PEP 654: Exception Groups and except* [REPOST]
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
Everyone, Given the resounding silence I'm inclined to submit this to the Steering Council. While I'm technically a co-author, Irit has done almost all the work, and she's done a great job. If there are no further issues I'll send this SC-wards on Monday. --Guido On Sat, Mar 20, 2021 at 10:05 AM Irit Katriel <iritkatriel@googlemail.com> 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
-- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-change-the-world/>
On 3/26/2021 7:19 PM, Guido van Rossum wrote:
Everyone,
Given the resounding silence I'm inclined to submit this to the Steering Council. While I'm technically a co-author, Irit has done almost all the work, and she's done a great job. If there are no further issues I'll send this SC-wards on Monday.
The current version looks very carefully done. I leave it to the SC to review the details. But two thoughts on other possible uses. 1. unittests already uses the idea of exception groups by catching test_xyz exceptions and adding them to a group, to be printed it when all those for a file are run. If the PEP is accepted, someone might request an option to have them packaged as an ExceptionGroup. 2. At least some compilers for other languages can report multiple syntax exceptions, but at least for C, the usefulness of more than 2 or 3 is crippled by the synchronization problem. Python's compile() only reports the first, but I imagine that the usually dependable parsing into statements might make multiple reports for Python useful. I am thinking of a 'multiple' mode that repeatedly and recursively did 'single' compiles. (There would still be only one report per statement or compound statement header.) I have seen newbie code posted on Stackoverflow that needed a multi-error report. -- Terry Jan Reedy
On Sat, 27 Mar 2021, 9:24 am Guido van Rossum, <guido@python.org> wrote:
Everyone,
Given the resounding silence I'm inclined to submit this to the Steering Council. While I'm technically a co-author, Irit has done almost all the work, and she's done a great job. If there are no further issues I'll send this SC-wards on Monday.
+1 from me. I remember talking to Nathaniel (and Yury IIRC) about early iterations of the MultiError handling in trio at PyCon years ago, so it's great to see the way the concept has matured in the intervening time. The one thing I did think was that it might be good to mention concurrent.futures and contextlib in the PEP as modules that could potentially take advantage of exception groups. I'm not sure it's really necessary though, as the presentation is already compelling as it is. Cheers, Nick. P.S. The potential use case I see in contextlib: the way ExitStack nests dynamic resource handling *works*, but the tracebacks when multiple clean up steps fail can be spectacularly cryptic. With exception groups available, it should be possible to define a new "context lib.ExitGroup" API that flattens out the resource clean up and collects any exceptions that occur into a more comprehensible exception group, rather than building a nested exception context tree the way ExitStack does.
--Guido
On Sat, Mar 20, 2021 at 10:05 AM Irit Katriel <iritkatriel@googlemail.com> 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
-- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-change-the-world/> _______________________________________________ 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/AQGDIZDZ... Code of Conduct: http://python.org/psf/codeofconduct/
Hello, On Fri, 26 Mar 2021 16:19:26 -0700 Guido van Rossum <guido@python.org> wrote:
Everyone,
Given the resounding silence I'm inclined to submit this to the Steering Council. While I'm technically a co-author, Irit has done almost all the work, and she's done a great job. If there are no further issues I'll send this SC-wards on Monday.
One issue with PEP654 is that it introduces pretty adhoc and complex-semantics concept (ExceptionGroup) on the language level. Here's an idea (maybe duplicate) on how to introduce a much simpler, and more generic concept on the language level, and let particular frameworks to introduce (and elaborate without further changing the language) adhoc concept they need. So, let's look how the usual "except MyExc as e" works: it performs "isinstance(e0, MyExc)" operation, where e0 is incoming exception (roughly, sys.exc_info[1]), and if it returns True, then assigns e0 to the "e" variable and executes handler body. "isinstance(e0, MyExc)" is formally known as an "exception filter". As we see, currently Python hardcodes isinstance() as exception filter. The idea is to allow to use an explicit exception filter. Let's reuse the same "except *" syntax to specify it. Also, it's more flexible instead of returning True/False from filter, to return either None (filter didn't match), or an exception object to make available to handler (which in general may be different than passed to the filter). With this, ExceptionGroup usecases should be covered. Examples: 1. Current implicit exception filter is equivalent to: def implicit(e0, excs): # one or tuple, as usual if isinstance(e0, excs): return e0 return None try: 1/0 except *implicit(ZeroDivisionError) as e: print(e) 2. Allow to catch main or chained exception (context manager example from PEP) def chained(e, excs): while e: if isinstance(e, excs): return e e = e.__cause__ # Simplified, should consider __context__ too try: tempfile.TemporaryDirectory(...) except *chained(OSError) as e: print(e) 3. Rough example of ExceptionGroup functionality (now not a language builtin, just implemented by framework(s) which need it, or as a separate module): class ExceptionGroup: ... @staticmethod def match(e0, excs): cur, rest = e0.split_by_types(excs) # That's how we allow an exception handler to re-raise either an # original group in full or just "unhandled" exception in the # group (or anything) - everything should be passed via # exception attributes (or computed by methods). cur.org = e0 cur.rest = rest return cur try: ... except *ExceptionGroup.match((TypeError, ValueError)) as e: # try to handle a subgroup with (TypeError, ValueError) here ... # now reraise a subgroup with unhandled exceptions from the # original group raise e.rest
--Guido
On Sat, Mar 20, 2021 at 10:05 AM Irit Katriel <iritkatriel@googlemail.com> 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
-- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-change-the-world/>
-- Best regards, Paul mailto:pmiscml@gmail.com
One of the motivations for introducing ExceptionGroup as a builtin is so that we won't have a different custom version in each library that needs it. So if you are writing a library the needs to raise multiple exceptions, and then you decide to call Trio, you don't need to translate Trio's MultiError into your own exception group type, because everybody uses the builtin. And your users don't need to learn how your particular exception group works because they know that you are using the builtin one. I see the aesthetic value of your suggestion, but does it have practical advantages in light of the above? Irit On Sat, Mar 27, 2021 at 10:31 AM Paul Sokolovsky <pmiscml@gmail.com> wrote:
Hello,
On Fri, 26 Mar 2021 16:19:26 -0700 Guido van Rossum <guido@python.org> wrote:
Everyone,
Given the resounding silence I'm inclined to submit this to the Steering Council. While I'm technically a co-author, Irit has done almost all the work, and she's done a great job. If there are no further issues I'll send this SC-wards on Monday.
One issue with PEP654 is that it introduces pretty adhoc and complex-semantics concept (ExceptionGroup) on the language level. Here's an idea (maybe duplicate) on how to introduce a much simpler, and more generic concept on the language level, and let particular frameworks to introduce (and elaborate without further changing the language) adhoc concept they need.
So, let's look how the usual "except MyExc as e" works: it performs "isinstance(e0, MyExc)" operation, where e0 is incoming exception (roughly, sys.exc_info[1]), and if it returns True, then assigns e0 to the "e" variable and executes handler body. "isinstance(e0, MyExc)" is formally known as an "exception filter".
As we see, currently Python hardcodes isinstance() as exception filter. The idea is to allow to use an explicit exception filter. Let's reuse the same "except *" syntax to specify it. Also, it's more flexible instead of returning True/False from filter, to return either None (filter didn't match), or an exception object to make available to handler (which in general may be different than passed to the filter). With this, ExceptionGroup usecases should be covered.
Examples:
1. Current implicit exception filter is equivalent to:
def implicit(e0, excs): # one or tuple, as usual if isinstance(e0, excs): return e0 return None
try: 1/0 except *implicit(ZeroDivisionError) as e: print(e)
2. Allow to catch main or chained exception (context manager example from PEP)
def chained(e, excs): while e: if isinstance(e, excs): return e e = e.__cause__ # Simplified, should consider __context__ too
try: tempfile.TemporaryDirectory(...) except *chained(OSError) as e: print(e)
3. Rough example of ExceptionGroup functionality (now not a language builtin, just implemented by framework(s) which need it, or as a separate module):
class ExceptionGroup:
...
@staticmethod def match(e0, excs): cur, rest = e0.split_by_types(excs) # That's how we allow an exception handler to re-raise either an # original group in full or just "unhandled" exception in the # group (or anything) - everything should be passed via # exception attributes (or computed by methods). cur.org = e0 cur.rest = rest return cur
try: ... except *ExceptionGroup.match((TypeError, ValueError)) as e: # try to handle a subgroup with (TypeError, ValueError) here ... # now reraise a subgroup with unhandled exceptions from the # original group raise e.rest
Hello, On Sat, 27 Mar 2021 10:55:40 +0000 Irit Katriel <iritkatriel@googlemail.com> wrote:
One of the motivations for introducing ExceptionGroup as a builtin is so that we won't have a different custom version in each library that needs it. So if you are writing a library the needs to raise multiple exceptions, and then you decide to call Trio, you don't need to translate Trio's MultiError into your own exception group type, because everybody uses the builtin.
