plumber

A geospatial data-engineering pipeline framework. A stable function-shaped phase interface (run(gdf, **params) -> gdf) separates three concerns that are usually tangled: what an algorithm does, how phases are sequenced and scaled, and how a pipeline is specified. That separation is meant to make both prototyping→production scaling and safe agentic pipeline generation tractable — new logic drops into a fixed slot instead of modifying a monolith.

Status

Phase 5 (Refinement), v0.1.2. run, check, localmp, checkpoints and partial runs work. See issue #1 for the wayfinder map and plan.md for the design.

Install

Plumber must be installed in the same Python environment as the code your pipeline imports. Your shims import your project, so a global pipx install will not work.

Plumber is not a dependency of your project. Do not add it to your project's pyproject.toml. Install it into the environment after the fact.

Use plumber (from PyPI)

Activate the environment your project already runs in, then install plumber into it. Pick the tool you like:

# pip
pip install ab-plumber

# uv (pass the venv, so uv does not pick another Python)
uv pip install --python .venv/bin/python ab-plumber

# conda
conda install pip && pip install ab-plumber

Check it: plumber --help.

Develop plumber (editable install)

Editable means the environment points at the source, so your edits show up with no reinstall. Activate your project's environment first, then pick one.

pip can install straight from GitHub. No manual clone. pip clones the repo for you into <venv>/src/ab-plumber, and that folder is a normal git clone you can edit and commit from:

pip install -e "git+https://github.com/austinbreunig/plumber@develop#egg=ab-plumber"

uv cannot. uv pip install -e only takes a local folder, so clone first:

git clone https://github.com/austinbreunig/plumber
uv pip install --python .venv/bin/python -e ./plumber

(With pip you can also clone first and run pip install -e ./plumber.)

To work on plumber itself (tests, lint), pick one:

# uv
uv run --extra dev pytest -q
uv run --extra dev ruff check .

# pip
pip install -e ".[dev]" && pytest -q && ruff check .

# Nix (adds Python, uv and system libraries for you)
nix develop --command uv run --extra dev pytest -q
nix develop --command uv run --extra dev ruff check .

Notes

• Nix is optional. Use it only if your project has a flake.nix. On Nix, run the commands above inside nix develop, and always pass --python <venv>/bin/python to uv pip.
• uv sync removes packages that are not in your lockfile, including plumber. Keep plumber in its own venv (for example .venv-plumber), or run uv sync --inexact.

Quickstart

First install plumber. Then, in any repo, write your input function, one or more phases, and an output function. A phase is run(gdf, **params) -> gdf. Put pipeline glue (input, output, shims) in a .plumber/ folder, or in the project folder. Then add a plumber.yaml:

input:
  path: myio:load_roads
  params: {path: data/roads.parquet}
phases:
  - path: phases.buffer:run
    partitionable: true
    params: {distance: 10}
output:
  path: myio:write_parquet
  params: {path: out/roads.parquet}

Check it, then run it:

plumber check                  # PASS/WARN/FAIL per phase, also writes report.json
plumber run                    # one process (local)
plumber run --strategy localmp --worker-count 4   # worker processes, same output

Optional: checkpoint: true on a slow phase saves its output, and plumber run --from <phase> --to <phase> runs a slice.

Execution strategy

Default is local (one process). Switch to localmp (worker processes) with no phase changes:

execution:
  strategy: localmp
  partition:            # pick exactly one
    by: [zone_id]       # one piece per unique value combination
    # chunk_size: 10000 # fixed-size row slices, in order
    # worker_count: 8   # N roughly equal slices

CLI flags override the config: --strategy localmp, --by zone_id, --chunk-size 10000, --worker-count 8. Any partition flag replaces the config's whole partition: block. A phase with partitionable: false gets the joined data once. The output keeps the input's row order, and the output function runs once on the joined data.

Wrapping your own function

See docs/adapter-shim.md.

Development workflow

Follows the ab_spatial playbook: Discovery → POC → Tracer → MVP → Refinement. See ../../playbook/phases.md.

Issue tracker

GitHub Issues in austinbreunig/plumber, via the gh CLI. See docs/agents/issue-tracker.md.