barca

The invisible asset orchestrator.
Rust plans it. Python runs it. You just write functions.

Barca is an asset orchestrator that adds zero perceptible overhead to your Python pipelines. A compiled Rust binary handles parsing, DAG construction, and execution planning. Python does what it's best at: running your code.

# 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
{"elapsed_seconds":0.27,"final_output":{"count":3,"total":6},"phases":1,"run_id":"...","steps_executed":2}

No config files. No YAML. No daemon. Just functions and a fast binary.

Install

Barca is designed for use with uv:

uv add barca

This gives you:

• The barca CLI binary (compiled Rust)
• Python API: barca.get(), barca.plan()
• Decorator stubs for @asset, @sensor, @task (IDE autocomplete + type checking)

For optional parquet (DataFrame) support:

uv add 'barca[parquet]'

All in one wheel, built with maturin. Requires Python >= 3.12.

From source

git clone https://github.com/barca-orc/barca.git
cd barca
uv sync
cargo build --release
maturin develop --release    # installs into current .venv

Quick start

# assets.py
from barca import asset

@asset()
def hello() -> dict:
    return {"message": "Hello from barca!"}

barca get assets.py

That's it. Barca parses your Python source with ruff's AST parser (no import, pure static analysis), builds a dependency graph, generates a phased execution plan, spawns Python workers, and persists results to a local SQLite database -- all in under 40ms for a trivial asset.

How it works

                    ┌─────────────────────────────────────┐
                    │          barca get pipeline.py       │
                    └──────────────┬──────────────────────┘
                                   │
                    ┌──────────────▼──────────────────────┐
                    │         Rust binary (barca)          │
                    │                                      │
                    │  1. Parse Python source (ruff AST)   │
                    │  2. Build DAG (petgraph)              │
                    │  3. Generate execution plan           │
                    │  4. Initialize DB (.barca/metadata.db)│
                    │  5. Spawn Python workers per phase    │
                    │  6. Collect outputs, persist to DB    │
                    └──────────────┬──────────────────────┘
                                   │
                    ┌──────────────▼──────────────────────┐
                    │      Python worker (per phase)       │
                    │                                      │
                    │  - Loads modules via importlib        │
                    │  - Executes steps in tier order       │
                    │  - LRU cache for in-process results   │
                    │  - Emits JSON lines to stdout         │
                    └─────────────────────────────────────┘

Key design decisions:

• Static analysis only -- Rust never imports your Python code. It parses source text and extracts decorator metadata from the AST.
• Phased execution -- The planner decomposes the DAG into sequential phases. Within each phase, independent streams run in parallel workers.
• No framework lock-in -- Decorators are identity functions. Your code runs standalone without barca installed.
• Single binary -- One pip install gives you everything. No JVM, no Docker, no scheduler service.

Decorators

from barca import asset, sensor, task, sink, unsafe
from barca import Always, Manual, Schedule
from barca import partitions, partitions_from, collect, asset_ref

@asset

Cached computation node. The workhorse.

@asset()
def prices() -> dict:
    return {"AAPL": 150, "MSFT": 380}

@asset(inputs={"data": prices})
def report(data: dict) -> str:
    return f"Tracked {len(data)} tickers"

@sensor

Observes external state. Returns (update_detected, output).

@sensor()
def inbox_files() -> tuple[bool, list[str]]:
    files = list(Path("inbox").glob("*.csv"))
    return bool(files), [str(f) for f in files]

A sensor always runs, and its returned value is part of the run hash of every asset that reads it: a new value re-runs those assets and everything downstream, the same value serves them from cache. Use one in front of data that changes in place (return the blob's etag); see barca docs cache.

@task

Workflow-management step — deploys, notifications, migrations, cache warming. Always re-runs (never cached). May appear anywhere in the graph and may depend on assets, sensors, or other tasks, but must not be an input to an asset or sensor (that would poison caching). Run a task with barca run <task>.

# A task consuming an upstream asset (asset → task).
@task(inputs={"report": report})
def publish(report: str) -> None:
    print(f"Publishing: {report}")

Declared environment variables

Assets that read environment variables declare them, so a changed value invalidates the cache:

@asset(env=["SOURCE_CSV", "API_TOKEN"])
def raw() -> dict:
    return load(os.environ["SOURCE_CSV"])

The values are folded into the run hash at plan time (unset is its own value) and reported per step in --agent lines and the JSON result. Names ending in _TOKEN, _SECRET, _KEY or _PASSWORD are hashed but redacted. Undeclared variables are not tracked.

@sink

Declares a sink output target (stacks on @asset; file writing coming soon).

