Parallel functions

def describe(name):
    return {"service": name, "region": env["REGION"]}

output = steps.parallel(
    tasks = [
        steps.task(name=n, function=describe, args=[n], timeout="30s")
        for n in ["api", "worker"]
    ],
    max_concurrency = 2,
)

Pass functions, not their evaluated results. Use either functions or tasks; named tasks must be unique. max_concurrency defaults to 4. Shared globals, closures, default arguments, task arguments, and returned branch results become immutable. Create branch-local collections and return them; do not append to a shared results list. Nested groups each have their own concurrency limit, so avoid unbounded nested fan-out.

Default behavior finishes all branches and aggregates failures. fail_fast=True cancels outstanding work only after a branch's retry policy is exhausted. The caller always joins running work. A task timeout spans all attempts and backoff.

Retry repeats the entire function. Make its side effects repeatable or place the retry around a narrower function. Use existing retry keys such as max_attempts, initial_delay, backoff_strategy, and max_delay; delay is not a valid key. Provide max_attempts or a task timeout: an explicit policy without an attempt limit retries until success, timeout, or cancellation. Output-regex conditions do not apply to function tasks; they work on exec.run(..., retry={...}), where they match the failed attempt's output. A task timeout failure reads task "<name>" timed out after <duration>.

Use steps.task(..., timeout="30s") for per-task limits. A timeout: on the YAML script step is enforced: the interpreter and its subprocesses are canceled and the step fails with step timed out.

Ctrl-C cancels the script, runs defer calls with a 30 s grace, and exits 130 without an error box; a second Ctrl-C exits at once. This holds in every embedded script session.

Script output is live: lines appear as they are produced. Inside steps.parallel with more than one task, each line is prefixed [<task name>] and lines never interleave mid-line. Errors are single-line messages with the Starlark traceback shown below.