Looking from a different angle shows a different perspective: 1. Trio devised an interesting concept of "nurseries" to deal with multiple tasks in asyncio-like programming. 2. It was all nice and beautiful until ... it came to error handling. Note that it's a rather typical situation - one can write nice, clean, beautiful code, which is not adequate in real-world scenarios, particularly because of error handling concerns. Oftentimes, such initial code/concepts are discarded, and more robust (though maybe not as beautiful) concepts/code is used. 3. But that's not what happened in the Trio case. The concept of nurseries got pushed forward, until it became clear that it requires changes on the programming language level. 4. That's how PEP654 was born, which, besides async scheduling of multiple tasks, bring *somewhat similar* usecases of e.g. raising exceptions thru context managers' own exceptions. Note that where errors and exceptions lead us is in questions on how to handle them. And beyond a couple of well known patterns ("dump and crash" or "dump and continue with next iteration"), error handling is very adhoc to a particular application and particular error(s). Seen from that angle, Trio wants to vendor-lock the entire language into its particular (experimental) way of handling multiple errors. So yes, maybe being well aware that you're handling exactly Trio's (or asyncio's, or context manager's) error isn't bad thing actually. And if it's clear that multiple asyncio frameworks are interested in sharing common exception base class for such usecases, then it could be introduced on the "asyncio" package level, or maybe even better, as a module similar to "concurrent.futures".
And your users don't need to learn how your particular exception group works because they know that you are using the builtin one.
I see the aesthetic value of your suggestion, but does it have practical advantages in light of the above?
The concern is that it codifies pretty complex and special-purpose things on the language level. And it seems that the whole concept is rather experimental and "original design". We can compare that with another sufficiently complex feature which landed recently: pattern matching. At all phases of the design and discussion of that feature, one of the guiding principles was: "Many other languages implement it, so we know it's generally useful, and have design space more or less charted, and can vary things to find local optimum for Python". Contrary to that, the original PEP654 didn't refer to (cross-language, cross-framework) prior art in handling multiple errors, and I don't see that changed in the latest version. So I'm sorry, but it seems like NIH feature of a specific 3rd-party framework being promoted to a whole language's way of doing things. Under such circumstance, I guess it would be good idea to try to decouple behavior of that feature from the languages core, and make aspects of behavior more explicit (following "explicit is better than implicit" principle), and allow to vary/evolve it without changing the core language. I tried to draft a scheme aspiring to allow that. (Which would definitely need more work to achieve parity with functionality in PEP654, because again, it tries to codify rather complex and "magic" behavior. Where complex and magic behavior in exception handling is itself of concern, so making it more explicit may be a good idea. (A good "how other languages deal with it") review would mention that Go, Rust, Zig, etc. don't have and condemn exception handling at all (normal simple classy exceptions, not magic we discuss here!)). Thanks, Paul
Irit
On Sat, Mar 27, 2021 at 10:31 AM Paul Sokolovsky <pmiscml@gmail.com> wrote:
Hello,
On Fri, 26 Mar 2021 16:19:26 -0700 Guido van Rossum <guido@python.org> wrote:
Everyone,
Given the resounding silence I'm inclined to submit this to the Steering Council. While I'm technically a co-author, Irit has done almost all the work, and she's done a great job. If there are no further issues I'll send this SC-wards on Monday.
One issue with PEP654 is that it introduces pretty adhoc and complex-semantics concept (ExceptionGroup) on the language level. Here's an idea (maybe duplicate) on how to introduce a much simpler, and more generic concept on the language level, and let particular frameworks to introduce (and elaborate without further changing the language) adhoc concept they need.
So, let's look how the usual "except MyExc as e" works: it performs "isinstance(e0, MyExc)" operation, where e0 is incoming exception (roughly, sys.exc_info[1]), and if it returns True, then assigns e0 to the "e" variable and executes handler body. "isinstance(e0, MyExc)" is formally known as an "exception filter".
As we see, currently Python hardcodes isinstance() as exception filter. The idea is to allow to use an explicit exception filter. Let's reuse the same "except *" syntax to specify it. Also, it's more flexible instead of returning True/False from filter, to return either None (filter didn't match), or an exception object to make available to handler (which in general may be different than passed to the filter). With this, ExceptionGroup usecases should be covered.
Examples:
1. Current implicit exception filter is equivalent to:
def implicit(e0, excs): # one or tuple, as usual if isinstance(e0, excs): return e0 return None
try: 1/0 except *implicit(ZeroDivisionError) as e: print(e)
2. Allow to catch main or chained exception (context manager example from PEP)
def chained(e, excs): while e: if isinstance(e, excs): return e e = e.__cause__ # Simplified, should consider __context__ too
try: tempfile.TemporaryDirectory(...) except *chained(OSError) as e: print(e)
3. Rough example of ExceptionGroup functionality (now not a language builtin, just implemented by framework(s) which need it, or as a separate module):
class ExceptionGroup:
...
@staticmethod def match(e0, excs): cur, rest = e0.split_by_types(excs) # That's how we allow an exception handler to re-raise either an # original group in full or just "unhandled" exception in the # group (or anything) - everything should be passed via # exception attributes (or computed by methods). cur.org = e0 cur.rest = rest return cur
try: ... except *ExceptionGroup.match((TypeError, ValueError)) as e: # try to handle a subgroup with (TypeError, ValueError) here ... # now reraise a subgroup with unhandled exceptions from the # original group raise e.rest
-- Best regards, Paul mailto:pmiscml@gmail.com
Hi Paul, IIUC, you are saying that exception group should not be a builtin type because it is (1) complex (2) special-purposed. Instead, you propose that we make exception handling pluggable. (1) I agree, it is somewhat complex - it took us several iterations to get from the idea of "a container of exceptions" to something that works correctly with split()/subgroup(), maintaining the integrity of __traceback__/__cause__/__context__ information, allowing subclassing and having a relatively simple API. To me this is a reason to provide it as a builtin rather than a reason not to (I think we should add this point to the PEP- thanks). You wrote:
So yes, maybe being well aware that you're handling exactly Trio's (or asyncio's, or context manager's) error isn't bad thing actually.
Maybe, in which case you can create a subclass of ExceptionGroup to mark the groups coming from your library. But also maybe not, in which case you don't have to. (2) special-purposed? How so? When I asked for a practical advantage of your pluggable solution I meant an example of a use case that our proposal doesn't accommodate well and where a pluggable one does better. The only example you suggested so far was the TemporaryDirectory one, which I don't find compelling:
def chained(e, excs): while e: if isinstance(e, excs): return e e = e.__cause__ # Simplified, should consider __context__ too
try: tempfile.TemporaryDirectory(...) except *chained(OSError) as e: print(e)
What if the user code raised an OSError too, so now that exception group has more than one? Even without that problem, why should I prefer this to our proposal? I am interested to know (1) what limitations you see in the ExceptionGroup type we are proposing (in what sense is it special-purposed? what purpose is it unsuitable for?) and (2) an example where the pluggable solution works better. Irit On Sat, Mar 27, 2021 at 12:42 PM Paul Sokolovsky <pmiscml@gmail.com> wrote:
Hello,
On Sat, 27 Mar 2021 10:55:40 +0000 Irit Katriel <iritkatriel@googlemail.com> wrote:
One of the motivations for introducing ExceptionGroup as a builtin is so that we won't have a different custom version in each library that needs it. So if you are writing a library the needs to raise multiple exceptions, and then you decide to call Trio, you don't need to translate Trio's MultiError into your own exception group type, because everybody uses the builtin.
Looking from a different angle shows a different perspective:
1. Trio devised an interesting concept of "nurseries" to deal with multiple tasks in asyncio-like programming.
2. It was all nice and beautiful until ... it came to error handling. Note that it's a rather typical situation - one can write nice, clean, beautiful code, which is not adequate in real-world scenarios, particularly because of error handling concerns. Oftentimes, such initial code/concepts are discarded, and more robust (though maybe not as beautiful) concepts/code is used.
3. But that's not what happened in the Trio case. The concept of nurseries got pushed forward, until it became clear that it requires changes on the programming language level.
4. That's how PEP654 was born, which, besides async scheduling of multiple tasks, bring *somewhat similar* usecases of e.g. raising exceptions thru context managers' own exceptions.
