-
-
Notifications
You must be signed in to change notification settings - Fork 343
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Document strict ExceptionGroups as the natural default #2984
Changes from 1 commit
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
Original file line number | Diff line number | Diff line change |
---|---|---|
|
@@ -639,8 +639,7 @@ crucial things to keep in mind: | |
* The nursery is marked as "closed", meaning that no new tasks can | ||
be started inside it. | ||
|
||
* Any unhandled exceptions are re-raised inside the parent task. If | ||
there are multiple exceptions, then they're collected up into a | ||
* Any unhandled exceptions are re-raised inside the parent task, grouped into a | ||
single :exc:`BaseExceptionGroup` or :exc:`ExceptionGroup` exception. | ||
|
||
Since all tasks are descendents of the initial task, one consequence | ||
|
@@ -733,9 +732,9 @@ If you want to reraise exceptions, or raise new ones, you can do so, but be awar | |
exceptions raised in ``except*`` sections will be raised together in a new exception | ||
group. | ||
|
||
But what if you can't use ``except*`` just yet? Well, for that there is the handy | ||
exceptiongroup_ library which lets you approximate this behavior with exception handler | ||
callbacks:: | ||
But what if you can't use Python 3.11, and therefore ``except*``, just yet? | ||
The same exceptiongroup_ library which backports `ExceptionGroup` also lets | ||
you approximate this behavior with exception handler callbacks:: | ||
|
||
from exceptiongroup import catch | ||
|
||
|
@@ -768,43 +767,91 @@ inside the handler function(s) with the ``nonlocal`` keyword:: | |
async with trio.open_nursery() as nursery: | ||
nursery.start_soon(broken1) | ||
|
||
.. _strict_exception_groups: | ||
.. _handling_exception_groups: | ||
|
||
Designing for multiple errors | ||
+++++++++++++++++++++++++++++ | ||
|
||
Structured concurrency is still a young design pattern, but there are a few patterns | ||
we've identified for how you (or your users) might want to handle groups of exceptions. | ||
Note that the final pattern, simply raising an `ExceptionGroup`, is the most common - | ||
and nurseries automatically do that for you. | ||
|
||
**First**, you might want to 'defer to' a particular exception type, raising just that if | ||
there is any such instance in the group. For example: `KeyboardInterrupt` has a clear | ||
meaning for the surrounding code, could reasonably take priority over errors of other | ||
types, and whether you have one or several of them doesn't really matter. | ||
|
||
This pattern can often be implemented using a decorator or a context manager, such | ||
as :func:`trio_util.multi_error_defer_to` or :func:`trio_util.defer_to_cancelled`. | ||
Note however that re-raising a 'leaf' exception will discard whatever part of the | ||
traceback is attached to the `ExceptionGroup` itself, so we don't recommend this for | ||
errors that will be presented to humans. | ||
|
||
"Strict" versus "loose" ExceptionGroup semantics | ||
++++++++++++++++++++++++++++++++++++++++++++++++ | ||
.. | ||
TODO: what about `Cancelled`? It's relevantly similar to `KeyboardInterrupt`, | ||
but if you have multiple Cancelleds destined for different scopes, it seems | ||
like it might be bad to abandon all-but-one of those - we might try to execute | ||
some more code which then itself gets cancelled again, and incur more cleanup. | ||
That's only a mild inefficiency though, and the semantics are fine overall. | ||
**Second**, you might want to treat the concurrency inside your code as an implementation | ||
detail which is hidden from your users - for example, abstracting a protocol which | ||
involves sending and receiving data to a simple receive-only interface, or implementing | ||
a context manager which maintains some background tasks for the length of the | ||
``async with`` block. | ||
|
||
The simple option here is to ``raise MySpecificError from group``, allowing users to | ||
handle your library-specific error. This is simple and reliable, but doesn't completely | ||
hide the nursery. *Do not* unwrap single exceptions if there could ever be multiple | ||
Comment on lines
+805
to
+806
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. it what way does it not hide the nursery? There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. # end-user code
with library_function_wrapping_a_nursery():
1/0 raises |
||
exceptions though; that always ends in latent bugs and then tears. | ||
|
||
The more complex option is to ensure that only one exception can in fact happen at a time. | ||
This is *very hard*, for example you'll need to handle `KeyboardInterrupt` somehow, and | ||
we strongly recommend having a ``raise PleaseReportBug from group`` fallback just in case | ||
you get a group containing more than one exception. | ||
This is useful when writing a context manager which starts some background tasks, and then | ||
yields to user code which effectively runs 'inline' in the body of the nursery block. | ||
In this case, the background tasks can be wrapped with e.g. the `outcome | ||
<https://pypi.org/project/outcome/>`__ library to ensure that only one exception | ||
can be raised (from end-user code); and then you can either ``raise SomeInternalError`` | ||
if a background task failed, or unwrap the user exception if that was the only error. | ||
|
||
.. | ||
TODO: rewrite this (and possible other) sections from the new strict-by-default perspective, under the heading "Deprecated: non-strict ExceptionGroups" - to explain that it only exists for backwards-compatibility, will be removed in future, and that we recommend against it for all new code. | ||
