async-pipeline 0.9.0

Lightweight composition of async pipelines where each stage receives the previous stage's output.

async-pipeline

A small library for composing async pipelines: each Stage receives the previous stage’s output, executed in order.

Requirements

• Python 3.14 or newer

Install with uv

In your project:

uv add async-pipeline

To work on this library:

git clone <repo-url>
cd async-pipeline
uv sync

Basic usage

from async_pipeline import Pipeline, Stage

async def add_one(value: int) -> int:
    return value + 1

async def multiply_by_two(value: int) -> int:
    return value * 2

pipeline = Pipeline([
    Stage("add_one", add_one),
    Stage("multiply_by_two", multiply_by_two),
])

result = await pipeline.run(10)
assert result == 22

Synchronous handlers are supported as well (the stage’s run method remains async):

def add_one(value: int) -> int:
    return value + 1

pipeline = Pipeline([
    Stage("add_one", add_one),
])

result = await pipeline.run(1)
assert result == 2

Errors

Failures inside a handler are surfaced as StageExecutionError, including the stage name and the original exception:

from async_pipeline import Pipeline, Stage, StageExecutionError

async def broken(value: int) -> int:
    raise RuntimeError("boom")

pipeline = Pipeline([
    Stage("broken", broken),
])

try:
    await pipeline.run(1)
except StageExecutionError as exc:
    assert exc.stage_name == "broken"
    assert isinstance(exc.original_error, RuntimeError)

A Pipeline with no stages raises ValueError at construction time.

Stage timeout

Per-stage timeouts apply only when the handler returns an awaitable (async handler). Synchronous handlers are unchanged; the timeout argument is ignored for them.

Timeouts use asyncio.timeout (not wait_for). If the awaitable runs longer than the limit, the stage raises StageExecutionError with original_error set to TimeoutError.

import asyncio

from async_pipeline import Pipeline, Stage

async def fetch_data(value: int) -> int:
    await asyncio.sleep(2)
    return value

pipeline = Pipeline([
    Stage("fetch_data", fetch_data, timeout=1.0),
])

Invalid values (timeout <= 0) raise ValueError with message timeout must be greater than 0.

Execution context

Pass a mutable dict (typically dict[str, Any]) to Pipeline.run(..., context=...) to share data across stages and hooks for that run. If context=None, the pipeline creates a new empty dict for that execution and reuses it for every stage in order.

Handlers can stay async def handler(value) or opt into async def handler(value, context) — the library uses inspect.signature: if the callable has at least two positional parameters, the context dict is passed as the second argument.

Hooks support the same pattern:

• before_stage(name, input) or before_stage(name, input, context) (three or more parameters → context is passed).
• after_stage(name, input, output, error) or after_stage(..., context) (five or more parameters → context is passed).

Pipeline.map accepts context= and applies a shallow copy (dict(template)) per item, so concurrent workers never share the same dict instance.

from async_pipeline import Pipeline, Stage


async def handler(value: int, context: dict[str, object]) -> int:
    context["user_id"] = 123
    return value


pipeline = Pipeline([Stage("step", handler)])
await pipeline.run(1, context={})

Retry

Configure automatic re-runs when a handler raises an Exception (including TimeoutError from asyncio.timeout). On success, the stage returns immediately. If every attempt fails, the library raises StageExecutionError with the last exception as original_error.

• retries — extra attempts after the first (retries=0 is the default: no retries). Total tries are 1 + retries.
• retry_delay — base seconds to wait after a failed attempt before the next one. If 0, no asyncio.sleep is used between attempts.
• backoff — "fixed" (same delay after each failure) or "exponential" (delay multiplies by 2 ** (failure_number - 1) from the base retry_delay, e.g. 0.5s, 1.0s, 2.0s, …).

CancelledError and KeyboardInterrupt are BaseException, not Exception, so they are not retried and propagate as usual.

from async_pipeline import Stage

stage = Stage(
    "api_call",
    api_call,
    retries=3,
    retry_delay=0.5,
    backoff="exponential",
    timeout=5.0,
)

With timeout: each attempt is wrapped in asyncio.timeout when timeout is set, so a slow await can fail with TimeoutError, trigger a retry (after the backoff sleep), and eventually surface as StageExecutionError if all tries time out or fail.

Invalid retries, retry_delay, or backoff values raise ValueError with a clear message.

Hooks

Pipeline accepts optional before_stage and after_stage callbacks. They run around each Stage inside run() (and therefore around map(), which calls run() per item).

• before_stage(stage_name, input_value) — runs immediately before stage.run(...) (optionally before_stage(..., context) — see Execution context).
• after_stage(stage_name, input_value, output_value, error) — runs after the stage finishes (optionally after_stage(..., context)). On success, output_value is the stage result and error is None. On failure, output_value is None and error is the exception raised by stage.run (typically StageExecutionError).

Hooks may be sync or async (if they return an awaitable, it is awaited). Failures inside hooks are ignored (they do not replace or mask stage errors, and they do not stop the pipeline). There is no built-in logging so the library stays opinion-free.

Typical uses: logging, metrics, tracing, auditing, and debugging without coupling that logic to stage handlers.

from async_pipeline import Pipeline, Stage


def before_stage(stage_name: str, input_value: object) -> None:
    print(f"Starting {stage_name}")


def after_stage(
    stage_name: str,
    input_value: object,
    output_value: object | None,
    error: Exception | None,
) -> None:
    if error:
        print(f"{stage_name} failed: {error}")
    else:
        print(f"{stage_name} finished: {output_value}")


async def add_one(value: int) -> int:
    return value + 1


pipeline = Pipeline(
    [Stage("add_one", add_one)],
    before_stage=before_stage,
    after_stage=after_stage,
)

Async example:

async def before_stage(stage_name: str, input_value: object) -> None:
    await audit_log(stage_name, input_value)

Middleware

Middlewares wrap each stage’s execution in a chain: the first item in middlewares runs outermost (before the second, and so on, until the stage). Each middleware receives next, a callable that continues the chain with the current input value (you may pass a different value into next to change what the stage sees). Use middlewares=[...] on Pipeline.

Unlike hooks, middlewares participate in the data path: they can transform inputs and outputs, catch or transform errors, and implement cross-cutting behavior (logging, tracing, policies) with full control over await next(value).

Hooks stay lightweight observers: they run before the middleware chain (before_stage) and after the full stage completes (after_stage), cannot replace the chain, and their own failures are ignored by design.

import time

from async_pipeline import Pipeline, Stage


async def timing_middleware(next, stage_name, value, context):
    start = time.perf_counter()
    result = await next(value)
    duration = time.perf_counter() - start
    print(f"{stage_name} took {duration:.3f}s")
    return result

Built-in middlewares

The library ships reusable middlewares under async_pipeline.middlewares. They compose with Pipeline.run, Pipeline.map, hooks, execution context, and optional OpenTelemetry.

from async_pipeline.middlewares import (
    LoggingMiddleware,
    RetryMiddleware,
    TimeoutMiddleware,
    TimingMiddleware,
)

Order matters: the first entry in middlewares=[...] is the outermost wrapper (runs first before the rest, closest to the caller). The last entry sits just before the stage. Example:

middlewares=[
    LoggingMiddleware(),
    TimingMiddleware(),
    RetryMiddleware(retries=3),
    TimeoutMiddleware(timeout=5.0),
]

LoggingMiddleware

pipeline = Pipeline(
    [...],
    middlewares=[LoggingMiddleware()],
)

TimingMiddleware

Stores per-stage durations (seconds) under context["timings"][stage_name] as a list (multiple runs append).

context = {}
pipeline = Pipeline(
    [...],
    middlewares=[TimingMiddleware()],
)
result = await pipeline.run(1, context=context)
print(context["timings"])

RetryMiddleware

Retries await next(...) (in addition to any Stage(..., retries=...) policy). Does not retry KeyboardInterrupt or asyncio.CancelledError.

pipeline = Pipeline(
    [...],
    middlewares=[
        RetryMiddleware(
            retries=3,
            delay=0.5,
            backoff="exponential",
        ),
    ],
)

TimeoutMiddleware

Wraps the next hop in asyncio.timeout (not wait_for). Compatible with Stage.timeout, which still applies inside the stage.

pipeline = Pipeline(
    [...],
    middlewares=[TimeoutMiddleware(timeout=5.0)],
)

OpenTelemetry

Tracing is optional. The core package does not depend on OpenTelemetry; install the extra when you want spans around each stage:

uv add "async-pipeline[otel]"

Use OpenTelemetryMiddleware in middlewares=[...]. It creates one span per stage, named {span_prefix}.{stage_name} (defaults: span_prefix="pipeline.stage" → e.g. pipeline.stage.fetch_api).

from async_pipeline import Pipeline, Stage
from async_pipeline.telemetry import OpenTelemetryMiddleware

pipeline = Pipeline(
    [
        Stage("add_one", add_one),
    ],
    middlewares=[OpenTelemetryMiddleware()],
)

You can attach simple attributes from the execution context under trace_attributes (only str, int, float, bool values are copied onto the span; other types are ignored):

result = await pipeline.run(
    1,
    context={
        "trace_attributes": {
            "request_id": "abc-123",
        }
    },
)

Batch processing

Run the same pipeline for many inputs in parallel, with a fixed concurrency limit and stable output order (aligned with the input sequence):

results = await pipeline.map([1, 2, 3], concurrency=5)

Implementation notes:

• Uses asyncio.TaskGroup to run one async worker per item (not gather).
• Uses asyncio.Semaphore so at most concurrency pipelines run at once; workers still start as tasks, but only concurrency of them proceed past the semaphore at a time.
• Each worker calls run() for its item and writes into a pre-sized list by index, so results stay in input order even when tasks finish out of order.

Errors (default): if any item fails, TaskGroup surfaces an ExceptionGroup (and cancels the other workers). StageExecutionError from a stage is propagated like in run() (wrapped inside the group as needed).

Errors (return_exceptions=True): failures are stored in the result list in the matching position as the exception object; the TaskGroup completes without raising, so you get list entries that are either normal outputs or an Exception (often StageExecutionError).

Development commands

uv sync
uv run pytest
uv run ruff check .
uv run mypy src

Changelog

Release history: CHANGELOG.md.

License

See the LICENSE file.