.arrow", 'r') as source2:

            table1 = pa.ipc.RecordBatchFileReader(source1).read_all()
            table2 = pa.ipc.RecordBatchFileReader(source2).read_all()
            combined_table = pa.concat_tables([table1, table2])
            print("RSS: {}MB".format(pa.total_allocated_bytes() >> 20))        #Zero-copy

            df1 = table1.to_pandas(zero_copy_only=True)
            df2 = table2.to_pandas(zero_copy_only=True)
            print("RSS: {}MB".format(pa.total_allocated_bytes() >> 20))       #Zero-copy

            #Use pandas to concat two zero-copy dataframes
            #But copy happens
            result_df = pd.concat([df1, df2], axis=0, copy=False)

            #Try to convert the arrow table to pandas directly
            #This will raise exception for chunk number is 2
            df3 = combined_table.to_pandas(zero_copy_only=True)

            # Combining chunks to one will cause copy
            combined_table = combined_table.combine_chunks()

[0x26res]

Beside the build size, there is a portability issue with pyarrow.

pyarrow does not provide wheels for as many environment as numpy.

For environments where pyarrow does not provide wheels, pyarrow has to be installed from source which is not simple.

[flying-sheep]

If this happens, would dtype='string' and dtype='string[pyarrow]' be merged into one implementation?

We’re currently thinking about coercing strings in our library, but hesitating because of the unclear future here.

[EwoutH]

pyarrow does not provide wheels for as many environment as numpy.

The fact that they still don’t have Python 3.12 wheels up is worrisome.

[h-vetinari]

The fact that they still don’t have Python 3.12 wheels up is worrisome.

Arrow is a beast to build, and even harder to fit into a wheel properly (so you get less features, and things like using the slimmed-down libarrow will be harder to pull off).

Conda-forge builds for py312 have been available for a month already though, and are ready in principle to ship pyarrow with a minimal libarrow. That still needs some usability improvements, but it's getting there.

[musicinmybrain]

Without weighing in on whether this is a good idea or a bad one, Fedora Linux already has a libarrow package that provides python3-pyarrow, so I think this shouldn’t be a real problem for us from a packaging perspective.

I’m not saying that Pandas is easy to keep packaged, up to date, and coordinated with its dependencies and reverse dependencies! Just that a hard dependency on PyArrow wouldn’t necessarily make the situation worse for us.

[ZupoLlask]

@h-vetinari Almost there? :-)

[raulcd]

@h-vetinari Almost there? :-)

There is still a lot of work to be done on the wheels side but for conda after the work we did to divide the CPP library, I created this PR which is currently under discussion in order to provide both a pyarrow-base that only depends on libarrow and libparquet and pyarrow which would pull all the Arrow CPP dependencies. Both have been built with support for everything so depending on pyarrow-base and libarrow-dataset would allow the use of pyarrow.dataset, etc.

[chris-vecchio]

Thanks for requesting feedback. I'm not well versed on the technicalities, but I strongly prefer to not require pyarrow as a dependency. It's better imo to let users choose to use PyArrow if they desire. I prefer to use the default NumPy object type or pandas' StringDType without the added complexity of PyArrow.

[Sai123-prathyu]

Thankyou its working

[dwgillies]

I too was hoping to use pandas in an embedded AWS lambda function. If the size explodes, this will be a huge overhead. I am currently using about 0.004% of the pandas library. From the looks of this discussion, my usage will not change nor will I ever need pyarrow but I will now be using 0.0015% of the pandas library, and paying dearly for it, probably by abandoning this bloated software.

I have found and verified that the deprecation warning can be suppressed with this : #54466 (comment)

Does anyone have a procedure for installing pyarrow in cygwin?

Note: straightforward installation does not work.

cygwin$ python3.9 -m pip install pyarrow
...
      -- Generator: Unix Makefiles
      -- Build output directory: /tmp/pip-install-obx0lyoa/pyarrow_5ca48afb32b3451db0badc556c1c74fc/build/temp.cygwin-3.5.0-x86_64-cpython-39/release
      -- Found Python3: /usr/bin/python3.9.exe (found version "3.9.16") found components: Interpreter Development.Module NumPy
      -- Found Python3Alt: /usr/bin/python3.9.exe
      CMake Error at CMakeLists.txt:268 (find_package):
        By not providing "FindArrow.cmake" in CMAKE_MODULE_PATH this project has
        asked CMake to find a package configuration file provided by "Arrow", but
        CMake did not find one.

        Could not find a package configuration file provided by "Arrow" with any of
        the following names:

          ArrowConfig.cmake
          arrow-config.cmake

        Add the installation prefix of "Arrow" to CMAKE_PREFIX_PATH or set
        "Arrow_DIR" to a directory containing one of the above files.  If "Arrow"
        provides a separate development package or SDK, be sure it has been
        installed.


      -- Configuring incomplete, errors occurred!
      See also "/tmp/pip-install-obx0lyoa/pyarrow_5ca48afb32b3451db0badc556c1c74fc/build/temp.cygwin-3.5.0-x86_64-cpython-39/CMakeFiles/CMakeOutput.log".
      error: command '/usr/bin/cmake' failed with exit code 1
      [end of output]

  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for pyarrow
Failed to build pyarrow
ERROR: Could not build wheels for pyarrow, which is required to install pyproject.toml-based projects

[maksym-petrenko]

What is the minimum version of PyArrow that will work with pandas?

[jsolbrig]

@dwgillies I don't use Cygwin so I can only help a little with your installation issue. Pyarrow doesn't provide a wheel for your OS and architecture. So, pip is trying to build a wheel from source. In order to build form source, pyarrow requires that you have libarrow installed. If you install libarrow, then try to pip install again, it might work.