Ideally, in some abstract sense we'd want everything that *can* raise an | ||
`ExceptionGroup` to *always* raise an `ExceptionGroup` (rather than, say, a single | ||
`ValueError`). Otherwise, it would be easy to accidentally write something like ``except | ||
ValueError:`` (not ``except*``), which works if a single exception is raised but fails to | ||
catch _anything_ in the case of multiple simultaneous exceptions (even if one of them is | ||
a ValueError). However, this is not how Trio worked in the past: as a concession to | ||
practicality when the ``except*`` syntax hadn't been dreamed up yet, the old | ||
``trio.MultiError`` was raised only when at least two exceptions occurred | ||
simultaneously. Adding a layer of `ExceptionGroup` around every nursery, while | ||
theoretically appealing, would probably break a lot of existing code in practice. | ||
|
||
Therefore, we've chosen to gate the newer, "stricter" behavior behind a parameter | ||
called ``strict_exception_groups``. This is accepted as a parameter to | ||
:func:`open_nursery`, to set the behavior for that nursery, and to :func:`trio.run`, | ||
to set the default behavior for any nursery in your program that doesn't override it. | ||
|
||
* With ``strict_exception_groups=True``, the exception(s) coming out of a nursery will | ||
always be wrapped in an `ExceptionGroup`, so you'll know that if you're handling | ||
single errors correctly, multiple simultaneous errors will work as well. | ||
|
||
* With ``strict_exception_groups=False``, a nursery in which only one task has failed | ||
will raise that task's exception without an additional layer of `ExceptionGroup` | ||
wrapping, so you'll get maximum compatibility with code that was written to | ||
support older versions of Trio. | ||
|
||
The default is set to ``strict_exception_groups=True``, in line with the default behaviour | ||
of ``TaskGroup`` in asyncio and anyio. We've also found that non-strict mode makes it | ||
too easy to neglect the possibility of several exceptions being raised concurrently, | ||
causing nasty latent bugs when errors occur under load. | ||
For more on this pattern, see https://github.com/python-trio/trio/issues/2929 | ||
and the linked issue on trio-websocket. We may want to provide a nursery mode | ||
which handles this automatically; it's annoying but not too complicated and | ||
seems like it might be a good feature to ship for such cases. | ||
**Third and most often**, the existence of a nursery in your code is not just an | ||
implementation detail, and callers *should* be prepared to handle multiple exceptions | ||
in the form of an `ExceptionGroup`, whether with ``except*`` or manual inspection | ||
or by just letting it propagate to *their* callers. Because this is so common, | ||
it's nurseries' default behavior and you don't need to do anything. | ||
Comment on lines
+826
to
+830
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. most often -> most common But if we're talking about an API, I don't think it's desirable for this to be the most common. I'm afraid it will become the most common though, because it takes a lot of work to do otherwise. I'd guess that an There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. For many functions, internal concurrency is part of the documented semantics - for example it's not just an implementation detail that I agree it's a fuzzy boundary though, and I'll remove the most-common claim. I don't think the burden is particularly heavy once you're on Python 3.11+ and can use There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. No doubt there are API's that need to expose concurrency, though we might disagree on how common that is. A few burden points I can think of:
|
||
|
||
.. _strict_exception_groups: | ||
|
||
Historical Note: "non-strict" ExceptionGroups | ||
+++++++++++++++++++++++++++++++++++++++++++++ | ||
|
||
In early versions of Trio, the ``except*`` syntax hadn't be dreamt up yet, and we | ||
hadn't worked with structured concurrency for long or in large codebases. | ||
As a concession to convenience, some APIs would therefore raise single exceptions, | ||
and only wrap concurrent exceptions in the old ``trio.MultiError`` type if there | ||
were two or more. | ||
|
||
Unfortunately, the results were not good: calling code often didn't realize that | ||
some function *could* raise a ``MultiError``, and therefore handle only the common | ||
case - with the result that things would work well in testing, and then crash under | ||
heavier load (typically in production). `asyncio.TaskGroup` learned from this | ||
experience and *always* wraps errors into an `ExceptionGroup`, as does ``anyio``, | ||
and as of Trio 0.25 that's our default behavior too. | ||
|
||
We currently support a compatibility argument ``strict_exception_groups=False`` to | ||
`trio.run` and `trio.open_nursery`, which restores the old behavior (although | ||
``MultiError`` itself has been fully removed). We strongly advise against it for | ||
new code, and encourage existing uses to migrate - we consider the option deprecated | ||
and plan to remove it after a period of documented and then runtime warnings. | ||
A5rocks marked this conversation as resolved.
Show resolved
Hide resolved
|
||
|
||
.. _exceptiongroup: https://pypi.org/project/exceptiongroup/ | ||
|
||
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
(unfortunately trio-util itself is not caught up with the MultiError/ExceptionGroup transition)