@asset()
@sink("output/data.json", serializer="json")
def my_data() -> dict:
    return {"rows": 42}

Freshness markers

Partitions

Fan a single asset definition into N independent materializations:

@asset(partitions={"ticker": partitions(["AAPL", "MSFT", "GOOG"])})
def prices(ticker: str) -> dict:
    return {"ticker": ticker, "price": get_price(ticker)}

CLI

barca get [target] [file.py|dir/ ...]      Get asset(s) — cache-aware; `a,b` gets several
barca run <task> [file.py|dir/ ...]        Run a task (always re-runs) and its cone; `a,b` runs
                                           several
  get/run: --refresh a,b                   re-run a, b and their downstream (--no-cascade: only a, b)
  get/run: --refresh-all                   re-run every asset in the cone
barca plan [file.py ...]                   Emit execution plan as JSON (experimental)
barca list [file.py|dir/ ...] [--json|--pretty] [--limit N]  List all definitions with deps and env
barca status [target] [file.py ...] [--json|--pretty]  Cache state, last run, artifact rows/columns per
                                           node; `a,b` shows several cones
barca sql "<query>" [file.py ...] [--json] Query cached results with DuckDB (experimental)
barca history [--limit N] [--json|--pretty] Show recent run history
barca stats <target> [file.py ...] [--json|--pretty]  Timing/cache stats for an asset
barca serve [file.py ...] [--port N]       Run the HTTP API server + cron scheduler
barca docs [topic] [--all] [--json]        Built-in manual: concepts, formats, examples
barca --help                               Show help (every command ends with examples)

Files are optional everywhere: without them barca reads every .py file under the project root (the nearest barca.toml) that imports barca. See barca docs discovery.

barca docs is the manual, compiled into the binary: topics for types and output formats, caching, tasks, partitions, scheduling, runnable examples, and conventions for scripts and AI agents (barca docs agents). AI agents: load SKILL.md (also barca docs skill), a ~1500-token Agent Skill with the commands, argument order, exit codes and guardrails. Results go to stdout: human-readable in a terminal, JSON when piped or captured (--json / --pretty or BARCA_OUTPUT=json|pretty override); progress and errors go to stderr. In JSON mode an error is one JSON line on stderr ({"error", "code", "kind", "remediation"}, plus node, traceback and artifact_dir when a step failed). Exit codes: 0 ok, 1 step failed, 2 usage error, 3 barca/infra failure, 130 cancelled.

List output is bounded: list shows 100 nodes and history 10 runs unless you pass --limit N or --all, and their JSON reports truncated and total. --fields a,b keeps only those keys on each item of any JSON output (barca list pipeline.py --fields id,inputs).

Shorthand: barca pipeline.py works as barca get pipeline.py (all assets and sensors; tasks are skipped, use barca run).

The CLI surface is written down as a contract: every command, flag, environment variable, exit code, JSON output schema and --agent line, each marked stable or experimental, in crates/barca-cli/docs/contract.md (also barca docs contract, and online). Snapshot tests fail CI on any change to it that the contract does not reflect. Before 1.0 a breaking change ships in a minor release with a "Breaking" line in the release notes; from 1.0 changes are additive only.

Scheduling

Barca doubles as a plain task scheduler. Decorate a function with a cron Schedule and leave barca serve running — each job fires on its own tick. No external cron, no YAML, no daemon service to install.

# job.py
from barca import task, Schedule

@task(freshness=Schedule("*/10 * * * *"))
def refresh() -> None:
    ...  # hit an API, rebuild a file, send a report — a task always re-runs

@task(freshness=Schedule("*/15 * * * * *"))   # 6-field cron → every 15 seconds
def heartbeat() -> None:
    ...

barca serve job.py

Standard 5-field cron (minute hour day-of-month month day-of-week) and a 6-field form with a leading seconds field are both supported — the scheduler evaluates at 1-second resolution. It's timezone-aware, fires once on restart to catch up a missed tick, skips a tick if the previous run is still going, and exposes live status at GET /schedule. Inspect the schedule any time with barca list job.py.

See the Scheduling guide for the full story (timezones, catch-up, keeping it alive under systemd). A minimal example lives in examples/scheduler.

