barca

An embedded orchestrator for Python: a binary and a directory, like DuckDB or SQLite.
No server, no config; it runs what changed and caches the rest.

Docs · Guide · Changelog · Contributing

Status: pre-1.0. A breaking change to the CLI ships in a minor release with a "Breaking" line in the changelog.

You mark Python functions with @asset, @sensor or @task and declare their inputs. The barca binary, written in Rust, reads your source without importing it, works out what is stale, and runs it in Python worker processes. Every result is a file under .barca/. barca serve adds cron schedules, an HTTP API and a web UI when you want them.

Why barca

• Reads your code without importing it. The graph comes from parsing the source, so list, status and --dry-run do not run your modules.
• One install, no server. uv add barca brings the binary, the decorators and a Python API; the state is one directory, .barca/.
• Recomputes only what changed. An asset runs again when its code, the project code it calls, or its inputs change; everything else comes from the cache.
• Milliseconds of overhead. A one-asset pipeline runs in 64 ms from process start to exit, against 1.07 s for Dagster and about 5 s for Prefect and Airflow on the same machine (comparison); 0.28.3 then cut barca's time for it from 86 to 47 ms.

Install

uv add barca          # or: pip install barca

Python 3.12 or later. The wheel contains the barca binary, the decorators and a Python API; wheels are published for macOS on Apple Silicon and for Linux on x86-64 and arm64, glibc and musl. Barca runs your steps with the Python of the environment it is installed in, or python3 on PATH.

Optional extras: barca[parquet] (pyarrow, for DataFrames), barca[sql] (DuckDB, for barca sql), barca[datadog], and barca[s3], barca[gcs], barca[azure] or barca[remote] for a remote store.

Example

# pipeline.py
from barca import asset


@asset()
def raw_data() -> list[dict]:
    return [{"x": 1}, {"x": 2}, {"x": 3}]


@asset(inputs={"data": raw_data})
def summary(data: list[dict]) -> dict:
    return {"count": len(data), "total": sum(d["x"] for d in data)}

$ barca get pipeline.py
[barca] 2/2 steps | done in 0.1s
Run dc8e11f9d49e | all assets in 0.064s (2 steps, 1 phase)

.barca/artifacts/pipeline.py--summary/2d10c4c5ef45...json
  {
    "count": 3,
    "total": 6
  }

$ barca get pipeline.py
Run dc8f6d863e28 | all assets in 0.003s (0 steps, 1 phase)
...

get prints the path of each result, with a preview of a small json value in a terminal. The second run executes 0 steps: both results come from the cache.

Change summary (say, "total" becomes "sum") and ask what is stale. status runs nothing:

$ barca status
NAME      KIND   STATE   WHY            LAST RUN                               SHAPE           DEPS
raw_data  asset  cached  materialized   success 2026-10-21 19:52:47 UTC 0.00s  3 rows x 1 col  -
summary   asset  stale   changed: code  success 2026-10-21 19:52:47 UTC 0.00s  dict (2 keys)   raw_data

1 cached, 1 stale, 0 never run, 0 partial, 0 unknown, 0 always run

$ barca get summary
[barca] 1/2 steps | done in 0.0s
Run dc8c75a5f03a | got 'summary' in 0.034s (1 step, 1 phase)
...

Only summary runs again; raw_data comes from the cache. To recompute by hand, name the assets: barca get summary --refresh raw_data (and everything downstream of it), or --refresh-all. Do not delete files under .barca/ to force a recompute.

An asset that reads a file, a bucket or a table in its own body has the same code and inputs on every run, so it is served from cache even after the data changes. Data from outside comes in through a @sensor, which runs every time and returns something identifying the current version; the assets that take it as an input re-run only when that value changes:

import hashlib
from pathlib import Path

from barca import asset, sensor


@sensor()
def orders_version() -> tuple[bool, str]:
    return True, hashlib.md5(Path("orders.csv").read_bytes()).hexdigest()


@asset(inputs={"version": orders_version})
def orders(version: str) -> list:
    return Path("orders.csv").read_text().splitlines()[1:]

More: barca docs cache ("External data that changes in place"), which also lists what the run hash does not see.

barca serve runs the same files on cron schedules, with an HTTP API and a web UI that shows what is cached, running and queued:

From Python, barca.get("summary") returns the value, and the decorators return the function unchanged, so steps can be called and unit tested as ordinary Python.

Commands

The target comes first, then files or directories. Both are optional: with no files, barca reads every .py file under the project root that imports barca (barca docs discovery). get is for assets and uses the cache; run is for tasks, which always run, and their upstream assets come from the cache. Every barca <command> --help ends with runnable examples.

Common flags on get and run: --refresh a,b, --no-cascade, --refresh-all, --dry-run, --env <name> (a separate cache and history), --partition key=value (only those keys, across the whole run; also on barca plan).

Scripts and AI agents. Results go to stdout: a table or summary in a terminal, JSON when piped or captured (--json and --pretty override). Progress and errors go to stderr; in JSON mode an error is one JSON line with error, code, kind and remediation. Exit codes: 0 ok, 1 a step failed, 2 usage error, 3 barca or infrastructure failure, 130 cancelled. An agent should load SKILL.md, also printed by barca docs skill.

More

barca docs lists the manual's topics and barca docs <topic> prints one; start with overview. The site, barca.sh, has the same material plus a guide and patterns.

Development

git clone https://github.com/barca-orc/barca.git
cd barca
uv venv
source .venv/bin/activate
uv pip install maturin
maturin develop --uv --release --extras test   # builds the binary and installs it into .venv
cargo test

See Architecture and Contributing. benchmarks/ holds the same pipelines written for barca, Dagster, Prefect and Airflow; benchmarks/RESULTS.md records measurements with their date and machine.

License

MIT