DataFusion in Python

This is a Python library that binds to Apache Arrow in-memory query engine DataFusion.

Like pyspark, it allows you to build a plan through SQL or a DataFrame API against in-memory data, parquet or CSV files, run it in a multi-threaded environment, and obtain the result back in Python.

It also allows you to use UDFs and UDAFs for complex operations.

The major advantage of this library over other execution engines is that this library achieves zero-copy between Python and its execution engine: there is no cost in using UDFs, UDAFs, and collecting the results to Python apart from having to lock the GIL when running those operations.

Its query engine, DataFusion, is written in Rust, which makes strong assumptions about thread safety and lack of memory leaks.

Technically, zero-copy is achieved via the c data interface.

How to use it

Simple usage:

import datafusion
import pyarrow

# an alias
f = datafusion.functions

# create a context
ctx = datafusion.ExecutionContext()

# create a RecordBatch and a new DataFrame from it
batch = pyarrow.RecordBatch.from_arrays(
    [pyarrow.array([1, 2, 3]), pyarrow.array([4, 5, 6])],
    names=["a", "b"],
)
df = ctx.create_dataframe([[batch]])

# create a new statement
df = df.select(
    f.col("a") + f.col("b"),
    f.col("a") - f.col("b"),
)

# execute and collect the first (and only) batch
result = df.collect()[0]

assert result.column(0) == pyarrow.array([5, 7, 9])
assert result.column(1) == pyarrow.array([-3, -3, -3])

UDFs

def is_null(array: pyarrow.Array) -> pyarrow.Array:
    return array.is_null()

udf = f.udf(is_null, [pyarrow.int64()], pyarrow.bool_())

df = df.select(udf(f.col("a")))

UDAF

import pyarrow
import pyarrow.compute


class Accumulator:
    """
    Interface of a user-defined accumulation.
    """
    def __init__(self):
        self._sum = pyarrow.scalar(0.0)

    def to_scalars(self) -> [pyarrow.Scalar]:
        return [self._sum]

    def update(self, values: pyarrow.Array) -> None:
        # not nice since pyarrow scalars can't be summed yet. This breaks on `None`
        self._sum = pyarrow.scalar(self._sum.as_py() + pyarrow.compute.sum(values).as_py())

    def merge(self, states: pyarrow.Array) -> None:
        # not nice since pyarrow scalars can't be summed yet. This breaks on `None`
        self._sum = pyarrow.scalar(self._sum.as_py() + pyarrow.compute.sum(states).as_py())

    def evaluate(self) -> pyarrow.Scalar:
        return self._sum


df = ...

udaf = f.udaf(Accumulator, pyarrow.float64(), pyarrow.float64(), [pyarrow.float64()])

df = df.aggregate(
    [],
    [udaf(f.col("a"))]
)

How to install (from pip)

pip install datafusion

How to develop

This assumes that you have rust and cargo installed. We use the workflow recommended by pyo3 and maturin.

Bootstrap:

# fetch this repo
git clone git@github.com:apache/arrow-datafusion.git

cd arrow-datafusion/python

# prepare development environment (used to build wheel / install in development)
python3 -m venv venv
# activate the venv
source venv/bin/activate
pip install -r requirements.txt

Whenever rust code changes (your changes or via git pull):

# make sure you activate the venv using "source venv/bin/activate" first
maturin develop
python -m pytest

How to update dependencies

To change test dependencies, change the requirements.in and run

# install pip-tools (this can be done only once), also consider running in venv
pip install pip-tools

# change requirements.in and then run
pip-compile --generate-hashes

To update dependencies, run

pip-compile update

More details here