[sahilfatima]

Importing model

[Soft-Buddy]

I don't consider this a good decision, a huge increment in the installation size will be there :(

[miraculixx]

@dwgillies #54466 (comment)

Does anyone have a procedure for installing pyarrow in cygwin? Note: straightforward installation does not work.

That's a great question - many companies rely on Python + Pandas running in cygwin, mingw (through git-bash) and Msys in their Windows work PCs. It is often the best way to have a useful Python dev env in a corporate environment.

Will Pandas+PyArrow be supported in these environments? If not there is a high risk of lots of outdated installations bc these environments are rather sticky once deployed, and there is no easy way to upgrade to Linux or WSL.

[Soft-Buddy]

@dwgillies #54466 (comment)

Does anyone have a procedure for installing pyarrow in cygwin? Note: straightforward installation does not work.

That's a great question - many companies rely on Python + Pandas running in cygwin, mingw (through git-bash) and Msys in their Windows work PCs. It is often the best way to have a useful Python dev env in a corporate environment.

Will Pandas+PyArrow be supported in these environments? If not there is a high risk of lots of outdated installations bc these environments are rather sticky once deployed, and there is no easy way to upgrade to Linux or WSL.

Our issues kinda match buddy. I use pandas in my android app which ships a cross compiled copy of python and of pandas compiled using crossenv. PyArrow's installation doesn't work there either... And triggers some weird errors

[jkmackie]

My general concern with the mandatory PyArrow dependency is chasing competing standards and dependency issues like bugs.

Kindly recall PDEP 10 lists three key benefits of pyarrow: (1) better pyarrow string memory/speed; (2) nested datatypes; and (3) interoperability.

PDEP Point 1 - Pandas 2.2.0 Performance
To make this less abstract, below are Pandas performance stats based on the 1brc challenge of aggregating 1 billion rows of city temperatures as fast as possible.

---

1brc INPUT - 1 billion rows
2 columns: city and temperature

OUTPUT - Temp mean/min/max by city

Metrics

---

Memory
These metrics use the default DataFrame format: city is 'object' and temperature is 'float64'.

Turns out the city column 'object' format hogs 🐷 90% of the 'deep' memory usage ⵜ. This is indeed an issue! The last 10% of memory is temperatures. Downcasting to 'float32' halves memory for the temperature column.

_{ⵜ Memory Footnote:
There's a mismatch between dataframe 'deep' memory usage (69GB) and the PC RAM increase I saw in Task Manager (about 23-24 GB) during pd.read_parquet(). My system memory is 64GB. Hard to believe 2GB memory compression accounts for the discrepancy.}

Speed
Reading from parquet is 2.5 times as fast as reading from csv and takes one-fifth the space (snappy compression). Mean/min/max aggregation time was reasonable at under one minute.

PDEP Point 2 - Nesting
The PDEP 10 nested datatype example saves [{'a': 1, 'b': 2}, {'a': 2, 'b': 99}] to a Series rather than a DataFrame. The pyarrow benefit is saving an unknown nested structure as speed/memory efficient strings.

The existing alternative is use pd.json_normalize() or pd.DataFrame() to load the example into a DataFrame with a column for eack key. Foreknowledge of the format is required. Then downcast numeric columns with pd.to_numeric(df[mycol], downcast=<'integer', 'signed', 'unsigned', or 'float'>).

PDEP Point 3 - Interoperability
What about potential PyArrow C++ binding issues? Is this straightforward to debug and fix?

---

TAKEAWAYS
Pandas stock performance is good. 😎 With foreknowledge of the nested format, data can be flattened into a DataFrame (with a column for each key). Numbers are downcasted one column at a time.

The standout issue to me is the dtype: object. Why not build a solution in Pandas or NumPy?

[hagenw]

BTW, reading in a CSV file or parquet file is still faster by a factor of 5 for me when I do the reading with pyarrow and then convert to a pandas.DataFrame (but yes, using pyarrow as datatype for string is then faster then using object), compared to reading directly with pandas using pyarrow as engine.

[jkmackie]

BTW, reading in a CSV file or parquet file is still faster by a factor of 5 for me when I do the reading with pyarrow and then convert to a pandas.DataFrame (but yes, using pyarrow as datatype for string is then faster then using object), compared to reading directly with pandas using pyarrow as engine.

@hagenw Would you kindly explain the below result? Looks like parquet uses a lot more peak RAM.

Windows users: In general, what is the large discrepancy between DataFrame memory shown by df.info(memory_usage = 'deep') versus Windows Task Manager (below is a Task Manager memory metric pic)? What is the right 'real-world' memory metric?

[hagenw]

I measured peak memory consumption with memray, but I'm not completely sure if I did it correctly.
I have some updated results in the dev branch (https://github.com/audeering/audb/tree/dev/benchmarks), there we see the following

So it seems to be more equal. The code I used to measure memory consumption is available at https://github.com/audeering/audb/blob/44de33f0fea1f4d003882d674dc696a8f0cfe95d/benchmarks/benchmark-dependencies-save-and-load.py. That uses memray and writes the results to binary files that you need to inspect afterwards to extract the result.

[ebuchlin]

There have been / are some efforts to reduce the size of pandas (#30741), these efforts should not be wasted by a dependency which could perhaps remain optional (although I have no idea whether this is feasible). +120MB multiplied by the number of installs/environments/images/CI runs is not so small. It takes more time to download and install, more network usage, more storage... It's neither green, nor inclusive for situations/people/institutes/countries where resources are not as easily available as where these decisions are taken.