24 KiB
tractor
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.
Install
No PyPi release yet!
pip install git+git://github.com/goodboy/tractor.git
Feel like saying hi?
This project is very much coupled to the ongoing development of trio
(i.e. tractor
gets all its ideas from that brilliant community). If you want to help, have suggestions or just want to say hi, please feel free to ping me on the trio gitter channel!
Philosophy
Our tenets non-comprehensively include:
- strict adherence to the concept-in-progress of structured concurrency
- no spawning of processes willy-nilly; causality is paramount!
- (remote) errors always propagate back to the parent supervisor
- verbatim support for
trio
's cancellation system - shared nothing architecture
- no use of proxy objects or shared references between processes
- an immersive debugging experience
- anti-fragility through chaos engineering
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.
Warning
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!
Examples
Note, if you are on Windows please be sure to see the gotchas
<windowsgotchas>
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:
../examples/a_trynamic_first_scene.py
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:
../examples/actor_spawning_and_causality.py
What's going on?
- an initial actor is started with
trio.run()
and told to execute its main task:main()
- inside
main()
an actor is spawned using anActorNusery
and is told to run a single function:cellar_door()
- a
portal
instance (we'll get to what it is shortly) returned fromnursery.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 executescellar_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). Thewith tractor.open_nursery()
exits only once all actors' main function/task complete (just like the nursery intrio
) - actors can be spawned to live forever using the
start_actor()
method and act like an RPC daemon that runs indefinitely (thewith tractor.open_nursery()
won't exit) until cancelled
Here is a similar example using the latter method:
../examples/actor_spawning_and_causality_with_daemon.py
The enable_modules
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 Portal.run()
. 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).
Cancellation
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.
../examples/remote_error_propagation.py
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 portal.run()
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 Portal.run()
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.
Streaming
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:
../examples/asynchronous_generators.py
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:
@tractor.stream
async def streamer(ctx: tractor.Context, rate: int = 2) -> None:
"""A simple web response streaming server.
"""
while True:
= await web_request('http://data.feed.com')
val
# 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 @tractor.stream
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,int = 2
rate: -> None:
) """A simple web response streaming server.
"""
while True:
= await web_request(url)
val
# this is the same as ``yield`` in the async gen case
await ctx.send_yield(val)
await trio.sleep(1 / rate)
@tractor.stream
async def stream_multiple_sources(
ctx: tractor.Context,str]
sources: List[-> 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:
../examples/full_fledged_streaming_service.py
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 (aka process global) variables
Although tractor
uses a shared-nothing architecture between processes you can of course share state between tasks running within an actor (since a trio.run() runtime is single threaded). trio
tasks spawned via multiple RPC calls to an actor can modify process-global-state defined using Python module attributes:
# a per process cache
str, bool] = {}
_actor_cache: Dict[
def ping_endpoints(endpoints: List[str]):
"""Start a polling process which runs completely separate
from our root actor/process.
"""
# This runs in a new process so no changes # will propagate
# back to the parent actor
while True:
for ep in endpoints:
= await check_endpoint_is_up(ep)
status = status
_actor_cache[ep]
await trio.sleep(0.5)
async def get_alive_endpoints():
nonlocal _actor_cache
return {key for key, value in _actor_cache.items() if value}
async def main():
async with tractor.open_nursery() as n:
= await n.run_in_actor(ping_endpoints)
portal
# print the alive endpoints after 3 seconds
await trio.sleep(3)
# this is submitted to be run in our "ping_endpoints" actor
print(await portal.run(get_alive_endpoints))
You can pass any kind of (msgpack) serializable data between actors using function call semantics but 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:
../examples/service_discovery.py
The name
value you should pass to find_actor()
is the one you passed as the first argument to either trio.run()
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:
import trio
import tractor
async def main():
async with tractor.open_root_actor(
=('192.168.0.10', 1616)
arbiter_addr
):await trio.sleep_forever()
trio.run(main)
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 tractor.open_nursery()
.
Currently the options available are:
trio
: atrio
-native spawner which is an async wrapper aroundsubprocess
spawn
: one of the stdlib'smultiprocessing
start methodsforkserver
: a fastermultiprocessing
variant that is Unix only
trio
The trio
backend offers a lightweight async wrapper around subprocess
from the standard library and takes advantage of the trio.
open_process API.
multiprocessing
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 __main__.py 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\_actor.py", 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/__main__.py
import trio
import tractor
import multiprocessing
from . import tractor_app
if __name__ == '__main__':
multiprocessing.freeze_support() trio.run(tractor_app.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:
../examples/__main__.py
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
= get_console_log('trace') log
What the future holds
Stuff I'd like to see tractor
do real soon:
- TLS, duh.
- erlang-like supervisors
- native support for nanomsg as a channel transport
- native gossip protocol support for service discovery and arbiter election
- a distributed log ledger for tracking cluster behaviour
- a slick multi-process aware debugger much like in celery but with better pdb++ support
- an extensive chaos engineering test suite
- support for reactive programming primitives and native support for asyncitertools like libs
- introduction of a capability-based security model