PDEP-10: PyArrow as a required dependency for default string inference implementation

Abstract

This PDEP proposes that:

This will bring immediate benefits to users, as well as opening up the door for significant further benefits in the future.

Background

PyArrow is an optional dependency of pandas that provides a wide range of supplemental features to pandas:

As of pandas 2.0, one can feasibly utilize PyArrow as an alternative data representation to NumPy with advantages such as:

  1. Consistent NA support for all data types;
  2. Broader support of data types such as decimal, date and nested types;
  3. Better interoperability with other dataframe libraries based on Arrow.

Motivation

While all the functionality described in the previous paragraph is currently optional, PyArrow has significant integration into many areas of pandas. With our roadmap noting that pandas strives for better Apache Arrow interoperability [^1] and many projects [^2], within or beyond the Python ecosystem, adopting or interacting with the Arrow format, making PyArrow a required dependency provides an additional signal of confidence in the Arrow ecosystem (as well as improving interoperability with it).

Immediate User Benefit 1: pyarrow strings

Currently, when users pass string data into pandas constructors without specifying a data type, the resulting data type is object, which has significantly much worse memory usage and performance as compared to pyarrow strings. With pyarrow string support available since 1.2.0, requiring pyarrow for 3.0 will allow pandas to default the inferred type to the more efficient pyarrow string type.

In [1]: import pandas as pd

In [2]: pd.Series(["a"]).dtype
# Current behavior
Out[2]: dtype('O')

# Future behavior in 3.0
Out[2]: string[pyarrow]

Dask developers investigated performance and memory of pyarrow strings here, and found them to be a significant improvement over the current object dtype.

Little demo:

import string
import random

import pandas as pd


def random_string() -> str:
    return "".join(random.choices(string.printable, k=random.randint(10, 100)))


ser_object = pd.Series([random_string() for _ in range(1_000_000)])
ser_string = ser_object.astype("string[pyarrow]")\

PyArrow backed strings are significantly faster than NumPy object strings:

str.len

In[1]: %timeit ser_object.str.len()
118 ms ± 260 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In[2]: %timeit ser_string.str.len()
24.2 ms ± 187 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

str.startswith

In[3]: %timeit ser_object.str.startswith("a")
136 ms ± 300 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In[4]: %timeit ser_string.str.startswith("a")
11 ms ± 19.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Immediate User Benefit 2: Nested Datatypes

Currently, if you try storing dicts in a pandas Series, you will again get the horrendous object dtype:

In [6]: pd.Series([{'a': 1, 'b': 2}, {'a': 2, 'b': 99}])
Out[6]:
0     {'a': 1, 'b': 2}
1    {'a': 2, 'b': 99}
dtype: object

If pyarrow were required, this could have been auto-inferred to be pyarrow.struct, which again would come with memory and performance improvements.

Immediate User Benefit 3: Interoperability

Other Arrow-backed dataframe libraries are growing in popularity. Having the same memory representation would improve interoperability with them, as operations such as:

import pandas as pd
import polars as pl

df = pd.DataFrame(
  {
    'a': ['one', 'two'],
    'b': [{'name': 'Billy', 'age': 3}, {'name': 'Bob', 'age': 4}],
  }
)
pl.from_pandas(df)

could be zero-copy. Users making use of multiple dataframe libraries would more easily be able to switch between them.

Future User Benefits:

Requiring PyArrow would simplify the related development within pandas and potentially improve NumPy functionality that would be better suited by PyArrow including:

Developer benefits

First, this would simplify development of pyarrow-backed datatypes, as it would avoid optional dependency checks.

Second, it could potentially remove redundant functionality: - fastparquet engine in read_parquet; - potentially simplifying the read_csv logic (needs more investigation); - factorization; - datetime/timezone ops.

Drawbacks

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.

Additionally, if a user is installing pandas in an environment where wheels are not available through a pip install or conda install, the user will need to also build Arrow C++ and related dependencies when installing from source. These environments include

Lastly, pandas development and releases will need to be mindful of PyArrow's development and release cadance. For example when supporting a newly released Python version, pandas will also need to be mindful of PyArrow's wheel support for that Python version before releasing a new pandas version.

F.A.Q.

Q: Why can't pandas just use numpy string and numpy void datatypes instead of pyarrow string and pyarrow struct?

A: NumPy strings aren't yet available, whereas pyarrow strings are. NumPy void datatype would be different to pyarrow struct, not bringing the same interoperabitlity benefit with other arrow-based dataframe libraries.

Q: Are all pyarrow dtypes ready? Isn't it too soon to make them the default?

A: They will likely be ready by 3.0 - however, we're not making them the default (yet). For example, pd.Series([1, 2, 3]) will continue to be auto-inferred to be np.int64. We will only change the default for dtypes which currently have no numpy-backed equivalent and which are stored as object dtype, such as strings and nested datatypes.

PDEP-10 History

[^1] https://pandas.pydata.org/docs/development/roadmap.html#apache-arrow-interoperability [^2] https://arrow.apache.org/powered_by/