Note that where errors and exceptions lead us is in questions on how to handle them. And beyond a couple of well known patterns ("dump and crash" or "dump and continue with next iteration"), error handling is very adhoc to a particular application and particular error(s). Seen from that angle, Trio wants to vendor-lock the entire language into its particular (experimental) way of handling multiple errors.
So yes, maybe being well aware that you're handling exactly Trio's (or asyncio's, or context manager's) error isn't bad thing actually. And if it's clear that multiple asyncio frameworks are interested in sharing common exception base class for such usecases, then it could be introduced on the "asyncio" package level, or maybe even better, as a module similar to "concurrent.futures".
And your users don't need to learn how your particular exception group works because they know that you are using the builtin one.
I see the aesthetic value of your suggestion, but does it have practical advantages in light of the above?
The concern is that it codifies pretty complex and special-purpose things on the language level. And it seems that the whole concept is rather experimental and "original design". We can compare that with another sufficiently complex feature which landed recently: pattern matching. At all phases of the design and discussion of that feature, one of the guiding principles was: "Many other languages implement it, so we know it's generally useful, and have design space more or less charted, and can vary things to find local optimum for Python".
Contrary to that, the original PEP654 didn't refer to (cross-language, cross-framework) prior art in handling multiple errors, and I don't see that changed in the latest version. So I'm sorry, but it seems like NIH feature of a specific 3rd-party framework being promoted to a whole language's way of doing things.
Under such circumstance, I guess it would be good idea to try to decouple behavior of that feature from the languages core, and make aspects of behavior more explicit (following "explicit is better than implicit" principle), and allow to vary/evolve it without changing the core language.
I tried to draft a scheme aspiring to allow that. (Which would definitely need more work to achieve parity with functionality in PEP654, because again, it tries to codify rather complex and "magic" behavior. Where complex and magic behavior in exception handling is itself of concern, so making it more explicit may be a good idea. (A good "how other languages deal with it") review would mention that Go, Rust, Zig, etc. don't have and condemn exception handling at all (normal simple classy exceptions, not magic we discuss here!)).
Thanks, Paul
On Sun, 28 Mar 2021, 12:34 am Irit Katriel via Python-Dev, < python-dev@python.org> wrote:
Hi Paul,
IIUC, you are saying that exception group should not be a builtin type because it is (1) complex (2) special-purposed. Instead, you propose that we make exception handling pluggable.
Note that (2) would be the *exact opposite* of the direction that we have taken exception handling in general: it used to be far more permissive than it is, and then got more restrictive over time to make structured exception handling more consistent. In particular, features like exception chaining and use case independent trace back printing only work because exception consumers can rely on all exceptions being instances of BaseException and having particular double-underscore attributes like "cause", "context", and "traceback" The lesson we've learned from Trio's experience is: * regular exceptions are still fine for most purposes * when they're not fine, they're not fine in a consistent way across a variety of use cases where the existing exception machinery currently loses or hides information, the problem isn't specific to Trio's nursery API I understand the temptation to generalise, but a general proposal would fail for the same reason the "user defined infix operators" PEPs failed: despite decades of experience, there just weren't compelling use cases for a general purpose feature. Instead, the proposal that succeeded was narrow: it targeted the feature that people actually wanted (an infix operator for matrix multiplication), rather than proposing more general machinery that could be used to build that feature. Cheers, Nick.
Hello, On Sat, 27 Mar 2021 14:29:21 +0000 Irit Katriel <iritkatriel@googlemail.com> wrote:
Hi Paul,
IIUC, you are saying that exception group should not be a builtin type because it is (1) complex (2) special-purposed. Instead, you propose that we make exception handling pluggable.
Yes, I wanted to mention that alternative possibility. I have to admit that I haven't read the entire discussion of the PEP, but the messages I read oftentimes were over-optimistic about the feature, rather than considering its issues/possible alternatives.
(1) I agree, it is somewhat complex - it took us several iterations to get from the idea of "a container of exceptions" to something that works correctly with split()/subgroup(), maintaining the integrity of __traceback__/__cause__/__context__ information, allowing subclassing and having a relatively simple API. To me this is a reason to provide it as a builtin rather than a reason not to (I think we should add this point to the PEP- thanks).
It definitely feels like a lot of effort went into devising and polishing ExceptionGroup's and except*, thanks. But I'm not sure if you gentlemen come up with the "ultimate" way to deal with multiple errors, which deserves being codified on the language level (vs be added as pluggable means indeed). That's why I asked if there's any prior art of a similar solution(s) in other languages and/or systems - that would definitely sooth concerns and add credibility. (And I'm not saying that Python can't lead there - just that more sustainable way would be to try and make it more generic/explicit rather than fix on a particular novel solution right away.)
You wrote:
So yes, maybe being well aware that you're handling exactly Trio's (or asyncio's, or context manager's) error isn't bad thing actually.
Maybe, in which case you can create a subclass of ExceptionGroup to mark the groups coming from your library. But also maybe not, in which case you don't have to.
(2) special-purposed? How so? When I asked for a practical advantage of your pluggable solution I meant an example of a use case that our proposal doesn't accommodate well and where a pluggable one does better. The only example you suggested so far was the TemporaryDirectory one, which I don't find compelling:
That's the impression I've got from reading the "Motivation" section. It cites Trio and multiple errors from async tasks as the guiding usecase. It also mentions a couple of cases in the stdlib, where my personal impression is that a simpler mechanism would suffice. Then it also cites some 3rd-party libs and their maintainers, who say such thing *might* be useful. But it's going to take practice to see if the actual thing added is useful. What if other projects turn out to want something slightly different (e.g. simpler?). Oh, they won't have much choice (to complain), as it's already shipped to them, not even in the stdlib, but in the language! Bottom line: this seems like a Trio's special-purpose feature, with good wishes of becoming the de facto standard. I find that somewhat ironical, as Trio's original motivation was to explore alternative ways of doing async framework in Python. And now it hardcodes its way of dealing with multiple errors in the language (no more exploration and alternatives).
def chained(e, excs): while e: if isinstance(e, excs): return e e = e.__cause__ # Simplified, should consider __context__ too
try: tempfile.TemporaryDirectory(...) except *chained(OSError) as e: print(e)
What if the user code raised an OSError too, so now that exception group has more than one?
I don't see how that's different from PEP654's ExceptionGroup's case. It's just in ExceptionGroup, there would be 2 (unrelated?) OSError's, while using normal chaining rules, there's at least context info which is preserved.
Even without that problem, why should I prefer this to our proposal?
My idea was to show an alternative not mentioned in PEP654's "Rejected Ideas" section.
I am interested to know (1) what limitations you see in the ExceptionGroup type we are proposing (in what sense is it special-purposed?
Just as a specific example, what if my application indeed needs ExceptionGroup's (and not just more flexible matching of existing exception chains), but I want it to be a bit simpler: a) to have a flat structure of contained exceptions; b) want to actually emphasize the users that they should not rely on the ordering, so want it to be a set. My understanding is that ExceptionGroup is carefully designed to allow subclassing it and override behavior, and the above should be possible by subclassing. But usually, a simpler base class is subclassed to have more advanced/complex behavior. So again, current ExceptionGroup feels kind of specialized with its careful attention to hierarchy and multi-story "3D" tracebacks. It will be good if you actually found the sweet spot. But what if not, and it will be a UX chore (Python's chore, nor a particular framework's chore)?
what purpose is it unsuitable for?) and (2) an example where the pluggable solution works better.
From my PoV, a solution which doesn't add particular complex behavior into the language core, but allows to plug it in, and keep whole thing more explicit, is something that "works better". Again, I wanted to mention that point, as between the initial PEP version and the latest one, I don't see that alternative to have been mentioned in the PEP (which might mean that it wasn't mentioned enough by other reviewers).
Irit
On Sat, Mar 27, 2021 at 12:42 PM Paul Sokolovsky <pmiscml@gmail.com> wrote:
Hello,
On Sat, 27 Mar 2021 10:55:40 +0000 Irit Katriel <iritkatriel@googlemail.com> wrote:
One of the motivations for introducing ExceptionGroup as a builtin is so that we won't have a different custom version in each library that needs it. So if you are writing a library the needs to raise multiple exceptions, and then you decide to call Trio, you don't need to translate Trio's MultiError into your own exception group type, because everybody uses the builtin.
