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 adheres 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.
- It's just a
trioAPI - 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. with msgspec), communications protocols, and environment specific IPC primitives
- Support for spawning process-level-SC, inter-loop one-to-one-task oriented
asyncioactors via "infectedasyncio" mode - structured chadcurrency from the ground up
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 $$" \
& python examples/parallelism/single_func.py \
&& kill $!
"""
import os
import tractor
import trio
async def burn_cpu():
pid = os.getpid()
# burn a core @ ~ 50kHz
for _ in range(50000):
await trio.sleep(1/50000/50)
return os.getpid()
async def main():
async with tractor.open_nursery() as n:
portal = await n.run_in_actor(burn_cpu)
# burn rubber in the parent too
await burn_cpu()
# wait on result from target function
pid = await portal.result()
# end of nursery block
print(f"Collected subproc {pid}")
if __name__ == '__main__':
trio.run(main)This runs burn_cpu() in a new process and reaps it on completion
of the nursery block.
If you only need to run a sync function and retreive a single result, you might want to check out trio-parallel.
tractor tries to protect you from zombies, no matter what.
"""
Run with a process monitor from a terminal using::
$TERM -e watch -n 0.1 "pstree -a $$" \
& 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.
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!
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!
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!
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!
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:
passFrom PyPi:
pip install tractor
From git:
pip install git+git://github.com/goodboy/tractor.git
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.
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:
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.
Help us push toward the future of distributed Python.
- Erlang-style supervisors via composed context managers (see #22)
- Typed messaging protocols (ex. via
msgspec.Struct, see #36) - Typed capability-based (dialog) protocols ( see #196 with draft work started in #311)
- We recently disabled CI-testing on windows and need help getting it running again! (see #327). We do have windows support (and have for quite a while) but since no active hacker exists in the user-base to help test on that OS, for now we're not actively maintaining testing due to the added hassle and general latency..
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!