Apache Arrow PostgreSQL connector

The goal of pgeon is to provide fast bulk data download from a PostgreSQL database into Apache Arrow tables. pgeon provides a C++ library and simple python bindings. Almost all PostgreSQL native types are supported (see below).

If you're looking to upload data, you might want to have a look at Arrow ADBC.

This project is similar to pg2arrow and is heavily inspired by it. The main differences are the use of COPY instead of FETCH and that our implementation uses the Arrow C++ API.

from pgeon import copy_query
db = "postgresql://postgres@localhost:5432/postgres"
query = "SELECT * FROM some_table"
tbl = copy_query(db, query)

The actual query performed is COPY ({query}) TO STDOUT (FORMAT binary), see this page for more information.

We provide pre-built binary wheels in the Release section. No dependencies are required. Conda users, please read below.

Building pgeon requires libpq to be available on your system.

git clone https://github.com/0x0L/pgeon.git
cd pgeon
pip install .

The pre-built binary wheels are built using the old C++ ABI as used by the pyarrow package available from pypi. Unfortunately the conda-forge pyarrow package uses the new C++ ABI. If you are using pyarrow from conda at runtime, you can install pgeon using

CONDA_BUILD=1 pip install .

This requires cmake and ninja. In addition you'll need to install libpq and the Arrow C++ libraries (e.g. arrow-cpp in conda)

mkdir build
cd build
cmake -GNinja ..
ninja

Elapsed time distributions of a query fetching 7 columns (1 timestamp, 2 ints, 4 reals) and around 4.5 million rows. The result is returned as a pandas.DataFrame in all cases.

• Queries using ROW (e.g. SELECT ROW('a', 1)) do not work (anonymous structs)

• SQL arrays are mapped to pyarrow.list_(...). Due to the PostgreSQL wire format, only 1D arrays are fully supported. Higher dimensional arrays will be flattened.

• BitString types output format is not really helpful

• tsvector types with letter weights are not supported

• PostgreSQL range and domain types are not supported

• Dynamic record types are not supported