Looking from a different angle shows a different perspective:
1. Trio devised an interesting concept of "nurseries" to deal with multiple tasks in asyncio-like programming.
2. It was all nice and beautiful until ... it came to error handling. Note that it's a rather typical situation - one can write nice, clean, beautiful code, which is not adequate in real-world scenarios, particularly because of error handling concerns. Oftentimes, such initial code/concepts are discarded, and more robust (though maybe not as beautiful) concepts/code is used.
3. But that's not what happened in the Trio case. The concept of nurseries got pushed forward, until it became clear that it requires changes on the programming language level.
4. That's how PEP654 was born, which, besides async scheduling of multiple tasks, bring *somewhat similar* usecases of e.g. raising exceptions thru context managers' own exceptions.
Note that where errors and exceptions lead us is in questions on how to handle them. And beyond a couple of well known patterns ("dump and crash" or "dump and continue with next iteration"), error handling is very adhoc to a particular application and particular error(s). Seen from that angle, Trio wants to vendor-lock the entire language into its particular (experimental) way of handling multiple errors.
So yes, maybe being well aware that you're handling exactly Trio's (or asyncio's, or context manager's) error isn't bad thing actually. And if it's clear that multiple asyncio frameworks are interested in sharing common exception base class for such usecases, then it could be introduced on the "asyncio" package level, or maybe even better, as a module similar to "concurrent.futures".
And your users don't need to learn how your particular exception group works because they know that you are using the builtin one.
I see the aesthetic value of your suggestion, but does it have practical advantages in light of the above?
The concern is that it codifies pretty complex and special-purpose things on the language level. And it seems that the whole concept is rather experimental and "original design". We can compare that with another sufficiently complex feature which landed recently: pattern matching. At all phases of the design and discussion of that feature, one of the guiding principles was: "Many other languages implement it, so we know it's generally useful, and have design space more or less charted, and can vary things to find local optimum for Python".
Contrary to that, the original PEP654 didn't refer to (cross-language, cross-framework) prior art in handling multiple errors, and I don't see that changed in the latest version. So I'm sorry, but it seems like NIH feature of a specific 3rd-party framework being promoted to a whole language's way of doing things.
Under such circumstance, I guess it would be good idea to try to decouple behavior of that feature from the languages core, and make aspects of behavior more explicit (following "explicit is better than implicit" principle), and allow to vary/evolve it without changing the core language.
I tried to draft a scheme aspiring to allow that. (Which would definitely need more work to achieve parity with functionality in PEP654, because again, it tries to codify rather complex and "magic" behavior. Where complex and magic behavior in exception handling is itself of concern, so making it more explicit may be a good idea. (A good "how other languages deal with it") review would mention that Go, Rust, Zig, etc. don't have and condemn exception handling at all (normal simple classy exceptions, not magic we discuss here!)).
Thanks, Paul
-- Best regards, Paul mailto:pmiscml@gmail.com
Hi Paul, On Sat, Mar 27, 2021 at 6:00 PM Paul Sokolovsky <pmiscml@gmail.com> wrote:
It definitely feels like a lot of effort went into devising and polishing ExceptionGroup's and except*, thanks. But I'm not sure if you gentlemen come up with the "ultimate" way to deal with multiple errors,
I've been mistaken for a man before, but no-one has ever confused me for gentle. I'll take that as a compliment.
which deserves being codified on the language level (vs be added as pluggable means indeed).
Pluggable is not without its problems. I'm all in favor of you developing this idea and proposing an alternative. As I said before, you just need to answer two questions: 1. show a limitation of our approach (the contrived flat-set is not one - see below) 2. describe how a pluggable approach works for that case
def chained(e, excs): while e: if isinstance(e, excs): return e e = e.__cause__ # Simplified, should consider __context__ too
try: tempfile.TemporaryDirectory(...) except *chained(OSError) as e: print(e)
What if the user code raised an OSError too, so now that exception group has more than one?
I don't see how that's different from PEP654's ExceptionGroup's case. It's just in ExceptionGroup, there would be 2 (unrelated?) OSError's, while using normal chaining rules, there's at least context info which is preserved.
It's different because your "match" function returns a single exception (the first one that is of OSError type). Any further OSErrors will be reraised. The PEP's except* knows how to match multiple exceptions of the relevant type.
I am interested to know (1) what limitations you see in the ExceptionGroup type we are proposing (in what sense is it special-purposed?
Just as a specific example, what if my application indeed needs ExceptionGroup's (and not just more flexible matching of existing exception chains), but I want it to be a bit simpler: a) to have a flat structure of contained exceptions; b) want to actually emphasize the users that they should not rely on the ordering, so want it to be a set.
My understanding is that ExceptionGroup is carefully designed to allow subclassing it and override behavior, and the above should be possible by subclassing.
Indeed: class PaulsExceptionGroup(ExceptionGroup): def __new__(cls, message, excs): for e in excs: if isinstance(e, BaseExceptionGroup): raise TypeError("me is a flat exception group") return super().__new__(cls, message, excs) @property def exceptions(self): return set(super().exceptions) def derive(self, excs): return PaulsExceptionGroup(self.message, excs) eg1 = PaulsExceptionGroup("Paul's group", [ValueError(12), TypeError(42)]) print(repr(eg1)) print(eg1.exceptions) match, rest = eg1.split(ValueError) print(f'{match=!r}\n{rest=!r}') print(f'{match.exceptions=}\n{rest.exceptions=}') Output: PaulsExceptionGroup("Paul's group", [ValueError(12), TypeError(42)]) {TypeError(42), ValueError(12)} match=PaulsExceptionGroup("Paul's group", [ValueError(12)]) rest=PaulsExceptionGroup("Paul's group", [TypeError(42)]) match.exceptions={ValueError(12)} rest.exceptions={TypeError(42)} See - split() just works for you out of the box. But usually, a simpler base class is subclassed to have
more advanced/complex behavior. So again, current ExceptionGroup feels kind of specialized with its careful attention to hierarchy and multi-story "3D" tracebacks.
So your concern with our design is that ExceptionGroup implements a generic split() that handles tracebacks and nested structure correctly, and you might not need that because maybe you don't nest or you don't care about tracebacks? If that is the case then I think you are confusing "generic" with "special-purpose". ExceptionGroup is generic, it works for general nested groups with arbitrary tracebacks and cause/context. That's the opposite of special-purpose. Irit
While we're talking about compelling use cases, does anyone have an actual, concrete use case for the proposed "except *" feature that's strong enough to justify new syntax? I'm fine with having ExceptionGroup as a built-in type. I'm not fine with adding new syntax that will apparently be used only in rare circumstances. Can code that's aware of the possibility of getting an ExceptionGroup not simply catch it as a normal exception and then pick it apart? Do we really need a whole new piece of machinery for this? -- Greg
Hi Greg, If all you want is to catch an exception group and process it, then except* does look like overkill. It gets more interesting if you want to handle only some of the exceptions and reraise the rest (without adding the current frame to the traceback), or when the exception handler raises/reraises exceptions (with 'raise e' or bare 'raise'). We can look to Trio to see what exception handling would be like if we add ExceptionGroup without except*. Trio's exception group is called MultiError, and this part of the tutorial explains how they can be handled: https://trio.readthedocs.io/en/stable/reference-core.html#trio.MultiError So asyncio (or some new module in the stdlib) would probably end up exposing something along the lines of MultiError.catch, a context manager where you register the exception handler as a callback and the unhandled exceptions are automatically reraised. You can see the trio code for that here: https://github.com/python-trio/trio/blob/master/trio/_core/_multierror.py#L1... The trick it does to reset __context__ in the finally block won't work for the __traceback__ (I tried). So it always adds the current frame (if it doesn't return before line 145). Also, it doesn't seem to support raising exceptions from the handler: https://github.com/python-trio/trio/blob/6754c74eacfad9cc5c92d5c24727a2f3b62... So yes, we can catch exception groups and handle them like normal exceptions, but I don't see a clean way to catch parts of them selectively or to raise exceptions from exception handling code. Irit On Sat, Mar 27, 2021 at 10:57 PM Greg Ewing <greg.ewing@canterbury.ac.nz> wrote:
While we're talking about compelling use cases, does anyone have an actual, concrete use case for the proposed "except *" feature that's strong enough to justify new syntax?
I'm fine with having ExceptionGroup as a built-in type. I'm not fine with adding new syntax that will apparently be used only in rare circumstances.
Can code that's aware of the possibility of getting an ExceptionGroup not simply catch it as a normal exception and then pick it apart? Do we really need a whole new piece of machinery for this?
-- Greg
_______________________________________________ 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/EIW6JGXR... Code of Conduct: http://python.org/psf/codeofconduct/
On Sat, Mar 27, 2021 at 3:56 PM Greg Ewing <greg.ewing@canterbury.ac.nz> wrote:
While we're talking about compelling use cases, does anyone have an actual, concrete use case for the proposed "except *" feature that's strong enough to justify new syntax?
I'm fine with having ExceptionGroup as a built-in type. I'm not fine with adding new syntax that will apparently be used only in rare circumstances.
Can code that's aware of the possibility of getting an ExceptionGroup not simply catch it as a normal exception and then pick it apart? Do we really need a whole new piece of machinery for this?
As Irit already wrote, the code to do this correctly is increasingly subtle and without special syntax it becomes too easy to do it wrong. Also note that multiple `except *` clauses can all run, since they handle (disjunct) different subsets of the group. -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-change-the-world/>
Just a few comments to add to Irit's response. On Sat, Mar 27, 2021 at 11:03 AM Paul Sokolovsky <pmiscml@gmail.com> wrote: [..]
Bottom line: this seems like a Trio's special-purpose feature, with good wishes of becoming the de facto standard.
The bottom line is that Trio nurseries were proven to be a very useful and intuitive primitive. But the error handling API around them is unintuitive and hard to use. Now that we want to add an equivalent of nurseries to asyncio (we'll likely call them Task Groups) we need to sort out the error handling mechanism, finally.
From my PoV, a solution which doesn't add particular complex behavior into the language core, but allows to plug it in, and keep whole thing more explicit, is something that "works better".
I understand your PoV, but there's a point where building explicit complex APIs becomes so detrimental to the overall usability that there's a warrant for solving the problem in syntax and with builtin types. Yury
Hello, On Sat, 27 Mar 2021 20:01:27 +0000 Irit Katriel <iritkatriel@googlemail.com> wrote: []
you gentlemen come up with the "ultimate" way to deal with multiple errors,
I've been mistaken for a man before, but no-one has ever confused me for gentle. I'll take that as a compliment.
Sorry, was just a figure of speech ;-). []
Pluggable is not without its problems. I'm all in favor of you developing this idea and proposing an alternative. As I said before, you just need to answer two questions: 1. show a limitation of our approach (the contrived flat-set is not one - see below) 2. describe how a pluggable approach works for that case
As I mentioned, I wanted to share concern about adding complex, arguably special-purpose semantics to the core language, and back that by an example on how to address some (not all) points of PEP 645 in a more lightweight (on the language more manner). It's hard to show a technical limitation of your approach - because again, it's already complex and special-purpose, and I can't say I've faced cases where regular application of it would be guaranteed and omissions would be visible. I vice versa can point to a limitation in "generic exception filters" approach - it handles "how to match exceptions more flexibly" issue, but doesn't cover "how to handle automagic reraising of multiple (unhandled) exceptions" part. In my list, that's good - in normal exception handling, exception are (re)raised explicitly. It's pretty big conceptual jump to make handling of that implicit (and exception handling in general is condemned as too-implicit control flow construct by some parties). Overall, my idea wasn't to come up with an alternative proposal so late in PEP654 lifecycle, nor even delay its submission to SC. I just decided to sounds these concern in the end (had them in my notes since initial posting), and ask for clarifications like if there's a cross-language prior art regarding it. []
def chained(e, excs): while e: if isinstance(e, excs): return e e = e.__cause__ # Simplified, should consider __context__
It's different because your "match" function returns a single exception (the first one that is of OSError type). Any further OSErrors will be reraised. The PEP's except* knows how to match multiple exceptions of the relevant type.
For this case, it was "chained" function (exception filter), which was concerned only with catching OSError anywhere in the exception chain. As mentioned in the other comment, if there's a need to reraise (parts) of original exception, in must be stored as an attribute of an exception returned from filter. All that's pretty clear and explicit, though I agree if it (reraising) needs to be done often, it may be cumbersome. If it needs to be done. We here touch the question of how exactly such (multiple) exceptions would be handled. And PEP654 is pretty bare on that too - it offers machinery to make that possible, but barely talks of actual usecases (where it starts talk about that is in "Rejected ideas" section, not in the main narrative).
My understanding is that ExceptionGroup is carefully designed to allow subclassing it and override behavior, and the above should be possible by subclassing.
Indeed:
class PaulsExceptionGroup(ExceptionGroup):
Thanks for confirming. []
So your concern with our design is that ExceptionGroup implements a generic split() that handles tracebacks and nested structure correctly, and you might not need that because maybe you don't nest or you don't care about tracebacks?
My concern is still that it adds quite a big chunk of complexity to the core semantics of language. PEP654's ExceptionGroup is clearly carefully designed to (try) to be general, there's no questions about that.
If that is the case then I think you are confusing "generic" with "special-purpose". ExceptionGroup is generic, it works for general nested groups with arbitrary tracebacks and cause/context. That's the opposite of special-purpose.
Sadly, to me it still looks that all its generic complexity is rooted in Trio and friends, and for other usecases simpler approach would suffice. Again, that's just personal opinion. I'm just worried to see "except*" in code where it's really needed (and be able to easily grasp what it does), and dread seeing it where it's not really needed (in that regard, PEP654 tries to delineate it clearly from the rest of Python, thanks).
Irit
-- Best regards, Paul mailto:pmiscml@gmail.com
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/MQ2UCSQ2... Code of Conduct: http://python.org/psf/codeofconduct/
-- Nathaniel J. Smith -- https://vorpus.org <http://vorpus.org>
On Mon, Apr 5, 2021 at 3:07 AM Nathaniel Smith <njs@pobox.com> wrote:
- 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 point reminded me again of this issue in the tracker ("Problems with recursive automatic exception chaining" from 2013): https://bugs.python.org/issue18861 I'm not sure if it's exactly the same, but you can see that a couple of the later comments there talk about "exception trees" and other types of annotations. If that issue were addressed after ExceptionGroups were introduced, does that mean there would then be two types of exception-related trees layered over each other (e.g. groups of trees, trees of groups, etc)? It makes me wonder if there's a more general tree structure that could accommodate both use cases... --Chris 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?
Hi Nathaniel, Thank you for your feedback. See a few comment below. On Mon, Apr 5, 2021 at 11:01 AM Nathaniel Smith <njs@pobox.com> wrote:
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.
People were uncomfortable about three main points: 1. In the first draft "except Exception" would not catch exception groups, this was changed in the second draft. 2. In the first draft exception groups were not subclassable, this also changed in the second draft. 3. Do we really need except*? Why is ExceptionGroup not enough? I think we replied, but this question may still be on some people's minds and if so we should continue discussing it. Was there anything else?
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.)
I'm not sure what you mean by "a placeholder". An exception group is an exception, and except* is a new syntax that helps you manipulate exception groups. I don't think the confusion you mention was due to the fact that except can catch EGs, but rather due to the fact that the simplest examples don't show you why except is not good enough, and why we need except*. Suppose we did as you suggest and made exception group a container which is not exception. Now suppose that an exception is raised in an except* block. What is the context of this exception? Do we change the requirement that an exception's context is always an exception? How do you raise an exception group if it's not an Exception? Do we go back to allowing random objects being raised?
- 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.
I don't understand this point. If you are using concurrency internally and don't want to raise EGs externally, then surely you will catch EGs, select one of the exceptions to raise and throw away all the others like in the old days when there weren't EGs (which is presumably what your caller is expecting). In other words, if you make drastic changes in a function's implementation, it is your responsibility to ensure that the old API is preserved.
- 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.
I agree with this point, my personal preference would be to have ExceptionGroup(BaseExceptionGroup) as in the first draft, with a runaway exception group being something you should fix. The current choice is a compromise for backwards compatibility, not so much with the language but with the practice of using "except Exception" to "catch almost everything, log it and move on". I don't think we can go against that at this point.
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...)
A Flat EG is in essence a "denormalized EG" - the tracebacks don't share parts, each leaf exception has its own linked list of frames as its traceback. It is easy enough to write a denormalize() function in traceback.py that constructs this from the current EG structure, if you need it (use the leaf_generator from the PEP). I'm not sure I see why we should trouble the interpreter with this. For display purposes, it is probably nicer to look at a normalized traceback where common parts are not repeated.
== 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?
I don't remember seeing this, so it wasn't deliberately dropped. It's not that we copied MultiError, or your sketches for MultiError2, and then modified them into what is now in the PEP. We learned from your experience, particularly the MutliError2 document which spells out the aspects of MultiError that didn't work well. But we designed ExceptionGroup pretty much from first principles (and then redesigned it several times as we learned our own lessons) . So you shouldn't assume that anything we didn't take from MultiError or the MultiError2 sketch was deliberately dropped. We may have just not paid attention to all of the details.
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.
It sounds like you want some way to enrich exceptions. This should be optional (we wouldn't want EG to require additional metadata for exceptions) so yeah, I agree it should sit on the leaf exception and not on the group. In that sense it's orthogonal to this PEP.
- 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?
Isn't this something that Trio/asyncio should handle, as in: if you catch Cancelled that has a __context__ then you need to put that __context__ into the exception group you are raising so that it's not lost?
- 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.
I agree that there are some edge cases in the interpreter that we will need to make a decision about. I would imagine an ExceptionGroup(StopIteration) is a bug of some sort, so we would probably want to not handle that as a StopIteration - just let the program fail and make the error obvious. ExceptionGroup(SystemExit) -- I don't have an opinion, I think it's the application's responsibility to unpack it, but we can bikeshed that.
- 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?
print_tb and print_exception are thin wrappers around these data structures, so this is implied but we can make it explicit in the PEP. Thanks again Irit
Just wanted to elaborate a little bit on StopIteration to add to Irit's reply: On Mon, Apr 5, 2021 at 9:52 AM Irit Katriel via Python-Dev <python-dev@python.org> wrote:
On Mon, Apr 5, 2021 at 11:01 AM Nathaniel Smith <njs@pobox.com> wrote:
- 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.
There's no need to do anything about ExceptionGroups potentially wrapping StopIteration or StopAsyncIteration exceptions. The point of PEP 479 was to solve a problem of one of nested frames raising a StopIteration (often by mistake) and the outer generator being stopped. That lead to some really tricky situations to debug. In our case, a rogue StopIteration wrapped in an EG would not stop a generator silently, it would do that loud and clear. As for SystemExit, we'll change "asyncio.run()" to unpack SystemExit and propagate them unwrapped, potentially allowing to dump the exception tree into a log file for later debug, if configured. Trio should do the same. Problem solved. It's important to understand that the PEP doesn't propose a magical mechanism to turn all exceptions into EGs automatically, it's up to the framework/user code how to build them and what to propagate out. In Python 3.9 you can just as well write `except BaseException: pass` and silence a SystemExit (and people do that from time to time!) Yury
On Mon, Apr 5, 2021 at 2:59 PM Chris Jerdonek <chris.jerdonek@gmail.com> wrote:
This point reminded me again of this issue in the tracker ("Problems with recursive automatic exception chaining" from 2013): https://bugs.python.org/issue18861 I'm not sure if it's exactly the same, but you can see that a couple of the later comments there talk about "exception trees" and other types of annotations.
If that issue were addressed after ExceptionGroups were introduced, does that mean there would then be two types of exception-related trees layered over each other (e.g. groups of trees, trees of groups, etc)? It makes me wonder if there's a more general tree structure that could accommodate both use cases...
--Chris
Interesting, I commented on that issue - I think we may be able to solve it without adding more trees. That said, we will have groups-of-trees/trees-of-groups. Already today, an exception plus its chained __cause__s and __context__s is the root of a binary tree of exceptions. The nodes of this tree represent the times that the exceptions were caught. An exception group is a tree where the nodes represent the times when exceptions were grouped together and raised. Irit
Irit, Yury and Guido, Thanks for submitting PEP 654 (Exception Groups and except *). The Steering Council isn’t done debating this PEP, but we wanted to give an update so everyone knows what to expect. We are mostly in favour of the PEP, but we’re worried about the timing -- not just because there is still some discussion going on, but also the time available for us to discuss the proposal, and to validate the new API once it’s accepted. We are rapidly approaching feature freeze for Python 3.10, and while we’re not particularly worried about the implementation of this PEP, we would like to see actual usage of the new API with enough time to tweak it before the official release. That is to say, we would like to see the "analysis of how exception groups will likely be used in asyncio programs" turn into "here is how exception groups are used in asyncio programs" before the final release, like in the asyncio module and the standard library (for example, with TaskGroups). There’s also a presentation scheduled in the PyCon US Language Summit, which may lead to more discussion of the PEP and perhaps evolution of its ideas. What would you and others feel about postponing the PEP until Python 3.11, getting it in as soon as the branch opens, so that potential users of the API -- asyncio or otherwise -- can provide us with practical experience to validate some of the assumptions made in the proposal, and to discover any warts or pitfalls we missed or misjudged? I don’t mean that as a slight on the PEP, which is thoughtful, thorough and excellently written, but as a practical observation of the complexity involved. Most of the PEP is more of a building block for other APIs than something many users will be directly exposed to, and we want to make sure the end result isn’t affected by unanticipated effects of these design decisions. (This would potentially also give third-party users more time to evaluate the API, although we know adoption of alpha releases isn’t big enough to make this a very strong argument for waiting.) Like I said, we’re still discussing the PEP, so this hasn’t been decided. If you or others feel strongly that the benefit of postponing is too small (or the cost too high), we’ll take that into consideration. The SC has been somewhat busy of late, and we don’t want to let our schedule be the deciding factor here, but this may come down to there just not being enough time before the feature freeze to make a thoughtful decision. For the SC, Thomas. -- Thomas Wouters <thomas@python.org> Hi! I'm an email virus! Think twice before sending your email to help me spread!
Thomas and the rest of the SC, Thank you for updating us and for your kind words about the PEP. We agree that it is safer to include this in 3.11 and not rush it into 3.10. As you say, having time to develop the integration with asyncio before finalizing the API will give us more confidence in the design. We would like to commit the code into the 3.11 development branch soon, as you suggest. We will be looking forward to the (provisional?) acceptance of the PEP by the SC so that we can do this. Best regards Irit, Yury & Guido On Thu, Apr 15, 2021 at 6:48 PM Thomas Wouters <thomas@python.org> wrote:
Irit, Yury and Guido,
Thanks for submitting PEP 654 (Exception Groups and except *). The Steering Council isn’t done debating this PEP, but we wanted to give an update so everyone knows what to expect.
We are mostly in favour of the PEP, but we’re worried about the timing -- not just because there is still some discussion going on, but also the time available for us to discuss the proposal, and to validate the new API once it’s accepted. We are rapidly approaching feature freeze for Python 3.10, and while we’re not particularly worried about the implementation of this PEP, we would like to see actual usage of the new API with enough time to tweak it before the official release. That is to say, we would like to see the "analysis of how exception groups will likely be used in asyncio programs" turn into "here is how exception groups are used in asyncio programs" before the final release, like in the asyncio module and the standard library (for example, with TaskGroups). There’s also a presentation scheduled in the PyCon US Language Summit, which may lead to more discussion of the PEP and perhaps evolution of its ideas.
What would you and others feel about postponing the PEP until Python 3.11, getting it in as soon as the branch opens, so that potential users of the API -- asyncio or otherwise -- can provide us with practical experience to validate some of the assumptions made in the proposal, and to discover any warts or pitfalls we missed or misjudged? I don’t mean that as a slight on the PEP, which is thoughtful, thorough and excellently written, but as a practical observation of the complexity involved. Most of the PEP is more of a building block for other APIs than something many users will be directly exposed to, and we want to make sure the end result isn’t affected by unanticipated effects of these design decisions. (This would potentially also give third-party users more time to evaluate the API, although we know adoption of alpha releases isn’t big enough to make this a very strong argument for waiting.)
Like I said, we’re still discussing the PEP, so this hasn’t been decided. If you or others feel strongly that the benefit of postponing is too small (or the cost too high), we’ll take that into consideration. The SC has been somewhat busy of late, and we don’t want to let our schedule be the deciding factor here, but this may come down to there just not being enough time before the feature freeze to make a thoughtful decision.
For the SC, Thomas. -- Thomas Wouters <thomas@python.org>
Hi! I'm an email virus! Think twice before sending your email to help me spread!
On Mon, Apr 5, 2021 at 9:48 AM Irit Katriel <iritkatriel@googlemail.com> wrote:
On Mon, Apr 5, 2021 at 11:01 AM Nathaniel Smith <njs@pobox.com> wrote:
- 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.)
I'm not sure what you mean by "a placeholder". An exception group is an exception, and except* is a new syntax that helps you manipulate exception groups. I don't think the confusion you mention was due to the fact that except can catch EGs, but rather due to the fact that the simplest examples don't show you why except is not good enough, and why we need except*.
Suppose we did as you suggest and made exception group a container which is not exception. Now suppose that an exception is raised in an except* block. What is the context of this exception? Do we change the requirement that an exception's context is always an exception?
How do you raise an exception group if it's not an Exception? Do we go back to allowing random objects being raised?
