Ideas for linting trio programs

@wgwz

I haven't given up on catching missing awaits at runtime (see #495), but another strategy that could be useful is to have a linter/static analyzer that catches them for you. In fact, you really want both, for all the reasons runtime + static checks are useful for other kinds of errors: runtime checks can be 100% reliable when the error actually happens, and are especially useful for beginners who can't be expected to set up fancy linting tools; static checks can catch mistakes early, even in untested code-paths.

Examples of this being a real issue:

Update VSTS to Azure DevOps and ignore issue with Heroku packages python/cpython#9168 (comment)
@wgwz gets bitten: https://gitter.im/python-trio/general?at=5b9dde717dca30650309ca3e
Yury says that specifically when a function is alone on a line, he accidentally writes foo() instead of await foo() about once a month.

I know absolutely nothing about static analysis, so I have no idea how to actually implement a lint for this, but let's have an issue to discuss and maybe we'll figure it out together.

In theory, this should be pretty straightforward, because in Trio, the property to enforce is: any function call which returns an awaitable, should have an await to immediately consume that awaitable. This makes things way simpler than asyncio, where it's common to create coroutine objects and then pass them somewhere else, and sometimes this is what you want (passing them into asyncio.run or asyncio.create_task or asyncio.gather) and other times its an accident because you forgot an await. So a lint for asyncio would require some pretty sophisticated data flow inference and knowledge of different APIs; for trio, once we've figured out which calls are to async functions/methods, the check itself requires only superficial knowledge of the surface AST. OTOH, this may make it more difficult to convince tools like mypy or pylint to include our check, because it won't work for asyncio users; and, if they do include it there will need to be some kind of configuration to turn it on or off. Maybe it would need to be a plugin, at least to start out.

I got a bit nerd-sniped by this last night. Findings so far:

pylint

Pylint has docs on how to add a new check, and an official plugin API, so we could distribute a checker as a third-party package. And, (through astroid), it has some reasonably sophisticated APIs to go from a call back to the definition of the function/method being called. So playing around, I came up with this: https://gist.github.com/njsmith/0ce8c661684ba4514b9ea4aac6182f3c

But unfortunately, it turns out that figuring out which functions return awaitables is a bit more complicated than it seems. In particular, the code in that gist can (I think) catch many cases where there's a call to a function, there's no await at the call site, and the function is defined with async def. However, this is still wrong in at least two cases:

It ignores @ decorators on the function. These can be pulled out of the AsyncFunctionDef by accessing its .decorators attribute, but... then what. (Maybe if we special-case @{en,dis}able_ki_protection and then otherwise ignore async defs with decorators, that's still enough to be useful?)
It doesn't distinguish between async functions and async generators. This is important, because both are defined using async def syntax, but async functions must be called with await, and async generators must not be called with await. We could do this by hand, though it's a bit tiresome: we'd need to walk the AST for the function body, looking for yield statements, except being careful not to look inside any nested functions.

So this is probably doable, but it was looking complicated enough that I paused to look at mypy instead.

mypy

In theory, mypy should solve this problem nicely, because it already has a stupendously sophisticated type inference engine that knows which calls return awaitables and which ones don't. So, we should be able to skip all the complications that we ran into with pylint, and just go straight to implementing our check. And in theory, this can even be more accurate, because it can take into account explicit type hints in cases where static analysis alone isn't enough.

Mypy also has a plugin API, but it's completely undocumented. That said, it looks like it is possible to ship a third-party plugin and then import it through the mypy config file by using the undocumented plugins = ... option in the config file, which contains a comma-separated list of setuptools-style entry points pointing to plugin objects.

AFAICT, though, currently the plugin API only allows you to define custom type inference rules (e.g., inferring the return value from open by checking whether the mode string has a b in it). I don't see any way to define a custom type checking rule.

It looks like the overall flow is mypy.main:main → mypy.build:build, whose BuildManager object together with dispatch runs 3 semantic analyzer passes (semanal_pass1.py, semanal.py, semanal_pass3.py) and then 2 passes of type checking which are performed using a mypy.checker:TypeChecker object. So I guess we'd want to hook in somewhere in those two passes?

Others?

Someone in the python/typing gitter claimed that pyre will have a check for this soon. Given that pyre is written by facebook and facebook uses asyncio, I suspect anything they come up with will be much cleverer than we need; and given that it's written in ocaml, I doubt we can easily write a plugin in python :-). Also they currently only support running checks on macOS and Linux, and I've never heard of it before now. So probably not what most of our users will be using?

Anything else?

I don't see any way to define a custom type checking rule.

We may not need a custom type checker; as soon as the Awaitable is processed further it'll be reported as an error by the regular mypy rules. (Assuming that the rest of your program is typed appropriately.)

I do hope that python/mypy#2499 will get some traction, though; it's old enough. :-P

as soon as the Awaitable is processed further it'll be reported as an error by the regular mypy rules.

As far as I know, none of the three bullet point examples above would be caught by mypy regularly. Two involve functions with no return value called for side effects, and in the other the return value is formatted intoa string, which isa legal operation on coroutine objects.

I'm thinking about how this policy will affect API's.

It's going to be common to have utility functions to compose awaitables. So e.g. an wait_all() would have to be:

await wait_all(some_event.wait, some_async_foo)

rather than

await wait_all(some_event.wait(), some_async_foo())

@belm0

@belm0 yep. Just like how trio already does it, and also all non-async python code :-)

So if the wait functions have parameters, we need to resort to partial, correct?

E.g. I have a ValueEvent which takes lambda etc. on wait.

await wait_all(partial(event.wait_value, lambda x: x > 10), event2.wait, ...)

Correct.

I'm seeing useful standard library functions which take raw coroutines:

https://docs.python.org/3/library/contextlib.html#contextlib.AsyncExitStack

@belm0

Those docs are being lazy and saying "coroutine", without clarifying whether they mean a "coroutine function" or a "coroutine object". This is one reason I always say "async function" or "coroutine object". (Also, "async function" is just a clearer name that avoids obscure jargon.)

If you look at the source, it turns out that AsyncExitStack's methods actually take "coroutine functions", i.e. you pass in an async function, not a coroutine object, so they follow Trio-style, not asyncio-style.

Someone should probably submit an issue or PR to clarify AsyncExitStack docs. [Edit: done, so @belm0's comment above will stop making sense soon :-)]

@njsmith

This still bites me on occasion. In my experience, the existing runtime check (the "coroutine was never awaited" warning) has always been good enough to spot the issue once I encounter it. Additional runtime checks might be good but I don't think there is much to be gained. On the other hand I would love a static warning!

I was planning to play around a bit with pylint to see what I could achieve and wow, @njsmith is always many steps ahead :) In my case, a best-effort pylint solution would already be a huge win. Pylint is already part of my workflow and I haven't looked into mypy yet.

Issue 1 (cannot detect decorated functions or other callable-then-awaitable patterns) is a limitation of course, but one I can live with for now.

Issue 2 (difference between async gens and async funcs): wouldn't it be enough for now to expect an async-something to be either awaited or async-iterated?

@sorcio

In my case, a best-effort pylint solution would already be a huge win. Pylint is already part of my workflow and I haven't looked into mypy yet.

There's no reason we have to pick one, either. If people care enough about pylint to make it work with pylint, then cool, pylint users get the benefit; it doesn't stop us from adding a check to mypy too, for mypy users.

Issue 1 (cannot detect decorated functions or other callable-then-awaitable patterns) is a limitation of course, but one I can live with for now.

Agreed.

Issue 2 (difference between async gens and async funcs): wouldn't it be enough for now to expect an async-something to be either awaited or async-iterated?

It's not always true that async generators get immediately iterated, e.g.:

async with aclosing(agen_fn()) as agen:
    async for ... in agen:
        ...

But also I suspect it's about as easy to implement a check for whether a function is an async generator, as it is to implement a check for whether it's being used in an async for. It's not hard, just annoying :-)

I guess something like:

class IsThisAnAsyncGenerator(ast.NodeVisitor):
    its_an_async_generator = False

    def visit_Yield(self, *args, **kwargs):
        self.its_an_async_generator = True

    # Can't happen on 3.7, but future-proofing in case async generators ever start supporting 'yield from'
    def visit_YieldFrom(self, *args, **kwargs):
        self.its_an_async_generator = True

    def visit_FunctionDef(self, *args, **kwargs):
        # Don't recurse into nested functions
        pass

    def visit_AsyncFunctionDef(self, *args, **kwargs):
        # Don't recurse into nested async functions either
        pass

def is_this_asyncfunctiondef_node_an_async_generator(node):
    visitor = IsThisAnAsyncGenerator()
    visitor.visit(node)
    return visitor.its_an_async_generator

(See: ast.NodeVisitor, list of node types.)

@sorcio any interest in collecting this all up into an installable pytest plugin and writing some tests?

Oh, yes, this is interesting. I don't have experience with pytest plugins and I'm on vacation for a couple weeks. I will see into it later on :) Thank you very much for the comments!

There's no reason we have to pick one, either. If people care enough about pylint to make it work with pylint, then cool, pylint users get the benefit; it doesn't stop us from adding a check to mypy too, for mypy users.

+1, that's what I meant as well.

I tried the pylint work in progress but it's giving me a false positive at every use of await in my codebase. What am I missing?

It's not even work-in-progress, more like thinking-with-an-editor-buffer-open. I've never even tried running it, so... I guess there's a bug ¯\(ツ)/¯

update on pending pylint missing-await checker (#700):

it's run against the entire trio codebase along with a few builtin checkers related to async/await
it properly ignores async generators
it has automated tests (I wrote a simple pytest test framework)

From chat:

# What you have:
async with trio.open_nursery() as nursery:
    async with await trio.open_process(hunspell_cmd, **proc_kwargs) as proc:
        nursery.start_soon(send_hunspell, proc, ['McLaughlin', 'SecureStorage'])
        nursery.start_soon(recv_hunspell, proc)
    # Process is closed here, after starting the two tasks but before they finish
# Then we wait for tasks to finish
        
# What you want:
async with await trio.open_process(hunspell_cmd, **proc_kwargs) as proc:
    async with trio.open_nursery() as nursery:
        nursery.start_soon(send_hunspell, proc, ['McLaughlin', 'SecureStorage'])
        nursery.start_soon(recv_hunspell, proc)
    # First we wait for tasks to finish
# Then the process is closed

Could a linter catch instances where a process or other task is closed before it finishes?

Good point! I guess the general case of detecting when something is closed while in use would require some super sophisticated data-flow analysis, but this specific pattern of async with x around a call to start_soon(..., x) seems like it could be simple enough for a linter to detect + a common enough error to be a useful lint.

Though we might want to special-case async with open_nursery() as nursery: nursery.start_soon(..., nursery), since that's not an error. More generally, I guess the point is that we need to also have some heuristic to rule out cases where the async with surrounds the nursery.

We might also want to warn if we see a nursery body that contains exactly one statement, and that statement is a call to nursery.start_soon ("consider using a regular function call instead")

From chat:

anything that awaits inside a finally: clause needs shielding. I habitually use with trio.fail_after(2,shield=True). Otherwise they could swallow errors.

I've just created feature request for Pyright about forgotten await: microsoft/pyright#1317

Personally I use Pyright all the time while writing Python code in VSCode. It's works great, especially in interactive use.

@erictraut

@erictraut suggested Pyright could maybe implement a check for "Awaitable object that is dropped on the floor (not assigned to a variable, passed as an argument, returned from the function, etc.)".

It will solve two of the three cases mentioned in the description of this issue (trio.sleep and Yury's confession):

trio.sleep(1) # could warn about missing `await` because the result is not used in any way

But of course it won't warn about this example:

result = some_awaitable() # can't warn in general case

What do you think? Any potential downsides for the general case? Is there anything that could help Pyright catch even more cases? (And avoid the need for a config option or a plugin.)

@shamrin

@shamrin I replied on the pyright issue.

Pyright author has just implemented reportUnusedCoroutine rule that makes it an error to write alone-on-a-line foo() when it should be await foo(). The rule will be enabled by default in all modes.

microsoft/pyright@fb4a4e9

Now that pyright reports alone-on-a-line missing await, the only remaining¹ real-life problem (that the type checker can't catch) is implicit string conversion:

async def getone(): return 1
name = getone()  # should have been `name = await getone()`
print(f'one is {name}')

$ python3 ha.py
one is <coroutine object getone at 0x1012b0e40>
sys:1: RuntimeWarning: coroutine 'getone' was never awaited

(I assume this is what broken cpython bot did: python/cpython#9168 (comment).)

How do we catch it?

Of course no checker should forbid a very common print(f'debug: {any_object}')². However, a warning could help. It could warn about implicit conversion to string for certain cases. The warning could be silenced by calling repr() or by special inline comment. Would this warning be too drastic?

¹ At least it's the only one I know. Have you seen other real-life examples of forgotten await?
² Starting from Python 3.9, print(f'debug: {any_object=}' would be better. No need to warn about this obviously.

Closing this because typecheckers and flake8-trio are going pretty well these days - I'll even be speaking about it at PyCon 🙂

Navigation Menu

Repository navigation

Ideas for linting trio programs #671

pylint

mypy

Others?

10 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Ideas for linting trio programs #671

Description

pylint

mypy

Others?

Activity

m-u-xyz commented on Sep 17, 2018

njsmith commented on Sep 17, 2018

belm0 commented on Sep 18, 2018

njsmith commented on Sep 18, 2018

belm0 commented on Sep 18, 2018

njsmith commented on Sep 18, 2018

belm0 commented on Sep 18, 2018

njsmith commented on Sep 18, 2018

sorcio commented on Sep 18, 2018

njsmith commented on Sep 19, 2018

sorcio commented on Sep 21, 2018

belm0 commented on Sep 29, 2018

njsmith commented on Sep 29, 2018

10 remaining items

belm0 commented on Apr 14, 2019

jtrakk commented on Sep 23, 2019

njsmith commented on Sep 23, 2019

njsmith commented on Sep 23, 2019

njsmith commented on Sep 23, 2019

jtrakk commented on Jan 27, 2020

shamrin commented on Dec 26, 2020

shamrin commented on Dec 27, 2020

njsmith commented on Dec 27, 2020

shamrin commented on Dec 27, 2020

shamrin commented on Jan 5, 2021

Zac-HD commented on Mar 17, 2023

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions