tractor: next-gen Python parallelism

tractor is a structured concurrent, multi-processing runtime built on trio.

Fundamentally tractor gives you parallelism via trio-"actors": our nurseries let you spawn new Python processes which each run a trio scheduled runtime - a call to trio.run().

We believe the system adhere's to the 3 axioms of an "actor model" but likely does not look like what you probably think an "actor model" looks like, and that's intentional.

The first step to grok tractor is to get the basics of trio down. A great place to start is the trio docs and this blog post.

Features

• It's just a trio API
• Infinitely nesteable process trees
• Builtin IPC streaming APIs with task fan-out broadcasting
• A (first ever?) "native" multi-core debugger UX for Python using pdb++
• Support for a swappable, OS specific, process spawning layer
• A modular transport stack, allowing for custom serialization (eg. msgspec), communications protocols, and environment specific IPC primitives
• Support for spawning process-level-SC, inter-loop one-to-one-task oriented asyncio actors via "infected asyncio" mode
• structured chadcurrency from the ground up

Run a func in a process

Use trio's style of focussing on tasks as functions:

"""
Run with a process monitor from a terminal using::

    $TERM -e watch -n 0.1  "pstree -a $$" \
        & python examples/parallelism/single_func.py \
        && kill $!

"""
import os

import tractor
import trio


async def burn_cpu():

    pid = os.getpid()

    # burn a core @ ~ 50kHz
    for _ in range(50000):
        await trio.sleep(1/50000/50)

    return os.getpid()


async def main():

    async with tractor.open_nursery() as n:

        portal = await n.run_in_actor(burn_cpu)

        #  burn rubber in the parent too
        await burn_cpu()

        # wait on result from target function
        pid = await portal.result()

    # end of nursery block
    print(f"Collected subproc {pid}")


if __name__ == '__main__':
    trio.run(main)

This runs burn_cpu() in a new process and reaps it on completion of the nursery block.

If you only need to run a sync function and retreive a single result, you might want to check out trio-parallel.

Zombie safe: self-destruct a process tree

tractor tries to protect you from zombies, no matter what.

"""
Run with a process monitor from a terminal using::

    $TERM -e watch -n 0.1  "pstree -a $$" \
        & python examples/parallelism/we_are_processes.py \
        && kill $!

"""
from multiprocessing import cpu_count
import os

import tractor
import trio


async def target():
    print(
        f"Yo, i'm '{tractor.current_actor().name}' "
        f"running in pid {os.getpid()}"
    )

   await trio.sleep_forever()


async def main():

    async with tractor.open_nursery() as n:

        for i in range(cpu_count()):
            await n.run_in_actor(target, name=f'worker_{i}')

        print('This process tree will self-destruct in 1 sec...')
        await trio.sleep(1)

        # raise an error in root actor/process and trigger
        # reaping of all minions
        raise Exception('Self Destructed')


if __name__ == '__main__':
    try:
        trio.run(main)
    except Exception:
        print('Zombies Contained')

If you can create zombie child processes (without using a system signal) it is a bug.

"Native" multi-process debugging

Using the magic of pdb++ and our internal IPC, we've been able to create a native feeling debugging experience for any (sub-)process in your tractor tree.

from os import getpid

import tractor
import trio


async def breakpoint_forever():
    "Indefinitely re-enter debugger in child actor."
    while True:
        yield 'yo'
        await tractor.breakpoint()


async def name_error():
    "Raise a ``NameError``"
    getattr(doggypants)


async def main():
    """Test breakpoint in a streaming actor.
    """
    async with tractor.open_nursery(
        debug_mode=True,
        loglevel='error',
    ) as n:

        p0 = await n.start_actor('bp_forever', enable_modules=[__name__])
        p1 = await n.start_actor('name_error', enable_modules=[__name__])

        # retreive results
        stream = await p0.run(breakpoint_forever)
        await p1.run(name_error)


if __name__ == '__main__':
    trio.run(main)

You can run this with:

>>> python examples/debugging/multi_daemon_subactors.py

And, yes, there's a built-in crash handling mode B)

We're hoping to add a respawn-from-repl system soon!

