A structured concurrent, async-native “actor model” built on trio and multiprocessing.

tractor is an attempt to bring trionic structured concurrency to distributed multi-core Python; it aims to be the Python multi-processing framework you always wanted.

tractor lets you spawn trio “actors”: processes which each run a trio scheduled task tree (also known as an async sandwich). Actors communicate by exchanging asynchronous messages and avoid sharing any state. This model allows for highly distributed software architecture which works just as well on multiple cores as it does over many hosts.

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.


No PyPi release yet!

pip install git+git://


Our tenets non-comprehensively include:

tractor is an actor-model-like system in the sense that it adheres to the 3 axioms but does not (yet) fulfil all “unrequirements” in practise. It is an experiment in applying structured concurrency constraints on a parallel processing system where multiple Python processes exist over many hosts but no process can outlive its parent. In erlang parlance, it is an architecture where every process has a mandatory supervisor enforced by the type system. The API design is almost exclusively inspired by trio’s concepts and primitives (though we often lag a little). As a distributed computing system tractor attempts to place sophistication at the correct layer such that concurrency primitives are powerful yet simple, making it easy to build complex systems (you can build a “worker pool” architecture but it’s definitely not required). There is first class support for inter-actor streaming using async generators and ongoing work toward a functional reactive style for IPC.


tractor is in alpha-alpha and is expected to change rapidly! Expect nothing to be set in stone. Your ideas about where it should go are greatly appreciated!


Note, if you are on Windows please be sure to see the gotchas section before trying these.

A trynamic first scene

Let’s direct a couple actors and have them run their lines for the hip new film we’re shooting:

import tractor

_this_module = __name__
the_line = 'Hi my name is {}'


async def hi():
    return the_line.format(tractor.current_actor().name)

async def say_hello(other_actor):
    async with tractor.wait_for_actor(other_actor) as portal:
        return await, 'hi')

async def main():
    """Main tractor entry point, the "master" process (for now
    acts as the "director").
    async with tractor.open_nursery() as n:
        print("Alright... Action!")

        donny = await n.run_in_actor(
            # arguments are always named
        gretchen = await n.run_in_actor(
        print(await gretchen.result())
        print(await donny.result())
        print("CUTTTT CUUTT CUT!!! Donny!! You're supposed to say...")

if __name__ == '__main__':

We spawn two actors, donny and gretchen. Each actor starts up and executes their main task defined by an async function, say_hello(). The function instructs each actor to find their partner and say hello by calling their partner’s hi() function using something called a portal. Each actor receives a response and relays that back to the parent actor (in this case our “director” executing main()).

Actor spawning and causality

tractor tries to take trio’s concept of causal task lifetimes to multi-process land. Accordingly, tractor’s actor nursery behaves similar to trio’s nursery. That is, tractor.open_nursery() opens an ActorNursery which must wait on spawned actors to complete (or error) in the same causal way trio waits on spawned subtasks. This includes errors from any one actor causing all other actors spawned by the same nursery to be cancelled.

To spawn an actor and run a function in it, open a nursery block and use the run_in_actor() method:

import tractor

def cellar_door():
   return "Dang that's beautiful"

async def main():
    """The main ``tractor`` routine.
    async with tractor.open_nursery() as n:

        portal = await n.run_in_actor('some_linguist', cellar_door)

    # The ``async with`` will unblock here since the 'some_linguist'
    # actor has completed its main task ``cellar_door``.

    print(await portal.result())

if __name__ == '__main__':

What’s going on?

  • an initial actor is started with and told to execute its main task: main()
  • inside main() an actor is spawned using an ActorNusery and is told to run a single function: cellar_door()
  • a portal instance (we’ll get to what it is shortly) returned from nursery.run_in_actor() is used to communicate with the newly spawned sub-actor
  • the second actor, some_linguist, in a new process running a new trio task then executes cellar_door() and returns its result over a channel back to the parent actor
  • the parent actor retrieves the subactor’s final result using portal.result() much like you’d expect from a future.

This run_in_actor() API should look very familiar to users of asyncio’s run_in_executor() which uses a concurrent.futures Executor.

