brevis

Pass context between the steps of a Brevis workflow, and read the clock the engine already worked out for the run. That is all it does.

pip install brevis

from brevis import context, metrics, run

bucket = context.get("extract.bucket")
since, until = run.window()          # the window this run covers

context.set(rows=48213, watermark="2026-09-07T03:00:00Z")
metrics.set("rows_loaded", 48213)    # reaches the engine's /metrics

No dependencies, ever. It reads one environment variable and writes one file. There is no API call, no token and no port: the engine hands a step its input when it starts it and reads the output back when it ends.

This is not a port of the Go SDK. No drivers, no pagination, no ingestion ids — Python already has better tools for those, and this exists so a team can use pandas, Polars or dbt and still get the one thing an orchestrator is for.

Reading

A key is always qualified by the step that published it:

context.get("extract.bucket")            # the value
context.get("extract.nope", default=0)   # absent is not an error
context.of("extract")                    # everything that step published

context.get("bucket") with no step is refused, naming the steps that published it. Steps are isolated — extract and transform can both publish bucket without either losing it — and a bare key throws that away the moment two of them do. Searching and picking one is precedence by accident: it works until it silently does not.

Reading a step this one does not depends_on is an error pointing at the fix.

Writing

context.set(name="Daniel")
context.set(label="Nome")
# both survive: {"name": "Daniel", "label": "Nome"}

Calls merge. Setting the same key twice keeps the last write; setting a different key leaves the first alone. There is no step argument — the only namespace a process can write is its own, which is what makes the isolation structural rather than a rule.

Nothing reaches the disk until the process exits, so a step that dies hard publishes nothing. That is correct: a crashed step's output described work that did not finish.

The clock this run should read

Every run also carries auto params: values the engine works out on its own. Nobody declares them, every run has them, and they answer the question a pipeline would otherwise answer with datetime.now().

from datetime import timedelta
from brevis import run

since, until = run.window() or (run.now() - timedelta(days=1), run.now())
df = fetch(since, until)
df.to_parquet(f"/data/{run.auto().date}.parquet")

run.now() is the slot on a scheduled run. That matters more than it sounds: a fetcher that reads datetime.now() and subtracts two hours is right on a run that starts on time and forty minutes wrong on one the queue delayed — and those forty minutes belong to no run at all, because the next slot reads its own now() too. Nothing fails, and the gap is found weeks later.

run.now()                   # the clock: the slot, or when the run started
run.window()                # (start, end), end excluded — or None
run.auto().date             # "2026-09-08": the partition, the folder, the WHERE
run.auto().delay_seconds    # how late this attempt was
run.auto().previous_error   # the run before this one did not succeed
run.context()               # id, attempt, trigger, params, and auto

window() comes from the cron, not from the history: a backfill of a slot from March produces the window March had. Asking for [start, end) never overlaps and never gaps, however late the run is and however often it retries.

It returns None rather than a pair of zero times when the workflow has no schedule — a query from the epoch selects everything, and that failure should not be silent. The or above is the whole fallback.

Run the script by hand and run.now() is the wall clock and window() is None, so there is no branch to write for local development.

One instance of a mapped step

Under for_each:, the engine runs one process per element and tells each one which it got:

partition = run.map_value()
if partition is None:
    raise SystemExit("this step is meant to run under `for_each:`")

load(f"/data/{partition}.csv")

A JSON string arrives without its quotes — for_each over ["2026-01"] hands the step 2026-01, not "2026-01". Anything else — an object, a number, a list — arrives as its JSON, so json.loads(run.map_value()).

run.map_index() is the position, and both are None on a step that is not mapped. None and not -1 or "": a sentinel is a number somebody eventually does arithmetic on, and an empty string is indistinguishable from an element that is the empty string.

A mapped step publishes no context downstream — four instances cannot share one key — so a step that needs to hand something on writes a file, or a step after it counts what landed.

Metrics your pipeline knows and the engine does not

from brevis import metrics

metrics.set("rows_loaded", 48213)          # a gauge: the last value wins
metrics.inc("vendor_rejected_total")       # a counter: it adds up
metrics.inc("bytes_discarded_total", 4096)

They come out of the engine's /metrics, on the scheduler's port, labelled with the workflow and the step:

brevis_step_rows_loaded{workflow="daily_sales",step="load"} 48213
brevis_step_vendor_rejected_total{workflow="daily_sales",step="load"} 5

This does not open a port, and it could not. A step runs in its own pod for forty seconds and exits; a port it opened would be scraped never, or once by luck. So the line goes to stdout — the pipe the engine already reads to draw a pipeline's phases — and the engine records it. Nothing to run, nothing to install, and the labels are the engine's because a step cannot know its own workflow slug.

An invalid name is refused, loudly. Prometheus accepts letters, digits and underscore, and a name it refuses costs the whole scrape — not just that metric. rows-loaded raises MetricError here rather than arriving under a name nobody wrote.

A counter cannot go down, a value has to be a number, and True is not a number: every rate() assumes the first, and metrics.set("ok", True) reporting a 1 is a number whose meaning depends on knowing that bool is an int in Python.

Run the script by hand and the line still goes to stdout, where it reads as what it is. Nothing collects it and nothing fails.

What it refuses

The ceiling is the platform's number, not ours. Every "context between tasks" feature fails the same way — somebody puts data where only context fits. Four kilobytes hold a watermark, a count or a list of partitions. Publish a reference and leave the rows where they are.

Running it by hand

Outside Brevis there is no input and nothing reads the output. get() returns the default, set() accepts and discards, and the library says so once at INFO. It does not raise: a script that cannot be run by hand cannot be developed. The validation still runs, so a value that works on your laptop works in production.

This is not a secret store

Everything published is visible to anyone who can see the run — it is in the pod's status and in the run history. Publish a path, not a signed URL.

Versions

CHANGELOG.md. Do not use 0.2.0 — it is wheel-only on PyPI, for the reason its entry there gives; 0.2.1 is the same code, complete.

This package versions on its own, separately from the engine and from the Go SDK — a fix here does not force an engine release. brevis.run needs an engine on v0.9.0 or newer to have anything to read; on an older one every field is empty and run.now() is the wall clock, which is the same thing it does outside the engine.

Testing your step

The whole contract is two environment variables, so there is nothing to mock:

def test_my_step(monkeypatch, tmp_path):
    monkeypatch.setenv("BREVIS_INPUT", '{"extract": {"rows": 42}}')
    monkeypatch.setenv("BREVIS_OUTPUT", str(tmp_path / "out"))
    assert context.get("extract.rows") == 42