trio
Welcome to the Trio tutorial! Trio is a modern Python library for writing asynchronous applications – that is, programs that want to do multiple things at the same time with parallelized I/O, like a web spider that fetches lots of pages in parallel, a web server juggling lots of simultaneous downloads... that sort of thing. Here we'll try to give a gentle introduction to asynchronous programming with Trio.
We assume that you're familiar with Python in general, but don't worry – we don't assume you know anything about asynchronous programming or Python's new async/await
feature.
Also, unlike many async/await
tutorials, we assume that your goal is to use Trio to write interesting programs, so we won't go into the nitty-gritty details of how async/await
is implemented inside the Python interpreter. The word "coroutine" is never mentioned. The fact is, you really don't need to know any of that stuff unless you want to implement a library like Trio, so we leave it out (though we'll throw in a few links for those who want to dig deeper).
Okay, ready? Let's get started.
- Make sure you're using Python 3.6 or newer.
python3 -m pip install --upgrade trio
(or on Windows, maybepy -3 -m pip install --upgrade trio
– details)- Can you
import trio
? If so then you're good to go!
...then we want to know! We have a friendly chat channel, you can ask questions using the "python-trio" tag on StackOverflow, or just file a bug (if our documentation is confusing, that's our fault, and we want to fix it!).
Python 3.5 added a major new feature: async functions. Using Trio is all about writing async functions, so let's start there.
An async function is defined like a normal function, except you write async def
instead of def
:
# A regular function
def regular_double(x):
return 2 * x
# An async function
async def async_double(x):
return 2 * x
"Async" is short for "asynchronous"; we'll sometimes refer to regular functions like regular_double
as "synchronous functions", to distinguish them from async functions.
From a user's point of view, there are two differences between an async function and a regular function:
- To call an async function, you have to use the
await
keyword. So instead of writingregular_double(3)
, you writeawait async_double(3)
. You can't use the
await
keyword inside the body of a regular function. If you try it, you'll get a syntax error:def print_double(x): print(await async_double(x)) # <-- SyntaxError here
But inside an async function,
await
is allowed:async def print_double(x): print(await async_double(x)) # <-- OK!
Now, let's think about the consequences here: if you need await
to call an async function, and only async functions can use await
... here's a little table:
If a function like this | wants to call a function like this | is it gonna happen? |
---|---|---|
sync | sync | ✓ |
sync | async | NOPE |
async | sync | ✓ |
async | async | ✓ |
So in summary: As a user, the entire advantage of async functions over regular functions is that async functions have a superpower: they can call other async functions.
This immediately raises two questions: how, and why? Specifically:
When your Python program starts up, it's running regular old sync code. So there's a chicken-and-the-egg problem: once we're running an async function we can call other async functions, but how do we call that first async function?
And, if the only reason to write an async function is that it can call other async functions, why on earth would we ever use them in the first place? I mean, as superpowers go this seems a bit pointless. Wouldn't it be simpler to just... not use any async functions at all?
This is where an async library like Trio comes in. It provides two things:
A runner function, which is a special synchronous function that takes and calls an asynchronous function. In Trio, this is
trio.run
:import trio async def async_double(x): return 2 * x trio.run(async_double, 3) # returns 6
So that answers the "how" part.
A bunch of useful async functions – in particular, functions for doing I/O. So that answers the "why": these functions are async, and they're useful, so if you want to use them, you have to write async code. If you think keeping track of these
async
andawait
things is annoying, then too bad – you've got no choice in the matter! (Well, OK, you could just not use Trio. That's a legitimate option. But it turns out that theasync/await
stuff is actually a good thing, for reasons we'll discuss a little bit later.)Here's an example function that uses
trio.sleep
. (trio.sleep
is liketime.sleep
, but with more async.)import trio async def double_sleep(x): await trio.sleep(2 * x) trio.run(double_sleep, 3) # does nothing for 6 seconds then returns
So it turns out our async_double
function is actually a bad example. I mean, it works, it's fine, there's nothing wrong with it, but it's pointless: it could just as easily be written as a regular function, and it would be more useful that way. double_sleep
is a much more typical example: we have to make it async, because it calls another async function. The end result is a kind of async sandwich, with Trio on both sides and our code in the middle:
trio.run -> double_sleep -> trio.sleep
This "sandwich" structure is typical for async code; in general, it looks like:
trio.run -> [async function] -> ... -> [async function] -> trio.whatever
It's exactly the functions on the path between trio.run
and trio.whatever
that have to be async. Trio provides the async bread, and then your code makes up the async sandwich's tasty async filling. Other functions (e.g., helpers you call along the way) should generally be regular, non-async functions.