Server

barca serve starts a long-running HTTP server that exposes the orchestrator as a JSON API — for triggering runs programmatically, polling status, and (in the future) a web UI. It also runs the built-in cron scheduler above. It binds to 127.0.0.1 by default (local only, no auth).

barca serve pipeline.py --port 8274      # default port 8274
barca serve pipeline.py --watch          # dev mode: re-parse DAG on file change
barca serve pipeline.py --no-schedule    # HTTP API only, don't fire scheduled jobs
barca serve pipeline.py --timezone utc   # evaluate cron in UTC (default: local)

Runs are async: POST returns a run_id immediately, then you poll /status/{run_id}.

curl localhost:8274/health                       # {"status":"ok","version":"0.13.1"}
curl localhost:8274/assets                       # list assets + deps
curl localhost:8274/plan                          # execution plan JSON
curl -XPOST localhost:8274/run                    # → {"run_id":"…"}; poll /status/<id>
curl -XPOST localhost:8274/get/summary            # run a single target
curl localhost:8274/status/<run_id>               # poll run status + result
curl -XDELETE localhost:8274/run/<run_id>         # cancel an in-flight run

See the Server API reference for the full endpoint reference.

Python API

import barca

# Get all assets in a file (returns the last asset's value; tasks are not run)
value = barca.get("pipeline.py")
print(value)  # {"count": 3, "total": 6}

# Get a specific asset's value (cache-aware)
value = barca.get("summary", "pipeline.py")
print(value)  # {"count": 3, "total": 6}

# Inspect the execution plan
plan = barca.plan("pipeline.py")
print(plan["total_steps"])  # 2

All output formats work transparently: dicts, lists, sets, DataFrames, and arbitrary Python objects are serialized as JSON, pickle, or parquet and deserialized automatically. Every asset is fully materialized to an artifact file at step boundaries — that persistence is the cache checkpoint. Type annotations on parameters (e.g. pl.DataFrame) select the parquet reader; they do not skip materialization.

barca plan -- inspect without running

$ barca plan pipeline.py
{
  "total_steps": 2,
  "phases": [
    {
      "reason": "Initial",
      "streams": [
        {
          "stream_id": "p0-w0",
          "steps": ["pipeline.py:raw_data", "pipeline.py:summary"]
        }
      ]
    }
  ]
}

barca get -- execute and get results

Parses source, builds DAG, spawns workers, collects outputs, persists to .barca/metadata.db. With a target, only the target's subgraph runs. Without a target, every asset and sensor runs and tasks are skipped (previously tasks ran too); stderr names the skipped tasks. Run tasks with barca run.

Output is a JSON summary:

{
  "status": "success",
  "run_id": "...",
  "elapsed_seconds": 0.27,
  "steps_executed": 2,
  "phases": 1,
  "final_output": {"count": 3, "total": 6}
}

Use --refresh-all to skip cache lookups and execute everything fresh (--refresh a,b for chosen assets and what is downstream of them).

Diagnostics go to stderr:

[barca] 2/2 steps | done in 0.0s

Benchmarks

All benchmarks measured with hyperfine (3 warmup runs, 10 measured runs) on the same machine. Barca is compared against Dagster and Prefect running equivalent pipelines.

Trivial (1 asset, zero work)

Measures pure framework overhead -- how long it takes to do nothing.

Barca's total overhead (parse + plan + spawn + persist) is 38ms. Dagster needs ~0.5s. Prefect needs ~4s.

Benchmark suite

The benchmarks/ directory covers a wide range of DAG topologies and workloads — overhead/scaling, DAG shapes, real workloads (ETL, ML pipelines), partitioned runs, and dynamic dispatch/resilience — each with equivalent Dagster and Prefect implementations for apples-to-apples comparison. See benchmarks/README.md for the full topology tables and benchmarks/RESULTS.md for current results.

Run any benchmark:

cd benchmarks/trivial
./bench.sh 10    # 10 measured runs

Architecture

Cargo.toml                  Rust workspace root
crates/
  barca-core/               Engine: parser, DAG, planner, dispatch, DB, cache
  barca-cli/                Thin CLI shell (clap → barca-core)
python/barca/
  __init__.py               Decorator stubs + API exports
  api.py                    Python API (get/plan via subprocess)
  _worker.py                Execution worker (invoked by Rust binary)
  _artifacts.py             Artifact serialization (json/pickle/parquet)
  py.typed                  PEP 561 marker
pyproject.toml              Maturin build config

Tech stack

Node kinds

Development

git clone https://github.com/barca-orc/barca.git
cd barca

# Build
cargo build --release
maturin develop --release

# Test
cargo test

# Run
barca get examples/basic_app/example_project/assets.py
barca plan examples/basic_app/example_project/assets.py

Project status

Barca is in active development. The core pipeline (parse -> DAG -> plan -> execute -> persist) is working and benchmarked. See the guide for a walkthrough.

License

MIT