Since you might also want to spawn long running worker or daemon actors, each actor’s lifetime can be determined based on the spawn method:

  • if the actor is spawned using run_in_actor() it terminates when its main task completes (i.e. when the (async) function submitted to it returns). The with tractor.open_nursery() exits only once all actors’ main function/task complete (just like the nursery in trio)
  • actors can be spawned to live forever using the start_actor() method and act like an RPC daemon that runs indefinitely (the with tractor.open_nursery() won’t exit) until cancelled

Here is a similar example using the latter method:

import tractor

def movie_theatre_question():
    """A question asked in a dark theatre, in a tangent
    (errr, I mean different) process.
    return 'have you ever seen a portal?'

async def main():
    """The main ``tractor`` routine.
    async with tractor.open_nursery() as n:

        portal = await n.start_actor(
            # enable the actor to run funcs from this current module

        print(await, 'movie_theatre_question'))
        # call the subactor a 2nd time
        print(await, 'movie_theatre_question'))

        # the async with will block here indefinitely waiting
        # for our actor "frank" to complete, but since it's an
        # "outlive_main" actor it will never end until cancelled
        await portal.cancel_actor()

if __name__ == '__main__':

The rpc_module_paths kwarg above is a list of module path strings that will be loaded and made accessible for execution in the remote actor through a call to For now this is a simple mechanism to restrict the functionality of the remote (and possibly daemonized) actor and uses Python’s module system to limit the allowed remote function namespace(s).

tractor is opinionated about the underlying threading model used for each actor. Since Python has a GIL and an actor model by definition shares no state between actors, it fits naturally to use a multiprocessing Process. This allows tractor programs to leverage not only multi-core hardware but also distribute over many hardware hosts (each actor can talk to all others with ease over standard network protocols).


tractor supports trio’s cancellation system verbatim. Cancelling a nursery block cancels all actors spawned by it. Eventually tractor plans to support different supervision strategies like erlang.

Remote error propagation

Any task invoked in a remote actor should ship any error(s) back to the calling actor where it is raised and expected to be dealt with. This way remote actors are never cancelled unless explicitly asked or there’s a bug in tractor itself.

import tractor

async def assert_err():
    assert 0

async def main():
    async with tractor.open_nursery() as n:
        real_actors = []
        for i in range(3):
            real_actors.append(await n.start_actor(

        # start one actor that will fail immediately
        await n.run_in_actor('extra', assert_err)

    # should error here with a ``RemoteActorError`` containing
    # an ``AssertionError`` and all the other actors have been cancelled

if __name__ == '__main__':
        # also raises
    except tractor.RemoteActorError:
        print("Look Maa that actor failed hard, hehhh!")

You’ll notice the nursery cancellation conducts a one-cancels-all supervisory strategy exactly like trio. The plan is to add more erlang strategies in the near future by allowing nurseries to accept a Supervisor type.

IPC using portals

tractor introduces the concept of a portal which is an API borrowed from trio. A portal may seem similar to the idea of a RPC future except a portal allows invoking remote async functions and generators and intermittently blocking to receive responses. This allows for fully async-native IPC between actors.

When you invoke another actor’s routines using a portal it looks as though it was called locally in the current actor. So when you see a call to await what you get back is what you’d expect to if you’d called the function directly in-process. This approach avoids the need to add any special RPC proxy objects to the library by instead just relying on the built-in (async) function calling semantics and protocols of Python.

Depending on the function type tries to correctly interface exactly like a local version of the remote built-in Python function type. Currently async functions, generators, and regular functions are supported. Inspiration for this API comes remote function execution but without the client code being concerned about the underlying channels system or shipping code over the network.

This portal approach turns out to be paricularly exciting with the introduction of asynchronous generators in Python 3.6! It means that actors can compose nicely in a data streaming pipeline.


By now you’ve figured out that tractor lets you spawn process based actors that can invoke cross-process (async) functions and all with structured concurrency built in. But the real cool stuff is the native support for cross-process streaming.

Asynchronous generators

The default streaming function is simply an async generator definition. Every value yielded from the generator is delivered to the calling portal exactly like if you had invoked the function in-process meaning you can async for to receive each value on the calling side.

As an example here’s a parent actor that streams for 1 second from a spawned subactor:

from itertools import repeat
import trio
import tractor


async def stream_forever():
    for i in repeat("I can see these little future bubble things"):
        # each yielded value is sent over the ``Channel`` to the
        # parent actor
        yield i
        await trio.sleep(0.01)

async def main():
    # stream for at most 1 seconds
    with trio.move_on_after(1) as cancel_scope:
        async with tractor.open_nursery() as n:
            portal = await n.start_actor(

            # this async for loop streams values from the above
            # async generator running in a separate process
            async for letter in await, 'stream_forever'):

    # we support trio's cancellation system
    assert cancel_scope.cancelled_caught
    assert n.cancelled

if __name__ == '__main__':, start_method='forkserver')

By default async generator functions are treated as inter-actor streams when invoked via a portal (how else could you really interface with them anyway) so no special syntax to denote the streaming service is necessary.

