s3fs-access-grants 0.1.1

Transparent S3 Access Grants for fsspec/s3fs — per-grant scoped credentials, no Access Grants logic in application code.

s3fs-access-grants

Transparent S3 Access Grants for fsspec / s3fs.

Register the implementation for s3:// once and every fsspec consumer — pandas, polars, universal-pathlib, fsspec.open, ... — routes each call through a credential scoped to the matching grant. No Access Grants logic leaks into application code.

How it works

ScopedS3FileSystem subclasses s3fs.S3FileSystem. On construction it enumerates the caller's grants (ListCallerAccessGrants) and builds a longest-prefix routing table. Each S3 operation is matched to a grant scope and served by a per-scope aiobotocore client whose credentials come from GetDataAccess and auto-refresh on expiry. Calls that match no grant fall through to the default IAM client (fail-closed), and cross-scope copies are rejected.

If the caller has no grants (or no permission to list them), register() installs nothing and plain s3fs handles s3:// with zero added overhead. The same holds when there is no AWS access at all (no credentials, no region, or an unreachable endpoint): register() fails open and leaves the default s3fs in place, so it is safe to call unconditionally in tests and CI.

Install

pip install s3fs-access-grants
# or
uv add s3fs-access-grants

Usage

Registration is explicit — importing the package does no AWS I/O.

import s3fs_access_grants

# Install the s3:// handler for the calling identity's grants.
s3fs_access_grants.register()

import polars as pl
df = pl.read_parquet("s3://bucket/teamA/data.parquet")  # scoped automatically

Pass the grants-instance account/region explicitly (e.g. in a notebook, or when the grants instance lives in a different account than your role):

import s3fs_access_grants

s3fs_access_grants.register(account_id="123456789012", region="eu-west-1")

register() is the only entry point and returns nothing — it's a one-time setup call. It resolves the account/region, then binds them into the handler it registers, so a later argless fsspec.filesystem("s3") (what pandas / polars use under the hood) builds the filesystem with the right values. After calling it, just use fsspec / pandas / polars as normal.

Configuration

register() resolves the grants-instance account and region in this order:

Setting	Resolution order
Account ID	explicit arg → S3FS_ACCESS_GRANTS_ACCOUNT_ID → STS caller
Region	explicit arg → S3FS_ACCESS_GRANTS_REGION → session default

The grants instance may live in a different account than the caller's role, so the explicit arg / env is authoritative; the STS caller account is only a same-account fallback.

Advanced: manual construction

register() covers the common case. To build a scoped filesystem by hand — e.g. pointing at a different grants account without touching the global s3:// registration — construct ScopedS3FileSystem directly:

from s3fs_access_grants import ScopedS3FileSystem

fs = ScopedS3FileSystem(grants_account_id="123456789012", grants_region="eu-west-1")

Compatibility

This subclass depends on s3fs internals (_call_s3 and _iterdir resolving their client via get_s3()). Validated against s3fs 2026.6.0 / aiobotocore 2.25.1. Re-check on any s3fs major bump.

Development

bun install          # installs JS tooling + runs uv sync + lefthook install
bun run check        # lint + typecheck (ruff, ty, editorconfig)
bun run test         # pytest with coverage
bun run build        # uv build (wheel + sdist)

Releases are managed with changesets: add one with bun run changeset:add, and merging the generated "version packages" PR bumps the version, syncs it into pyproject.toml, and publishes to PyPI.

License

MIT