Ah, right. I'm talking about the conceptual/semantic level, not the implementation. For the implementation, backcompat certainly means we need some single object that can represent the whole group of exceptions, so that it can be returned by sys.exc_info(), C code can access it through the tstate, etc. But the more important question is how we explain this thing to users, and make it easy for them to do the right thing. Is the EG the exception they're trying to catch or otherwise work with? or is the EG just a detail that most users should ignore because they're focused on the leaf exceptions inside? For the concurrency use case (asyncio/trio/etc.), it's all about the leaf exceptions. ExceptionGroup([SomeError()]) and SomeError() both mean exactly the same thing, and treating them differently is basically always a bug. The ExceptionGroup object doesn't tell you anything about what went wrong, like normal exceptions do. It just tells you that there was some frame in between where the exception was raised and where it was caught where some other exception *could* have happened, but didn't. You *can* give 'except ExceptionGroup' a meaning, but that meaning doesn't make sense -- it's like saying "I want to catch any exceptions that was raised by code that *could* have raised a different exception instead" or "I want to catch all exceptions whose traceback contains an entry that's a certain subclass of TracebackType". Similarly, it doesn't make sense to attach error strings to an EG, or define custom subclasses, etc.. Ideally, plain 'except' would (somehow) handle 'EG([SomeError()])' and 'SomeError()' in exactly the same way; and if it doesn't, then using plain 'except' in concurrent code is going to usually be a bug. For other use cases, it does make sense to think of EG as a regular exception. Like, if Hypothesis wants to report that it ran some tests and there were failures, then modelling that as a single HypothesisError seems like a nice API. You never need 'except*' semantics to catch part of a HypothesisError. 'except HypothesisError' is totally sensible. The only real value you get from the PEP is that if you *don't* catch the HypothesisError, then the default traceback machinery can now automatically include info about the individual test failures. If we were starting from scratch, I don't think this would be a big conflict. I think the PEP would focus 100% on the concurrency case. That's the one that's really difficult to solve without language support. And, if you solve that, then there's a very natural way to handle the Hypothesis case too: do 'raise HypothesisError from EG(individual failures)'. That makes it explicit that HypothesisError is a single exception that should be handled as a unit, while still letting you introspect the individual exceptions and including them in the default traceback output. (In fact Trio uses this trick right now, in it's TCP connection code [1]. We report a single OSError("connection failed"), but then include all the details about each individual failure in its __cause__.) The actual problem is that we're not starting from scratch; we're trying to retrofit the concurrency-style semantics onto a system that's currently completely focused around the Hypothesis-style semantics. So the PEP ends up with a kind of mish-mash of the two approaches -- 'except*' gives you the semantics that concurrency use cases want, 'except' gives you the semantics that Hypothesis-style use cases want, and users are forced to understand all these subtleties and figure out which one is appropriate on a case-by-case basis. It's not very "one obvious way to do it". And again -- I'm *not* saying the PEP sucks or anything like that. Backcompat is a real problem, and the PEP might already be the best it can be given the tradeoffs. I'm just explaining why I want to make sure we consider all the alternatives. [1] https://github.com/python-trio/trio/blob/cb06242a78d850baaba4057f7dbd22d2a21...
- 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.
I don't understand this point. If you are using concurrency internally and don't want to raise EGs externally, then surely you will catch EGs, select one of the exceptions to raise and throw away all the others like in the old days when there weren't EGs (which is presumably what your caller is expecting). In other words, if you make drastic changes in a function's implementation, it is your responsibility to ensure that the old API is preserved.
The problem is that most of the time, even if you're using concurrency internally so multiple things *could* go wrong at once, only one thing actually *does* go wrong. So it's unfortunate if some existing code is prepared for a specific exception that it knows can be raised, that exact exception is raised... and the existing code fails to catch it because now it's wrapped in an EG.
== 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...)
A Flat EG is in essence a "denormalized EG" - the tracebacks don't share parts, each leaf exception has its own linked list of frames as its traceback.
Yes! "Denormalized" is a perfect word for what I'm talking about :-).
It is easy enough to write a denormalize() function in traceback.py that constructs this from the current EG structure, if you need it (use the leaf_generator from the PEP). I'm not sure I see why we should trouble the interpreter with this.
In the current design, once an exception is wrapped in an EG, then it can never be unwrapped, because its traceback information is spread across the individual exception + the EG tree around it. This is confusing to users ("how do I check errno?"), and makes the design more complicated (the need for topology-preserving .split(), the inability to define a sensible EG.__iter__, ...). The advantage of making the denormalized form the native form is that now the leaf exceptions would be self-contained objects like they are now, so you don't need EG nesting at all, and users can write intuitive code like: except OSError as *excs: remainder = [exc for exc in excs if exc.errno != ...] if remainder: raise ExceptionGroup(remainder)
For display purposes, it is probably nicer to look at a normalized traceback where common parts are not repeated.
Yeah, I agree; display code would want to re-normalize before printing. But now it's only the display code that needs to care about figuring out shared parts of the traceback, rather than something that has to be maintained as an invariant everywhere.
- 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?
I don't remember seeing this, so it wasn't deliberately dropped. It's not that we copied MultiError, or your sketches for MultiError2, and then modified them into what is now in the PEP. We learned from your experience, particularly the MutliError2 document which spells out the aspects of MultiError that didn't work well. But we designed ExceptionGroup pretty much from first principles (and then redesigned it several times as we learned our own lessons) . So you shouldn't assume that anything we didn't take from MultiError or the MultiError2 sketch was deliberately dropped. We may have just not paid attention to all of the details.
Huh! Then it's interesting that after all that you ended up with almost exactly the same design as MultiError v2 :-).
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.
It sounds like you want some way to enrich exceptions. This should be optional (we wouldn't want EG to require additional metadata for exceptions) so yeah, I agree it should sit on the leaf exception and not on the group. In that sense it's orthogonal to this PEP.
Well, the extra metadata would specifically be at "join" points in the traceback, which are a thing that this PEP is creating :-). And it's useful for every user of EGs, since by definition, an EG is multiplexing exceptions from multiple sources, so it's nice to tell the user which sources those are. That said, you're right, if we want to handle this by defining a new kind of traceback entry that code like Trio/asyncio/hypothesis can manually attach to exceptions, then that could be written as a separate-but-complementary PEP. In my original design, instead of defining a new kind of traceback entry, I was storing this on the EG itself, so that's why I was thinking about it needing to be part of this PEP.
- 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?
Isn't this something that Trio/asyncio should handle, as in: if you catch Cancelled that has a __context__ then you need to put that __context__ into the exception group you are raising so that it's not lost?
It's something that Trio/asyncio would handle manually, yeah, but Trio/asyncio need somewhere to put the metadata, that the traceback printing code will look at :-). Something like a __preempted__ field on exceptions, or another kind of extended traceback entry ("when passing through this frame, this exception preempted another exception"). -n -- Nathaniel J. Smith -- https://vorpus.org
On Tue, Apr 20, 2021 at 2:48 AM Nathaniel Smith <njs@pobox.com> wrote:
The problem is that most of the time, even if you're using concurrency internally so multiple things *could* go wrong at once, only one thing actually *does* go wrong. So it's unfortunate if some existing code is prepared for a specific exception that it knows can be raised, that exact exception is raised... and the existing code fails to catch it because now it's wrapped in an EG.
Yes, this was discussed at length on this list. Raising an exception group is an API-breaking change. If a function starts raising exception groups its callers need to be prepared for that. Realistically we think exception groups will be raised by new APIs. We tried and were unable to define exception group semantics for except that would be reasonable and backwards compatible. That's why we added except*.
It is easy enough to write a denormalize() function in traceback.py that constructs this from the current EG structure, if you need it (use the leaf_generator from the PEP). I'm not sure I see why we should trouble the interpreter with this.
In the current design, once an exception is wrapped in an EG, then it can never be unwrapped, because its traceback information is spread across the individual exception + the EG tree around it. This is confusing to users ("how do I check errno?"), and makes the design more complicated (the need for topology-preserving .split(), the inability to define a sensible EG.__iter__, ...). The advantage of making the denormalized form the native form is that now the leaf exceptions would be self-contained objects like they are now, so you don't need EG nesting at all, and users can write intuitive code like:
except OSError as *excs: remainder = [exc for exc in excs if exc.errno != ...] if remainder: raise ExceptionGroup(remainder)
We have this precise example in the PEP: match, rest = excs.split(lambda e: e.errno != ...) You use split() instead of iteration for that. split() preserves all __context__, __cause__ and __traceback__ information, on all leaf and non-leaf exceptions.
For display purposes, it is probably nicer to look at a normalized traceback where common parts are not repeated.