Channels and Contexts

If you aren’t fond of having to write an async generator to stream data between actors (or need something more flexible) you can instead use a Context. A context wraps an actor-local spawned task and a Channel so that tasks executing across multiple processes can stream data to one another using a low level, request oriented API.

A Channel wraps an underlying transport and interchange format to enable inter-actor-communication. In its present state tractor uses TCP and msgpack.

As an example if you wanted to create a streaming server without writing an async generator that yields values you instead define a decorated async function:
async def streamer(ctx: tractor.Context, rate: int = 2) -> None:
   """A simple web response streaming server.
   while True:
      val = await web_request('')

      # this is the same as ``yield`` in the async gen case
      await ctx.send_yield(val)

      await trio.sleep(1 / rate)

You must decorate the function with and declare a ctx argument as the first in your function signature and then tractor will treat the async function like an async generator - as a stream from the calling/client side.

This turns out to be handy particularly if you have multiple tasks pushing responses concurrently:

async def streamer(
   ctx: tractor.Context,
   rate: int = 2
) -> None:
   """A simple web response streaming server.
   while True:
      val = await web_request(url)

      # this is the same as ``yield`` in the async gen case
      await ctx.send_yield(val)

      await trio.sleep(1 / rate)
async def stream_multiple_sources(
   ctx: tractor.Context,
   sources: List[str]
) -> None:
   async with trio.open_nursery() as n:
      for url in sources:
         n.start_soon(streamer, ctx, url)

The context notion comes from the context in nanomsg.

A full fledged streaming service

Alright, let’s get fancy.

Say you wanted to spawn two actors which each pull data feeds from two different sources (and wanted this work spread across 2 cpus). You also want to aggregate these feeds, do some processing on them and then deliver the final result stream to a client (or in this case parent) actor and print the results to your screen:

import time
import trio
import tractor

# this is the first 2 actors, streamer_1 and streamer_2
async def stream_data(seed):
    for i in range(seed):
        yield i
        await trio.sleep(0)  # trigger scheduler

# this is the third actor; the aggregator
async def aggregate(seed):
    """Ensure that the two streams we receive match but only stream
    a single set of values to the parent.
    async with tractor.open_nursery() as nursery:
        portals = []
        for i in range(1, 3):
            # fork point
            portal = await nursery.start_actor(


        send_chan, recv_chan = trio.open_memory_channel(500)

        async def push_to_chan(portal, send_chan):
            async with send_chan:
                async for value in await
                    __name__, 'stream_data', seed=seed
                    # leverage trio's built-in backpressure
                    await send_chan.send(value)

            print(f"FINISHED ITERATING {}")

        # spawn 2 trio tasks to collect streams and push to a local queue
        async with trio.open_nursery() as n:

            for portal in portals:
                n.start_soon(push_to_chan, portal, send_chan.clone())

            # close this local task's reference to send side
            await send_chan.aclose()

            unique_vals = set()
            async with recv_chan:
                async for value in recv_chan:
                    if value not in unique_vals:
                        # yield upwards to the spawning parent actor
                        yield value

                assert value in unique_vals

            print("FINISHED ITERATING in aggregator")

        await nursery.cancel()
        print("WAITING on `ActorNursery` to finish")

# this is the main actor and *arbiter*
async def main():
    # a nursery which spawns "actors"
    async with tractor.open_nursery() as nursery:

        seed = int(1e3)
        import time
        pre_start = time.time()

        portal = await nursery.run_in_actor(

        start = time.time()
        # the portal call returns exactly what you'd expect
        # as if the remote "aggregate" function was called locally
        result_stream = []
        async for value in await portal.result():

        print(f"STREAM TIME = {time.time() - start}")
        print(f"STREAM + SPAWN TIME = {time.time() - pre_start}")
        assert result_stream == list(range(seed))
        return result_stream

if __name__ == '__main__':
    final_stream =, arbiter_addr=('', 1616))

Here there’s four actors running in separate processes (using all the cores on you machine). Two are streaming by yielding values from the stream_data() async generator, one is aggregating values from those two in aggregate() (also an async generator) and shipping the single stream of unique values up the parent actor (the 'MainProcess' as multiprocessing calls it) which is running main().

