Barca vs Prefect

Last measured: 2026-10-21, with barca 0.28.0 and Prefect 3.9.0, both from PyPI with uv, Python 3.13, on Linux (AMD Ryzen AI MAX+ 395). Setup and commands are in benchmarks/RESULTS.md. It replaces the first comparison, made on 2026-06-10 with barca 0.2.0 and Prefect 3.7.4 on macOS.

Install footprint

Metric	barca 0.28.0	Prefect 3.9.0
Packages installed	1	103
Venv size (du -sm)	31 MB	147 MB
Lockfile entries	2	106
Import time (python -c "import X")	42 ms	502 ms

Each column is uv pip install X into an empty venv. Import times are medians of 20 runs; python -c pass took 13 ms. The barca wheel contains the Rust binary, the web UI that barca serve shows and the Python decorator stubs, and declares no dependencies.

Steps to first output

Barca

uv add barca
# write pipeline.py
barca get pipeline.py

Prefect

pipeline.py

from prefect import flow, task

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

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

@flow
def pipeline():
    return summary(raw_data())

if __name__ == "__main__":
    print(pipeline())

python pipeline.py

Prefect needs no separate command: the flow runs with python pipeline.py. Each run starts a temporary local server first; in the logged run below, 1.6 seconds passed between “Starting temporary server” and the start of the flow.

Running the same pipeline

The pipeline is two steps: one returns three rows, the other counts and sums them.

Metric	barca 0.28.0	Prefect 3.9.0
Command	barca get pipeline.py	python pipeline.py
Time, nothing cached	73 ms (67–76)	4.59 s (4.37–4.91)
Output	1 JSON line on stdout, 2 progress lines on stderr	the result on stdout, 7 log lines on stderr
Error output	1 JSON line on stdout, 4 lines on stderr	111 lines on stderr

Times are medians of 10 runs, the two tools taking turns run by run; the range is the fastest and slowest run. barca’s .barca/ was deleted before each run.

Barca 0.28.0

[barca] importing pipeline.py
[barca] 2/2 steps | done in 0.0s
{"elapsed_seconds":0.0609,"final_output":{"_barca_artifact":{"format":"json","path":".barca/artifacts/pipeline.py--summary/2d10c4….json","size_bytes":24}},"phases":1,"run_id":"…","status":"success","steps":[…],"steps_executed":2,"warnings":[]}

barca get pipeline.py -o value prints the value itself. The JSON shape is in the CLI contract.

Prefect 3.9.0

03:53:34.483 | INFO    | prefect - Starting temporary server on http://127.0.0.1:8235
See https://docs.prefect.io/v3/concepts/server#how-to-guides for more information on running a dedicated Prefect server.
03:53:36.095 | INFO    | Flow run 'fantastic-clam' - Beginning flow run 'fantastic-clam' for flow 'pipeline'
03:53:36.109 | INFO    | Task run 'raw_data-c2b' - Finished in state Completed()
03:53:36.111 | INFO    | Task run 'summary-72d' - Finished in state Completed()
03:53:37.106 | INFO    | Flow run 'fantastic-clam' - Finished in state Completed()
{'count': 3, 'total': 6}
03:53:37.121 | INFO    | prefect - Stopping temporary server on http://127.0.0.1:8235

Error output

The same step raising ValueError("something went wrong").

Barca 0.28.0

Exit code 1. On stdout, one JSON object with "status": "failed", "error": "ValueError: something went wrong" and "failed_node". On stderr, a progress line, one JSON line with the error, a remediation and the traceback frames from the user’s file, and a [barca] run failed: ... line. See the CLI contract.

Prefect 3.9.0: 111 lines

Exit code 1. The error was printed three times:

1. a task-level ERROR with a traceback (task_engine.py, call_task_fn, …);
2. a flow-level ERROR with a traceback;
3. the unhandled exception with a traceback, starting at broken.py, line 15.

The frame from the user’s file (broken.py, line 6) comes after the Prefect frames in each.

Caching

Barca caches every asset by its run hash, a hash of the function’s code and its inputs, with no option to set. A second run of the same pipeline on barca 0.28.0 took 13 ms (median of 10) and ran no step. barca docs cache says what the hash covers.

Prefect has cache_policy=INPUTS. With both tasks set to it and two python pipeline.py runs on Prefect 3.9.0 under the temporary server, the second run reported summary as Cached(type=COMPLETED) and ran raw_data, which takes no inputs, again. On 2026-06-10, with Prefect 3.7.4, neither task was cached on the second run.

Other things observed on 2026-06-10

These are notes from using Prefect 3.7.4 on 2026-06-10. They have not been checked against a later Prefect release, except where a note dated 2026-10-21 says so.

• process_customer.map(customer_ids) fans one task out over a list in one call. barca’s equivalents are parallel_map inside a task and partitions on an asset (Parallel tasks).
• Flow runs get generated names such as “tricky-kiwi”. barca run ids are hexadecimal strings.
• prefect flow-run ls listed past runs with their status across separate invocations. barca’s equivalent is barca history.
• main.serve(name="my-deployment", cron="0 8 * * *") turned a flow into a scheduled long-running process. barca’s equivalent is Schedule("0 8 * * *") on the function and barca serve (Scheduling).
• Every python file.py run and every prefect CLI command started a temporary HTTP server, which took 2 to 3 seconds. A flow computing 1 + 2 took 3.5 seconds. Note, 2026-10-21: with Prefect 3.9.0 on the Linux machine, the two-task pipeline above took 4.6 seconds and the server start 1.6 seconds of it.
• prefect --help listed 31 subcommands. The names recorded were: api, artifact, automation, block, cloud, concurrency-limit, config, dashboard, deploy, deployment, dev, events, experimental, flow, flow-run, global-concurrency-limit, init, plugins, profile, sdk, server, shell, task, task-run, transfer, variable, version, work-pool, work-queue, worker.
• prefect init --help answered “Unknown option: –help. Did you mean –field?”.
• Every run printed the server start and stop lines, the flow run line and one line per task. No option to suppress them was found during the test.
• @flow and @task wrap the function, and the original is reached through .fn. barca’s decorators return the function unchanged, so a decorated function can be called as plain Python; barca then does no caching and records nothing.

Why barca has no python file.py mode

In barca the Rust binary does the parsing, planning, caching and persistence, so a pipeline is run through it: barca get on the command line, or barca.get(...) from Python, which starts the same binary. Running python pipeline.py executes the functions as ordinary Python with none of that.