⚡️ fastdsu

A fast disjoint sets implementation for Python, backed by Rust.

Accepts Arrow arrays via the C Data Interface (__arrow_c_array__) for zero-copy ingestion and zero-copy output. Works with any library that exports Arrow — PyArrow, Polars, pandas, and so on.

Why another implementation?

If you need disjoint sets / connected components over Arrow in Python, the common options are:

• A pure Python implementation
• SciPy's connected_components
• A graph library such as NetworkX or rustworkx

Pure Python requires leaving Arrow. SciPy requires leaving Arrow and constructing a sparse matrix — a large intermediate allocation entirely separate from the operation you actually want. Graph libraries bring heavy dependencies for what is fundamentally a simple algorithm.

fastdsu accepts Arrow arrays directly: no intermediate objects, no unnecessary allocations.

Requirements

Inputs must be non-nullable Arrow arrays, both of the same data type, using a fixed-width integer type (int8/int16/int32/int64/uint8/uint16/uint32/uint64).

fastdsu has no required Python dependencies — inputs and outputs use the Arrow C Data Interface protocol, so any Arrow-compatible library works at call time without being a declared dependency.

Usage

import pyarrow as pa
from fastdsu import DSU

dsu = DSU()

# Feed edge batches as they arrive: zero-copy in
dsu.union(batch_1_src, batch_1_dst)
dsu.union(batch_2_src, batch_2_dst)

# Extract components: zero-copy out, a two-column (key, label) Arrow table
components = dsu.components()
pa.record_batch(components)  # consume with PyArrow
pl.from_arrow(components)    # or Polars, or any Arrow-compatible library

Works naturally with Polars:

import polars as pl
from fastdsu import DSU

edges = pl.DataFrame({"src": [0, 1, 3, 4], "dst": [1, 2, 4, 5]}).cast(pl.UInt32)
dsu = DSU()
dsu.union(edges["src"].to_arrow(), edges["dst"].to_arrow())

components = pl.from_arrow(dsu.components())
result = edges.join(components, left_on="src", right_on="key", how="left")

One-shot convenience function

For a single batch of edges, connected_components skips constructing a DSU:

import pyarrow as pa
from fastdsu import connected_components

components = connected_components(src, dst)
pa.record_batch(components)

Contributing

pre-commit is mandatory and must be turned on.

pre-commit install --install-hooks --overwrite -t commit-msg -t pre-commit

This repo uses just as its task runner.