Now would be a good time to open up a Python prompt and experiment a little with writing simple async functions and running them with trio.run
.
At some point in this process, you'll probably write some code like this, that tries to call an async function but leaves out the await
:
import time
import trio
async def broken_double_sleep(x):
print("*yawn* Going to sleep")
start_time = time.perf_counter()
# Whoops, we forgot the 'await'!
trio.sleep(2 * x)
sleep_time = time.perf_counter() - start_time
print("Woke up after {:.2f} seconds, feeling well rested!".format(sleep_time))
trio.run(broken_double_sleep, 3)
You might think that Python would raise an error here, like it does for other kinds of mistakes we sometimes make when calling a function. Like, if we forgot to pass trio.sleep
its required argument, then we would get a nice TypeError
saying so. But unfortunately, if you forget an await
, you don't get that. What you actually get is:
>>> trio.run(broken_double_sleep, 3)
*yawn* Going to sleep
Woke up after 0.00 seconds, feeling well rested!
__main__:4: RuntimeWarning: coroutine 'sleep' was never awaited
>>>
This is clearly broken – 0.00 seconds is not long enough to feel well rested! Yet the code acts like it succeeded – no exception was raised. The only clue that something went wrong is that it prints RuntimeWarning: coroutine 'sleep' was never awaited
. Also, the exact place where the warning is printed might vary, because it depends on the whims of the garbage collector. If you're using PyPy, you might not even get a warning at all until the next GC collection runs:
# On PyPy:
>>>> trio.run(broken_double_sleep, 3)
*yawn* Going to sleep
Woke up after 0.00 seconds, feeling well rested!
>>>> # what the ... ?? not even a warning!
>>>> # but forcing a garbage collection gives us a warning:
>>>> import gc
>>>> gc.collect()
/home/njs/pypy-3.8-nightly/lib-python/3/importlib/_bootstrap.py:191: RuntimeWarning: coroutine 'sleep' was never awaited
if _module_locks.get(name) is wr: # XXX PyPy fix?
0
>>>>
(If you can't see the warning above, try scrolling right.)
Forgetting an await
like this is an incredibly common mistake. You will mess this up. Everyone does. And Python will not help you as much as you'd hope 😞. The key thing to remember is: if you see the magic words RuntimeWarning: coroutine '...' was never awaited
, then this always means that you made the mistake of leaving out an await
somewhere, and you should ignore all the other error messages you see and go fix that first, because there's a good chance the other stuff is just collateral damage. I'm not even sure what all that other junk in the PyPy output is. Fortunately I don't need to know, I just need to fix my function!
("I thought you said you weren't going to mention coroutines!" Yes, well, I didn't mention coroutines, Python did. Take it up with Guido! But seriously, this is unfortunately a place where the internal implementation details do leak out a bit.)
Why does this happen? In Trio, every time we use await
it's to call an async function, and every time we call an async function we use await
. But Python's trying to keep its options open for other libraries that are ahem a little less organized about things. So while for our purposes we can think of await trio.sleep(...)
as a single piece of syntax, Python thinks of it as two things: first a function call that returns this weird "coroutine" object:
>>> trio.sleep(3)
<coroutine object sleep at 0x7f5ac77be6d0>
and then that object gets passed to await
, which actually runs the function. So if you forget await
, then two bad things happen: your function doesn't actually get called, and you get a "coroutine" object where you might have been expecting something else, like a number:
>>> async_double(3) + 1
TypeError: unsupported operand type(s) for +: 'coroutine' and 'int'
If you didn't already mess this up naturally, then give it a try on purpose: try writing some code with a missing await
, or an extra await
, and see what you get. This way you'll be prepared for when it happens to you for real.
And remember: watch out for RuntimeWarning: coroutine '...' was never awaited
; it means you need to find and fix your missing await
.
So now we've started using Trio, but so far all we've learned to do is write functions that print things and sleep for various lengths of time. Interesting enough, but we could just as easily have done that with time.sleep
. async/await
is useless!
Well, not really. Trio has one more trick up its sleeve, that makes async functions more powerful than regular functions: it can run multiple async functions at the same time. Here's an example:
tutorial/tasks-intro.py
There's a lot going on in here, so we'll take it one step at a time. In the first part, we define two async functions child1
and child2
. These should look familiar from the last section:
tutorial/tasks-intro.py
Next, we define parent
as an async function that's going to call child1
and child2
concurrently:
tutorial/tasks-intro.py
It does this by using a mysterious async with
statement to create a "nursery", and then "spawns" child1
and child2
into the nursery.
Let's start with this async with
thing. It's actually pretty simple. In regular Python, a statement like with someobj: ...
instructs the interpreter to call someobj.__enter__()
at the beginning of the block, and to call someobj.__exit__()
at the end of the block. We call someobj
a "context manager". An async with
does exactly the same thing, except that where a regular with
statement calls regular methods, an async with
statement calls async methods: at the start of the block it does await someobj.__aenter__()
and at that end of the block it does await someobj.__aexit__()
. In this case we call someobj
an "async context manager". So in short: with
blocks are a shorthand for calling some functions, and since with async/await Python now has two kinds of functions, it also needs two kinds of with
blocks. That's all there is to it! If you understand async functions, then you understand async with
.
Note
This example doesn't use them, but while we're here we might as well mention the one other piece of new syntax that async/await added: async for
. It's basically the same idea as async with
versus with
: An async for
loop is just like a for
loop, except that where a for
loop does iterator.__next__()
to fetch the next item, an async for
does await async_iterator.__anext__()
. Now you understand all of async/await. Basically just remember that it involves making sandwiches and sticking the word "async" in front of everything, and you'll do fine.
Now that we understand async with
, let's look at parent
again:
tutorial/tasks-intro.py
There are only 4 lines of code that really do anything here. On line 17, we use trio.open_nursery
to get a "nursery" object, and then inside the async with
block we call nursery.start_soon
twice, on lines 19 and 22. There are actually two ways to call an async function: the first one is the one we already saw, using await async_fn()
; the new one is nursery.start_soon(async_fn)
: it asks Trio to start running this async function, but then returns immediately without waiting for the function to finish. So after our two calls to nursery.start_soon
, child1
and child2
are now running in the background. And then at line 25, the commented line, we hit the end of the async with
block, and the nursery's __aexit__
function runs. What this does is force parent
to stop here and wait for all the children in the nursery to exit. This is why you have to use async with
to get a nursery: it gives us a way to make sure that the child calls can't run away and get lost. One reason this is important is that if there's a bug or other problem in one of the children, and it raises an exception, then it lets us propagate that exception into the parent; in many other frameworks, exceptions like this are just discarded. Trio never discards exceptions.
Ok! Let's try running it and see what we get:
parent: started!
parent: spawning child1...
parent: spawning child2...
parent: waiting for children to finish...
child2: started! sleeping now...
child1: started! sleeping now...
[... 1 second passes ...]
child1: exiting!
child2: exiting!
parent: all done!
(Your output might have the order of the "started" and/or "exiting" lines swapped compared to mine.)
Notice that child1
and child2
both start together and then both exit together. And, even though we made two calls to trio.sleep(1)
, the program finished in just one second total. So it looks like child1
and child2
really are running at the same time!
Now, if you're familiar with programming using threads, this might look familiar – and that's intentional. But it's important to realize that there are no threads here. All of this is happening in a single thread. To remind ourselves of this, we use slightly different terminology: instead of spawning two "threads", we say that we spawned two "tasks". There are two differences between tasks and threads: (1) many tasks can take turns running on a single thread, and (2) with threads, the Python interpreter/operating system can switch which thread is running whenever they feel like it; with tasks, we can only switch at certain designated places we call "checkpoints"
<checkpoints>
. In the next section, we'll dig into what this means.
The big idea behind async/await-based libraries like Trio is to run lots of tasks simultaneously on a single thread by switching between them at appropriate places – so for example, if we're implementing a web server, then one task could be sending an HTTP response at the same time as another task is waiting for new connections. If all you want to do is use Trio, then you don't need to understand all the nitty-gritty detail of how this switching works – but it's very useful to have at least a general intuition about what Trio is doing "under the hood" when your code is executing. To help build that intuition, let's look more closely at how Trio ran our example from the last section.
Fortunately, Trio provides a rich set of tools for inspecting
and debugging your programs <instrumentation>
. Here we want to watch trio.run
at work, which we can do by writing a class we'll call Tracer
, which implements Trio's ~trio.abc.Instrument
interface. Its job is to log various events as they happen:
tutorial/tasks-with-trace.py
Then we re-run our example program from the previous section, but this time we pass trio.run
a Tracer
object:
tutorial/tasks-with-trace.py
This generates a lot of output, so we'll go through it one step at a time.
First, there's a bit of chatter while Trio gets ready to run our code. Most of this is irrelevant to us for now, but in the middle you can see that Trio has created a task for the __main__.parent
function, and "scheduled" it (i.e., made a note that it should be run soon):
$ python3 tutorial/tasks-with-trace.py
!!! run started
### new task spawned: <init>
### task scheduled: <init>
### doing a quick check for I/O
### finished I/O check (took 1.1122087016701698e-05 seconds)
>>> about to run one step of task: <init>
### new task spawned: <call soon task>
### task scheduled: <call soon task>
### new task spawned: __main__.parent
### task scheduled: __main__.parent
<<< task step finished: <init>
### doing a quick check for I/O
### finished I/O check (took 6.4980704337358475e-06 seconds)
Once the initial housekeeping is done, Trio starts running the parent
function, and you can see parent
creating the two child tasks. Then it hits the end of the async with
block, and pauses:
>>> about to run one step of task: __main__.parent
parent: started!
parent: spawning child1...
### new task spawned: __main__.child1
### task scheduled: __main__.child1
parent: spawning child2...
### new task spawned: __main__.child2
### task scheduled: __main__.child2
parent: waiting for children to finish...
<<< task step finished: __main__.parent
Control then goes back to trio.run
, which logs a bit more internal chatter:
>>> about to run one step of task: <call soon task>
<<< task step finished: <call soon task>
### doing a quick check for I/O
### finished I/O check (took 5.476875230669975e-06 seconds)
And then gives the two child tasks a chance to run:
>>> about to run one step of task: __main__.child2
child2 started! sleeping now...
<<< task step finished: __main__.child2
>>> about to run one step of task: __main__.child1
child1: started! sleeping now...
<<< task step finished: __main__.child1
Each task runs until it hits the call to trio.sleep
, and then suddenly we're back in trio.run
deciding what to run next. How does this happen? The secret is that trio.run
and trio.sleep
work together to make it happen: trio.sleep
has access to some special magic that lets it pause its entire call stack, so it sends a note to trio.run
requesting to be woken again after 1 second, and then suspends the task. And once the task is suspended, Python gives control back to trio.run
, which decides what to do next. (If this sounds similar to the way that generators can suspend execution by doing a yield
, then that's not a coincidence: inside the Python interpreter, there's a lot of overlap between the implementation of generators and async functions.)
Note
You might wonder whether you can mix-and-match primitives from different async libraries. For example, could we use trio.run
together with asyncio.sleep
? The answer is no, we can't, and the paragraph above explains why: the two sides of our async sandwich have a private language they use to talk to each other, and different libraries use different languages. So if you try to call asyncio.sleep
from inside a trio.run
, then Trio will get very confused indeed and probably blow up in some dramatic way.
Only async functions have access to the special magic for suspending a task, so only async functions can cause the program to switch to a different task. What this means if a call doesn't have an await
on it, then you know that it can't be a place where your task will be suspended. This makes tasks much easier to reason about than threads, because there are far fewer ways that tasks can be interleaved with each other and stomp on each others' state. (For example, in Trio a statement like a += 1
is always atomic – even if a
is some arbitrarily complicated custom object!) Trio also makes some further guarantees beyond that <checkpoints>
, but that's the big one.
And now you also know why parent
had to use an async with
to open the nursery: if we had used a regular with
block, then it wouldn't have been able to pause at the end and wait for the children to finish; we need our cleanup function to be async, which is exactly what async with
gives us.
Now, back to our execution trace. To recap: at this point parent
is waiting on child1
and child2
, and both children are sleeping. So trio.run
checks its notes, and sees that there's nothing to be done until those sleeps finish – unless possibly some external I/O event comes in. If that happened, then it might give us something to do. Of course we aren't doing any I/O here so it won't happen, but in other situations it could. So next it calls an operating system primitive to put the whole process to sleep:
### waiting for I/O for up to 0.9999009938910604 seconds
And in fact no I/O does arrive, so one second later we wake up again, and Trio checks its notes again. At this point it checks the current time, compares it to the notes that trio.sleep
sent saying when the two child tasks should be woken up again, and realizes that they've slept for long enough, so it schedules them to run soon:
### finished I/O check (took 1.0006483688484877 seconds)
### task scheduled: __main__.child1
### task scheduled: __main__.child2
And then the children get to run, and this time they run to completion. Remember how parent
is waiting for them to finish? Notice how parent
gets scheduled when the first child exits:
>>> about to run one step of task: __main__.child1
child1: exiting!
### task scheduled: __main__.parent
### task exited: __main__.child1
<<< task step finished: __main__.child1
>>> about to run one step of task: __main__.child2
child2 exiting!
### task exited: __main__.child2
<<< task step finished: __main__.child2
Then, after another check for I/O, parent
wakes up. The nursery cleanup code notices that all its children have exited, and lets the nursery block finish. And then parent
makes a final print and exits:
### doing a quick check for I/O
### finished I/O check (took 9.045004844665527e-06 seconds)
>>> about to run one step of task: __main__.parent
parent: all done!
### task scheduled: <init>
### task exited: __main__.parent
<<< task step finished: __main__.parent
And finally, after a bit more internal bookkeeping, trio.run
exits too:
### doing a quick check for I/O
### finished I/O check (took 5.996786057949066e-06 seconds)
>>> about to run one step of task: <init>
### task scheduled: <call soon task>
### task scheduled: <init>
<<< task step finished: <init>
### doing a quick check for I/O
### finished I/O check (took 6.258022040128708e-06 seconds)
>>> about to run one step of task: <call soon task>
### task exited: <call soon task>
<<< task step finished: <call soon task>
>>> about to run one step of task: <init>
### task exited: <init>
<<< task step finished: <init>
!!! run finished
You made it!
That was a lot of text, but again, you don't need to understand everything here to use Trio – in fact, Trio goes to great lengths to make each task feel like it executes in a simple, linear way. (Just like your operating system goes to great lengths to make it feel like your single-threaded code executes in a simple linear way, even though under the covers the operating system juggles between different threads and processes in essentially the same way Trio does.) But it is useful to have a rough model in your head of how the code you write is actually executed, and – most importantly – the consequences of that for parallelism.
Alternatively, if this has just whetted your appetite and you want to know more about how async/await
works internally, then this blog post is a good deep dive, or check out this great walkthrough to see how to build a simple async I/O framework from the ground up.
Speaking of parallelism – let's zoom out for a moment and talk about how async/await compares to other ways of handling concurrency in Python.
As we've already noted, Trio tasks are conceptually rather similar to Python's built-in threads, as provided by the threading
module. And in all common Python implementations, threads have a famous limitation: the Global Interpreter Lock, or "GIL" for short. The GIL means that even if you use multiple threads, your code still (mostly) ends up running on a single core. People tend to find this frustrating.
But from Trio's point of view, the problem with the GIL isn't that it restricts parallelism. Of course it would be nice if Python had better options for taking advantage of multiple cores, but that's an extremely difficult problem to solve, and in the meantime there are lots of problems where a single core is totally adequate – or where if it isn't, then process-level or machine-level parallelism works fine.
No, the problem with the GIL is that it's a lousy deal: we give up on using multiple cores, and in exchange we get... almost all the same challenges and mind-bending bugs that come with real parallel programming, and – to add insult to injury – pretty poor scalability. Threads in Python just aren't that appealing.
Trio doesn't make your code run on multiple cores; in fact, as we saw above, it's baked into Trio's design that when it has multiple tasks, they take turns, so at each moment only one of them is actively running. We're not so much overcoming the GIL as embracing it. But if you're willing to accept that, plus a bit of extra work to put these new async
and await
keywords in the right places, then in exchange you get:
- Excellent scalability: Trio can run 10,000+ tasks simultaneously without breaking a sweat, so long as their total CPU demands don't exceed what a single core can provide. (This is common in, for example, network servers that have lots of clients connected, but only a few active at any given time.)
- Fancy features: most threading systems are implemented in C and restricted to whatever features the operating system provides. In Trio our logic is all in Python, which makes it possible to implement powerful and ergonomic features like
Trio's cancellation system <cancellation>
. - Code that's easier to reason about: the
await
keyword means that potential task-switching points are explicitly marked within each function. This can make Trio code dramatically easier to reason about than the equivalent program using threads.
Certainly it's not appropriate for every app... but there are a lot of situations where the trade-offs here look pretty appealing.
There is one downside that's important to keep in mind, though. Making checkpoints explicit gives you more control over how your tasks can be interleaved – but with great power comes great responsibility. With threads, the runtime environment is responsible for making sure that each thread gets its fair share of running time. With Trio, if some task runs off and does stuff for seconds on end without executing a checkpoint, then... all your other tasks will just have to wait.
Here's an example of how this can go wrong. Take our example
from above <tutorial-example-tasks-intro>
, and replace the calls to trio.sleep
with calls to time.sleep
. If we run our modified program, we'll see something like:
parent: started!
parent: spawning child1...
parent: spawning child2...
parent: waiting for children to finish...
child2 started! sleeping now...
[... pauses for 1 second ...]
child2 exiting!
child1: started! sleeping now...
[... pauses for 1 second ...]
child1: exiting!
parent: all done!
One of the major reasons why Trio has such a rich instrumentation API <tutorial-instrument-example>
is to make it possible to write debugging tools to catch issues like this.
Now let's take what we've learned and use it to do some I/O, which is where async/await really shines.
The traditional application for demonstrating network APIs is an "echo server": a program that accepts arbitrary data from a client, and then sends that same data right back. (Probably a more relevant example these days would be an application that does lots of concurrent HTTP requests, but for that you need an HTTP library such as asks, so we'll stick with the echo server tradition.)
To start with, here's an example echo client, i.e., the program that will send some data at our echo server and get responses back:
tutorial/echo-client.py
The overall structure here should be familiar, because it's just like our last example <tutorial-example-tasks-intro>
: we have a parent task, which spawns two child tasks to do the actual work, and then at the end of the async with
block it switches into full-time parenting mode while waiting for them to finish. But now instead of just calling trio.sleep
, the children use some of Trio's networking APIs.
Let's look at the parent first:
tutorial/echo-client.py
First we call trio.open_tcp_stream
to make a TCP connection to the server. 127.0.0.1
is a magic IP address meaning "the computer I'm running on", so this connects us to whatever program on the local computer is using PORT
as its contact point. This function returns an object implementing Trio's ~trio.abc.Stream
interface, which gives us methods to send and receive bytes, and to close the connection when we're done. We use an async with
block to make sure that we do close the connection – not a big deal in a toy example like this, but it's a good habit to get into, and Trio is designed to make with
and async with
blocks easy to use.
Finally, we start up two child tasks, and pass each of them a reference to the stream. (This is also a good example of how nursery.start_soon
lets you pass positional arguments to the spawned function.)
Our first task's job is to send data to the server:
tutorial/echo-client.py
It uses a loop that alternates between calling await client_stream.send_all(...)
to send some data (this is the method you use for sending data on any kind of Trio stream), and then sleeping for a second to avoid making the output scroll by too fast on your terminal.
And the second task's job is to process the data the server sends back:
tutorial/echo-client.py
It uses an async for
loop to fetch data from the server. Alternatively, it could use ~trio.abc.ReceiveStream.receive_some, which is the opposite of ~trio.abc.SendStream.send_all, but using async for
saves some boilerplate.
And now we're ready to look at the server.
As usual, let's look at the whole thing first, and then we'll discuss the pieces:
tutorial/echo-server.py
Let's start with main
, which is just one line long:
tutorial/echo-server.py
What this does is call serve_tcp
, which is a convenience function Trio provides that runs forever (or at least until you hit control-C or otherwise cancel it). This function does several helpful things:
- It creates a nursery internally, so that our server will be able to handle multiple connections at the same time.
- It listens for incoming TCP connections on the specified
PORT
. - Whenever a connection arrives, it starts a new task running the function we pass (in this example it's
echo_server
), and passes it a stream representing that connection. - When each task exits, it makes sure to close the corresponding connection. (That's why you don't see any
async with server_stream
in the server –serve_tcp
takes care of this for us.)
So serve_tcp
is pretty handy! This part works pretty much the same for any server, whether it's an echo server, HTTP server, SSH server, or whatever, so it makes sense to bundle it all up together in a helper function like this.
Now let's look at echo_server
, which handles each client connection – so if there are multiple clients, there might be multiple calls to echo_server
running at the same time. This is where we implement our server's "echo" behavior. This should be pretty straightforward to understand, because it uses the same stream functions we saw in the last section:
tutorial/echo-server.py
The argument server_stream
is provided by serve_tcp
, and is the other end of the connection we made in the client: so the data that the client passes to send_all
will come out here. Then we have a try
block discussed below, and finally the server loop which alternates between reading some data from the socket and then sending it back out again (unless the socket was closed, in which case we quit).
So what's that try
block for? Remember that in Trio, like Python in general, exceptions keep propagating until they're caught. Here we think it's plausible there might be unexpected exceptions, and we want to isolate that to making just this one task crash, without taking down the whole program. For example, if the client closes the connection at the wrong moment then it's possible this code will end up calling send_all
on a closed connection and get a BrokenResourceError
; that's unfortunate, and in a more serious program we might want to handle it more explicitly, but it doesn't indicate a problem for any other connections. On the other hand, if the exception is something like a KeyboardInterrupt
, we do want that to propagate out into the parent task and cause the whole program to exit. To express this, we use a try
block with an except Exception:
handler.
In general, Trio leaves it up to you to decide whether and how you want to handle exceptions, just like Python in general.
Open a few terminals, run echo-server.py
in one, run echo-client.py
in another, and watch the messages scroll by! When you get bored, you can exit by hitting control-C.
Some things to try:
- Open several terminals, and run multiple clients at the same time, all talking to the same server.
- See how the server reacts when you hit control-C on the client.
- See how the client reacts when you hit control-C on the server.
Here's a question you might be wondering about: why does our client use two separate tasks for sending and receiving, instead of a single task that alternates between them – like the server has? For example, our client could use a single task like:
# Can you spot the two problems with this code?
async def send_and_receive(client_stream):
while True:
data = ...
await client_stream.send_all(data)
received = await client_stream.receive_some()
if not received:
sys.exit()
await trio.sleep(1)
It turns out there are two problems with this – one minor and one major. Both relate to flow control. The minor problem is that when we call receive_some
here we're not waiting for all the data to be available; receive_some
returns as soon as any data is available. If data
is small, then our operating systems / network / server will probably keep it all together in a single chunk, but there's no guarantee. If the server sends hello
then we might get hello
, or hel
lo
, or h
e
l
l
o
, or ... bottom line, any time we're expecting more than one byte of data, we have to be prepared to call receive_some
multiple times.
And where this would go especially wrong is if we find ourselves in the situation where data
is big enough that it passes some internal threshold, and the operating system or network decide to always break it up into multiple pieces. Now on each pass through the loop, we send len(data)
bytes, but read less than that. The result is something like a memory leak: we'll end up with more and more data backed up in the network, until eventually something breaks.
Note
If you're curious how things break, then you can use ~trio.abc.ReceiveStream.receive_some's optional argument to put a limit on how many bytes you read each time, and see what happens.
We could fix this by keeping track of how much data we're expecting at each moment, and then keep calling receive_some
until we get it all:
expected = len(data)
while expected > 0:
received = await client_stream.receive_some(expected)
if not received:
sys.exit(1)
expected -= len(received)
This is a bit cumbersome, but it would solve this problem.
There's another problem, though, that's deeper. We're still alternating between sending and receiving. Notice that when we send data, we use await
: this means that sending can potentially block. Why does this happen? Any data that we send goes first into an operating system buffer, and from there onto the network, and then another operating system buffer on the receiving computer, before the receiving program finally calls receive_some
to take the data out of these buffers. If we call send_all
with a small amount of data, then it goes into these buffers and send_all
returns immediately. But if we send enough data fast enough, eventually the buffers fill up, and send_all
will block until the remote side calls receive_some
and frees up some space.
Now let's think about this from the server's point of view. Each time it calls receive_some
, it gets some data that it needs to send back. And until it sends it back, the data is sitting around takes up memory. Computers have finite amounts of RAM, so if our server is well behaved then at some point it needs to stop calling receive_some
until it gets rid of some of the old data by doing its own call to send_all
. So for the server, really the only viable option is to alternate between receiving and sending.
But we need to remember that it's not just the client's call to send_all
that might block: the server's call to send_all
can also get into a situation where it blocks until the client calls receive_some
. So if the server is waiting for send_all
to finish before it calls receive_some
, and our client also waits for send_all
to finish before it calls receive_some
,... we have a problem! The client won't call receive_some
until the server has called receive_some
, and the server won't call receive_some
until the client has called receive_some
. If our client is written to alternate between sending and receiving, and the chunk of data it's trying to send is large enough (e.g. 10 megabytes will probably do it in most configurations), then the two processes will deadlock.
Moral: Trio gives you powerful tools to manage sequential and concurrent execution. In this example we saw that the server needs send
and receive_some
to alternate in sequence, while the client needs them to run concurrently, and both were straightforward to implement. But when you're implementing network code like this then it's important to think carefully about flow control and buffering, because it's up to you to choose the right execution mode!
Other popular async libraries like Twisted and asyncio
tend to paper over these kinds of issues by throwing in unbounded buffers everywhere. This can avoid deadlocks, but can introduce its own problems and in particular can make it difficult to keep memory usage and latency under control. While both approaches have their advantages, Trio takes the position that it's better to expose the underlying problem as directly as possible and provide good tools to confront it head-on.
Note
If you want to try and make the deadlock happen on purpose to see for yourself, and you're using Windows, then you might need to split the send_all
call up into two calls that each send half of the data. This is because Windows has a somewhat unusual way of handling buffering.
TODO: give an example using fail_after
TODO: explain Cancelled
TODO: explain how cancellation is also used when one child raises an exception
TODO: show an example MultiError
traceback and walk through its structure
TODO: maybe a brief discussion of KeyboardInterrupt
handling?