Actor local variables

Although tractor uses a shared-nothing architecture between processes you can of course share state between tasks running within an actor. trio tasks spawned via multiple RPC calls to an actor can access global state using the per actor statespace dictionary:

statespace = {'doggy': 10}

def check_statespace():
    # Remember this runs in a new process so no changes
    # will propagate back to the parent actor
    assert tractor.current_actor().statespace == statespace

async def main():
    async with tractor.open_nursery() as n:
        await n.run_in_actor(

Of course you don’t have to use the statespace variable (it’s mostly a convenience for passing simple data to newly spawned actors); building out a state sharing system per-actor is totally up to you.

Service Discovery

Though it will be built out much more in the near future, tractor currently keeps track of actors by (name: str, id: str) using a special actor called the arbiter. Currently the arbiter must exist on a host (or it will be created if one can’t be found) and keeps a simple dict of actor names to sockets for discovery by other actors. Obviously this can be made more sophisticated (help me with it!) but for now it does the trick.

To find the arbiter from the current actor use the get_arbiter() function and to find an actor’s socket address by name use the find_actor() function:

import tractor


async def main(service_name):

    async with tractor.open_nursery() as an:
        await an.start_actor(service_name)

        async with tractor.get_arbiter('', 1616) as portal:
            print(f"Arbiter is listening on {}")

        async with tractor.wait_for_actor(service_name) as sockaddr:
            print(f"my_service is found at {sockaddr}")

        await an.cancel()

if __name__ == '__main__':, 'some_actor_name')

The name value you should pass to find_actor() is the one you passed as the first argument to either or ActorNursery.start_actor().

Running actors standalone

You don’t have to spawn any actors using open_nursery() if you just want to run a single actor that connects to an existing cluster. All the comms and arbiter registration stuff still works. This can somtimes turn out being handy when debugging mult-process apps when you need to hop into a debugger. You just need to pass the existing arbiter’s socket address you’d like to connect to:, arbiter_addr=('', 1616))

Choosing a process spawning backend

tractor is architected to support multiple actor (sub-process) spawning backends. Specific defaults are chosen based on your system but you can also explicitly select a backend of choice at startup via a start_method kwarg to

Currently the options available are:

  • trio: a trio-native spawner which is an async wrapper around subprocess
  • spawn: one of the stdlib’s multiprocessing start methods
  • forkserver: a faster multiprocessing variant that is Unix only


The trio backend offers a lightweight async wrapper around subprocess from the standard library and takes advantage of the trio. open_process API.


There is support for the stdlib’s multiprocessing start methods. Note that on Windows spawn it the only supported method and on *nix systems forkserver is the best method for speed but has the caveat that it will break easily (hangs due to broken pipes) if spawning actors using nested nurseries.

In general, the multiprocessing backend has not proven reliable for handling errors from actors more then 2 nurseries deep (see #89). If you for some reason need this consider sticking with alternative backends.

Windows “gotchas”

On Windows (which requires the use of the stdlib’s multiprocessing package) there are some gotchas. Namely, the need for calling freeze_support() inside the __main__ context. Additionally you may need place you tractor program entry point in a seperate module in your package in order to avoid an error like the following

Traceback (most recent call last):
  File "C:\ProgramData\Miniconda3\envs\tractor19030601\lib\site-packages\tractor\", line 234, in _get_rpc_func
    return getattr(self._mods[ns], funcname)
KeyError: '__mp_main__'

To avoid this, the following is the only code that should be in your main python module of the program:

# application/
import tractor
import multiprocessing
from . import tractor_app

if __name__ == '__main__':

And execute as:

python -m application

As an example we use the following code to test all documented examples in the test suite on windows:

Needed on Windows.

This module is needed as the program entry point for invocation
with ``python -m <modulename>``. See the solution from @chrizzFTD

if __name__ == '__main__':
    import multiprocessing
    # ``tests/`` will copy each
    # script from this examples directory into a module in a new
    # temporary dir and name it We import that script
    # module here and invoke it's ``main()``.
    from . import test_example, start_method='spawn')

See #61 and #79 for further details.

Enabling logging

Considering how complicated distributed software can become it helps to know what exactly it’s doing (even at the lowest levels). Luckily tractor has tons of logging throughout the core. tractor isn’t opinionated on how you use this information and users are expected to consume log messages in whichever way is appropriate for the system at hand. That being said, when hacking on tractor there is a prettified console formatter which you can enable to see what the heck is going on. Just put the following somewhere in your code:

from tractor.log import get_console_log
log = get_console_log('trace')

What the future holds

Stuff I’d like to see tractor do real soon: