OK, better late than never... here's a much-delayed review of the PEP. Thank you Irit and Guido for carrying this forward while I've been AWOL! It's fantastic to see my old design sketches turned into something like, actually real.

== Overall feelings ==

Honestly, I have somewhat mixed feelings ExceptionGroups. I don't see any way around adding ExceptionGroups in some form, because it's just a fact of life that in a concurrent program, multiple things can go wrong at once, and we want Python to be usable for writing concurrent programs. Right now the state of the art is "exceptions in background threads/tasks get dropped on the floor", and almost anything is better than that. The current PEP is definitely better than that. But at the same time, there are a lot of compromises needed to retrofit this onto Python's existing system, and the current proposal feels like a bunch of awkward hacks with hacks on top. That's largely my fault for not thinking of something better, and maybe there is nothing better. But I still wish we could come up with something more elegant, and I do see why this proposal has made people uncomfortable. For example:

- I'm uncomfortable with how in some contexts we treat EG's as placeholders for the contained exceptions, and other places we treat them like a single first-class exceptions. (Witness all the feedback about "why not just catch the ExceptionGroup and handle it by hand?", and imagine writing the docs explaining all the situations where that is or isn't a good idea and the pitfalls involved...) If we could somehow pick one and stick to it then I think that would make it easier for users to grasp. (My gut feeling is that making them pure containers is better, which to me is why it makes sense for them to be @final and why I keep hoping we can figure out some better way for plain 'except' and EGs to interact.)

- If a function wants to start using concurrency internally, then now *all* its exceptions have to get wrapped in EGs and callers have to change *all* their exception handling code to use except* or similar. You would think this was an internal implementation detail that the caller shouldn't have to care about, but instead it forces a major change on the function's public API. And this is because regular 'except' can't do anything useful with EGs.

- We have a special-case hack to keep 'except Exception' working, but it has tricky edge cases (Exceptions can still sneak past if they're paired up with a BaseException), and it really is specific to 'except Exception'; it doesn't work for any other 'except SomeError' code. This smells funny.

Anyway, that's just abstract context to give an idea where I'm coming from. Maybe we just have to accept these trade-offs, but if anyone has any ideas, speak up...

== Most important comment ==

Flat ExceptionGroups: there were two basic design approaches we discussed last year, which I'll call "flat" vs "nested". The current PEP uses the nested design, where ExceptionGroups form a tree, and traceback information is distributed in pieces over this tree. This is the source of a lot of the complexity in the current PEP: for example, it's why EG's don't have one obvious iteration semantics, and it's why once an exception is wrapped in an EG, it can never be unwrapped again (because it would lose traceback information).

The idea of the "flat" design is to instead store all the traceback info directly on the leaf exceptions, so the EG itself can be just a pure container holding a list of exceptions, that's it, with no nesting. The downside is that it requires changes to the interpreter's code for updating __traceback__ attributes, which is currently hard-coded to only update one __traceback__ at a time.

For a third-party library like Trio, changing the interpreter is obviously impossible, so we never considered it seriously. But in a PEP, changing the interpreter is possible. And now I'm worried that we ruled out a better solution early on for reasons that no longer apply. The more I think about it, the more I suspect that flat EGs would end up being substantially simpler all around? So I think we should at least think through what that would look like (and Irit, I'd love your thoughts here now that you're the expert on the CPython details!), and document an explicit decision one way or another. (Maybe we should do a call or something to go over the details? I'm trying to keep this email from ballooning out of control...)

== Smaller points ==

- In my original proposal, EGs didn't just hold a list of exceptions, but also a list of "origins" for each exception. The idea being that if, say, you attempt to connect to a host with an IPv4 address and an IPv6 address, and they raised two different OSErrors that got bundled together into one EG, then it would be nice to know which OSError came from which attempt. Or in asyncio/trio it would be nice if tracebacks could show which task each exception came from. It seems like this got dropped at some point?

  On further consideration, I think this might be better handled as a special kind of traceback entry that we can attach to each exception, that just holds some arbitrary text that's inserted into the traceback at the appropriate place? But either way, I think it would be good to be able to attach this kind of information to tracebacks somehow.

- Recording pre-empted exceptions: This is another type of metadata that would be useful to print along with the traceback. It's non-obvious and a bit hard to explain, but multiple trio users have complained about this, so I assume it will bite asyncio users too as soon as TaskGroups are added. The situation is, you have a parent task P and two child tasks C1 and C2:

        P
       / \
      C1  C2


  C1 terminates with an unhandled exception E1, so in order to continue unwinding, the nursery/taskgroup in P cancels C2. But, C2 was itself in the middle of unwinding another, different exception E2 (so e.g. the cancellation arrived during a `finally` block). E2 gets replaced with a `Cancelled` exception whose __context__=E2, and that exception unwinds out of C2 and the nursery/taskgroup in P catches the `Cancelled` and discards it, then re-raises E1 so it can continue unwinding.

  The problem here is that E2 gets "lost" -- there's no record of it in the final output. Basically E1 replaced it. And that can be bad: for example, if the two children are interacting with each other, then E2 might be the actual error that broke the program, and E1 is some exception complaining that the connection to C2 was lost. If you have two exceptions that are triggered from the same underlying event, it's a race which one survives.

  This is conceptually similar to the way an exception in an 'except' block used to cause exceptions to be lost, so we added __context__ to avoid that. And just like for __context__, it would be nice if we could attach some info to E1 recording that E2 had happened and then got preempted. But I don't see how we can reuse __context__ itself for this, because it's a somewhat different relationship: __context__ means that an exception happened in the handler for another exception, while in this case you might have multiple preempted exceptions, and they're associated with particular points in the stack trace where the preemption occurred.

  This is a complex issue and maybe we should call it out-of-scope for the first version of ExceptionGroups. But I mention it because it's a second place where adding some extra annotations to the traceback info would be useful, and maybe we can keep it simple by adding some minimal hooks in the core traceback machinery and let libraries like trio/asyncio handle the complicated parts?

- There are a number of places where the Python VM itself catches exceptions and has hard-coded handling for certain exception types. For example:

  - Unhandled exceptions that reach the top of the main thread generally cause a traceback to be printed, but if the exception is SystemExit then the interpreter instead exits silently with status exc.args[0].

  - 'for' loops call iter.__next__, and catch StopIteration while allowing other exceptions to escape.

  - Generators catch StopIteration from their bodies and replace it with RuntimeError (PEP 479)

  With this PEP, it's now possible for the main thread to terminate with ExceptionGroup(SystemExit), __next__ to raise ExceptionGroup(StopIteration), a generator to raise ExceptionGroup(StopIteration), either alone or mixed with other exceptions. How should the VM handle these new cases? Should they be using except* or except?

  I don't think there's an obvious answer here, and possibly the answer is just "don't do that". But I feel like the PEP should say what the language semantics are in these cases, one way or another.

- traceback module API changes: The PEP notes that traceback.print_tb and traceback.print_exception will be updated to handle ExceptionGroups. The traceback module also has some special data structures for representing "pre-processed" stack traces, via the traceback.StackSummary type. This is used to capture tracebacks in a structured way but without holding onto the full frame objects. Maybe this API should also be extended somehow so it can also represent traceback trees?


On Sat, Mar 20, 2021 at 10:06 AM Irit Katriel via Python-Dev <python-dev@python.org> wrote:

We would like to present for feedback a new version of PEP 654, which incorporates the feedback we received in the discussions so far:
https://www.python.org/dev/peps/pep-0654/
The reference implementation has also been updated along with the PEP.

The changes we made since the first post are:

1. Instead of ExceptionGroup(BaseException), we will have two new builtin types: BaseExceptionGroup(BaseException) and ExceptionGroup(BaseExceptionGroup, Exception).
This is so that "except Exception" catches ExceptionGroups (but not BaseExceptionGroups). BaseExceptionGroup.__new__ inspects the wrapped exceptions, and if they are all Exception subclasses, it creates an ExceptionGroup instead of a BaseExceptionGroup.  

2. The exception group classes are not final - they can be subclassed and split()/subgroup() work correctly if the subclass overrides the derive() instance method as described here: https://www.python.org/dev/peps/pep-0654/#subclassing-exception-groups

3. We had some good suggestions on formatting exception groups, which we have implemented as you can see in the output shown for the examples in the PEP.

4. We expanded the section on handling Exception Groups, to show how subgroup can be used (with side effects) to do something for each leaf exception, and how to iterate correctly when the tracebacks of leaf exceptions are needed: https://www.python.org/dev/peps/pep-0654/#handling-exception-groups

5. We expanded the sections on rationale and backwards compatibility to explain our premise and expectations regarding how exception groups will be used and how the transition to using them will be managed.

6. We added several items to the rejected ideas section.

We did not receive any comments (or make any changes) to the proposed semantics of except*, hopefully this is because everyone thought they are sensible.

Irit, Yury and Guido

_______________________________________________
Python-Dev mailing list -- python-dev@python.org
To unsubscribe send an email to python-dev-leave@python.org
https://mail.python.org/mailman3/lists/python-dev.python.org/
Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/MQ2UCSQ2ZC4FIGT7KSVI6BJA4FCXSOCL/
Code of Conduct: http://python.org/psf/codeofconduct/


--
Nathaniel J. Smith -- https://vorpus.org