SC compatible bi-directional streaming

Yes, you saw it here first; we provide 2-way streams with reliable, transitive setup/teardown semantics.

Our nascent api is remniscent of trio.Nursery.start() style invocation:

import trio
import tractor


@tractor.context
async def simple_rpc(

    ctx: tractor.Context,
    data: int,

) -> None:
    '''Test a small ping-pong 2-way streaming server.

    '''
    # signal to parent that we're up much like
    # ``trio_typing.TaskStatus.started()``
    await ctx.started(data + 1)

    async with ctx.open_stream() as stream:

        count = 0
        async for msg in stream:

            assert msg == 'ping'
            await stream.send('pong')
            count += 1

        else:
            assert count == 10


async def main() -> None:

    async with tractor.open_nursery() as n:

        portal = await n.start_actor(
            'rpc_server',
            enable_modules=[__name__],
        )

        # XXX: this syntax requires py3.9
        async with (

            portal.open_context(
                simple_rpc,
                data=10,
            ) as (ctx, sent),

            ctx.open_stream() as stream,
        ):

            assert sent == 11

            count = 0
            # receive msgs using async for style
            await stream.send('ping')

            async for msg in stream:
                assert msg == 'pong'
                await stream.send('ping')
                count += 1

                if count >= 9:
                    break


        # explicitly teardown the daemon-actor
        await portal.cancel_actor()


if __name__ == '__main__':
    trio.run(main)

See original proposal and discussion in #53 as well as follow up improvements in #223 that we'd love to hear your thoughts on!

Worker poolz are easy peasy

The initial ask from most new users is "how do I make a worker pool thing?".

tractor is built to handle any SC (structured concurrent) process tree you can imagine; a "worker pool" pattern is a trivial special case.

We have a full worker pool re-implementation of the std-lib's concurrent.futures.ProcessPoolExecutor example for reference.

You can run it like so (from this dir) to see the process tree in real time:

$TERM -e watch -n 0.1  "pstree -a $$" \
    & python examples/parallelism/concurrent_actors_primes.py \
    && kill $!

This uses no extra threads, fancy semaphores or futures; all we need is tractor's IPC!

"Infected asyncio" mode

Have a bunch of asyncio code you want to force to be SC at the process level?

Check out our experimental system for guest-mode controlled asyncio actors:

import asyncio
from statistics import mean
import time

import trio
import tractor


async def aio_echo_server(
    to_trio: trio.MemorySendChannel,
    from_trio: asyncio.Queue,
) -> None:

    # a first message must be sent **from** this ``asyncio``
    # task or the ``trio`` side will never unblock from
    # ``tractor.to_asyncio.open_channel_from():``
    to_trio.send_nowait('start')

    # XXX: this uses an ``from_trio: asyncio.Queue`` currently but we
    # should probably offer something better.
    while True:
        # echo the msg back
        to_trio.send_nowait(await from_trio.get())
        await asyncio.sleep(0)


@tractor.context
async def trio_to_aio_echo_server(
    ctx: tractor.Context,
):
    # this will block until the ``asyncio`` task sends a "first"
    # message.
    async with tractor.to_asyncio.open_channel_from(
        aio_echo_server,
    ) as (first, chan):

        assert first == 'start'
        await ctx.started(first)

        async with ctx.open_stream() as stream:

            async for msg in stream:
                await chan.send(msg)

                out = await chan.receive()
                # echo back to parent actor-task
                await stream.send(out)


async def main():

    async with tractor.open_nursery() as n:
        p = await n.start_actor(
            'aio_server',
            enable_modules=[__name__],
            infect_asyncio=True,
        )
        async with p.open_context(
            trio_to_aio_echo_server,
        ) as (ctx, first):

            assert first == 'start'

            count = 0
            async with ctx.open_stream() as stream:

                delays = []
                send = time.time()

                await stream.send(count)
                async for msg in stream:
                    recv = time.time()
                    delays.append(recv - send)
                    assert msg == count
                    count += 1
                    send = time.time()
                    await stream.send(count)

                    if count >= 1e3:
                        break

        print(f'mean round trip rate (Hz): {1/mean(delays)}')
        await p.cancel_actor()


if __name__ == '__main__':
    trio.run(main)

Yes, we spawn a python process, run asyncio, start trio on the asyncio loop, then send commands to the trio scheduled tasks to tell asyncio tasks what to do XD

We need help refining the asyncio-side channel API to be more trio-like. Feel free to sling your opinion in #273!

Higher level "cluster" APIs

To be extra terse the tractor devs have started hacking some "higher level" APIs for managing actor trees/clusters. These interfaces should generally be condsidered provisional for now but we encourage you to try them and provide feedback. Here's a new API that let's you quickly spawn a flat cluster:

import trio
import tractor


async def sleepy_jane():
    uid = tractor.current_actor().uid
    print(f'Yo i am actor {uid}')
    await trio.sleep_forever()


async def main():
    '''
    Spawn a flat actor cluster, with one process per
    detected core.

    '''
    portal_map: dict[str, tractor.Portal]
    results: dict[str, str]

    # look at this hip new syntax!
    async with (

        tractor.open_actor_cluster(
            modules=[__name__]
        ) as portal_map,

        trio.open_nursery() as n,
    ):

        for (name, portal) in portal_map.items():
            n.start_soon(portal.run, sleepy_jane)

        await trio.sleep(0.5)

        # kill the cluster with a cancel
        raise KeyboardInterrupt


if __name__ == '__main__':
    try:
        trio.run(main)
    except KeyboardInterrupt:
        pass

Install

From PyPi:

pip install tractor

To try out the (optionally) faster msgspec codec instead of the default msgpack lib:

pip install tractor[msgspec]

From git:

pip install git+git://github.com/goodboy/tractor.git

Under the hood

tractor is an attempt to pair trionic structured concurrency with distributed Python. You can think of it as a trio -across-processes or simply as an opinionated replacement for the stdlib's multiprocessing but built on async programming primitives from the ground up.

Don't be scared off by this description. tractor is just trio but with nurseries for process management and cancel-able streaming IPC. If you understand how to work with trio, tractor will give you the parallelism you may have been needing.

Wait, huh?! I thought "actors" have messages, and mailboxes and stuff?!

Let's stop and ask how many canon actor model papers have you actually read ;)

From our experience many "actor systems" aren't really "actor models" since they don't adhere to the 3 axioms and pay even less attention to the problem of unbounded non-determinism (which was the whole point for creation of the model in the first place).

From the author's mouth, the only thing required is adherance to the 3 axioms, and that's it.

tractor adheres to said base requirements of an "actor model":

In response to a message, an actor may:

- send a finite number of new messages
- create a finite number of new actors
- designate a new behavior to process subsequent messages

and requires no further api changes to accomplish this.

If you want do debate this further please feel free to chime in on our chat or discuss on one of the following issues after you've read everything in them:

• https://github.com/goodboy/tractor/issues/210
• https://github.com/goodboy/tractor/issues/18

Let's clarify our parlance

Whether or not tractor has "actors" underneath should be mostly irrelevant to users other then for referring to the interactions of our primary runtime primitives: each Python process + trio.run() + surrounding IPC machinery. These are our high level, base runtime-units-of-abstraction which both are (as much as they can be in Python) and will be referred to as our "actors".

The main goal of tractor is is to allow for highly distributed software that, through the adherence to structured concurrency, results in systems which fail in predictable, recoverable and maybe even understandable ways; being an "actor model" is just one way to describe properties of the system.

What's on the TODO:

Help us push toward the future.

• (Soon to land) asyncio support allowing for "infected" actors where trio drives the asyncio scheduler via the astounding "guest mode"
• Typed messaging protocols (ex. via msgspec, see #36)
• Erlang-style supervisors via composed context managers

Feel like saying hi?

This project is very much coupled to the ongoing development of trio (i.e. tractor gets most of its ideas from that brilliant community). If you want to help, have suggestions or just want to say hi, please feel free to reach us in our matrix channel. If matrix seems too hip, we're also mostly all in the the trio gitter channel!