Yeah, I agree; display code would want to re-normalize before printing. But now it's only the display code that needs to care about figuring out shared parts of the traceback, rather than something that
has to be maintained as an invariant everywhere.
If you *do* want iteration, we show in the PEP how to write a leaf_generator() that gives you the leaf exceptions with their tracebacks (as a list of chunks). It is easy to copy the chunks into a single flat traceback. We didn't propose to add it to traceback.py yet because the use case is unclear but if people need it we're talking about 10-15 lines in traceback.py. So for your suggestion: Pros: 1. Those who want a denormalized traceback (if they exist) won't need to call traceback.denormalize(). Cons: 1. A significant change in the interpreter that will make it less efficient (both time and space). 2. Display code will need to normalize the traceback, which is much more complicated than denormalizing because you need to discover the shared parts. Am I missing something?
It sounds like you want some way to enrich exceptions. This should be optional (we wouldn't want EG to require additional metadata for exceptions) so yeah, I agree it should sit on the leaf exception and not on the group. In that sense it's orthogonal to this PEP.
Well, the extra metadata would specifically be at "join" points in the traceback, which are a thing that this PEP is creating :-). And it's useful for every user of EGs, since by definition, an EG is multiplexing exceptions from multiple sources, so it's nice to tell the user which sources those are.
That said, you're right, if we want to handle this by defining a new
kind of traceback entry that code like Trio/asyncio/hypothesis can manually attach to exceptions, then that could be written as a separate-but-complementary PEP. In my original design, instead of defining a new kind of traceback entry, I was storing this on the EG itself, so that's why I was thinking about it needing to be part of this PEP.
You can also create an ExceptionGroup subclass with whatever extra data you want to include. Irit
Hi Irit, reading this subthread specifically, I just got a wild idea and I couldn‘t find any related information in the PEP: Why not extending BaseException by __group__ among __cause__ and __context__? Would this reduce some of the added complexity and thus increase broader acceptance? Cheers, Sven
On 7. Apr 2021, at 20:26, Irit Katriel via Python-Dev <python-dev@python.org> wrote:
On Mon, Apr 5, 2021 at 2:59 PM Chris Jerdonek <chris.jerdonek@gmail.com> wrote:
This point reminded me again of this issue in the tracker ("Problems with recursive automatic exception chaining" from 2013): https://bugs.python.org/issue18861 I'm not sure if it's exactly the same, but you can see that a couple of the later comments there talk about "exception trees" and other types of annotations.
If that issue were addressed after ExceptionGroups were introduced, does that mean there would then be two types of exception-related trees layered over each other (e.g. groups of trees, trees of groups, etc)? It makes me wonder if there's a more general tree structure that could accommodate both use cases...
--Chris
Interesting, I commented on that issue - I think we may be able to solve it without adding more trees.
That said, we will have groups-of-trees/trees-of-groups. Already today, an exception plus its chained __cause__s and __context__s is the root of a binary tree of exceptions. The nodes of this tree represent the times that the exceptions were caught.
An exception group is a tree where the nodes represent the times when exceptions were grouped together and raised.
Irit _______________________________________________ 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/BZWELKDU... Code of Conduct: http://python.org/psf/codeofconduct/
Hi Sven, I don’t follow. What would the value of __group__ be and how would it work? Irit
On 20 Apr 2021, at 20:44, srkunze@mail.de wrote:
Hi Irit,
reading this subthread specifically, I just got a wild idea and I couldn‘t find any related information in the PEP:
Why not extending BaseException by __group__ among __cause__ and __context__?
Would this reduce some of the added complexity and thus increase broader acceptance?
Cheers, Sven
So, forgive me my relatively simple mental model about ExceptionGroup. I still try to create one for daily use. As noted in the discussion, an EG provides a way to collect exceptions from different sources and raise them as a bundle. They have no apparent relation up until this point in time (for whatever reason they have been separate and for whatever reason they are bundled now). The result would be a tree graph in any case. A usual datastructure for a tree is to store all child nodes at the parent node. That was the idea behind the content of BaseException.__group__: it’s the list of child exceptions bundled at a specific point in time and raise as such a bundle. So all exceptions could become EGs with the additional semantics you‘ve described in the PEP. Illustrative Example:
bundle_exc.__group__ [IOError(123), RuntimerError(‘issue somewhere’)]
I was wondering what of the PEP could be removed to make it simpler and more acceptable/less confusing (also looking at reactions from Twitter etc.) and I found these additional classes to be a part of it. Additionally, I fail to see how to access these bundled exceptions in an easy manner like __cause__ and __context__. (As the PEP also referring to them). So, I removed the classes and added a regular attribute. The reason I brought this up what the section “rejected ideas” didn’t showed anything in this direction (or I managed to missed that). Sven
On 20. Apr 2021, at 22:05, Irit Katriel <iritkatriel@yahoo.com> wrote:
Hi Sven,
I don’t follow. What would the value of __group__ be and how would it work?
Irit
On 20 Apr 2021, at 20:44, srkunze@mail.de wrote: Hi Irit, reading this subthread specifically, I just got a wild idea and I couldn‘t find any related information in the PEP: Why not extending BaseException by __group__ among __cause__ and __context__? Would this reduce some of the added complexity and thus increase broader acceptance? Cheers, Sven
I don’t see what this simplifies. We still need to implement split, and to worry about wrapping or not wrapping BaseExceptions and we still need to define exception handling semantics (except/except*).
On 20 Apr 2021, at 22:12, srkunze@mail.de wrote:
So, forgive me my relatively simple mental model about ExceptionGroup. I still try to create one for daily use.
As noted in the discussion, an EG provides a way to collect exceptions from different sources and raise them as a bundle. They have no apparent relation up until this point in time (for whatever reason they have been separate and for whatever reason they are bundled now). The result would be a tree graph in any case.
A usual datastructure for a tree is to store all child nodes at the parent node.
That was the idea behind the content of BaseException.__group__: it’s the list of child exceptions bundled at a specific point in time and raise as such a bundle. So all exceptions could become EGs with the additional semantics you‘ve described in the PEP.
Illustrative Example:
bundle_exc.__group__ [IOError(123), RuntimerError(‘issue somewhere’)]
I was wondering what of the PEP could be removed to make it simpler and more acceptable/less confusing (also looking at reactions from Twitter etc.) and I found these additional classes to be a part of it. Additionally, I fail to see how to access these bundled exceptions in an easy manner like __cause__ and __context__. (As the PEP also referring to them). So, I removed the classes and added a regular attribute.
The reason I brought this up what the section “rejected ideas” didn’t showed anything in this direction (or I managed to missed that).
Sven
On 20. Apr 2021, at 22:05, Irit Katriel <iritkatriel@yahoo.com> wrote:
Hi Sven,
I don’t follow. What would the value of __group__ be and how would it work?
Irit
On 20 Apr 2021, at 20:44, srkunze@mail.de wrote: Hi Irit, reading this subthread specifically, I just got a wild idea and I couldn‘t find any related information in the PEP: Why not extending BaseException by __group__ among __cause__ and __context__? Would this reduce some of the added complexity and thus increase broader acceptance? Cheers, Sven
Removing two concepts and preserving semantics simplifies the matter for users. People need less to memorize and less to learn. Or am I missing something here? Couldn’t we achieve our goal without these two new classes? Splitting, wrapping and exception handling semantics are perfectly fine and serve their purposes. So, of course this still needs to be implemented.
On 20. Apr 2021, at 23:26, Irit Katriel <iritkatriel@googlemail.com> wrote:
I don’t see what this simplifies. We still need to implement split, and to worry about wrapping or not wrapping BaseExceptions and we still need to define exception handling semantics (except/except*).
On 20 Apr 2021, at 22:12, srkunze@mail.de wrote:
So, forgive me my relatively simple mental model about ExceptionGroup. I still try to create one for daily use.
As noted in the discussion, an EG provides a way to collect exceptions from different sources and raise them as a bundle. They have no apparent relation up until this point in time (for whatever reason they have been separate and for whatever reason they are bundled now). The result would be a tree graph in any case.
A usual datastructure for a tree is to store all child nodes at the parent node.
That was the idea behind the content of BaseException.__group__: it’s the list of child exceptions bundled at a specific point in time and raise as such a bundle. So all exceptions could become EGs with the additional semantics you‘ve described in the PEP.
Illustrative Example:
bundle_exc.__group__ [IOError(123), RuntimerError(‘issue somewhere’)]
I was wondering what of the PEP could be removed to make it simpler and more acceptable/less confusing (also looking at reactions from Twitter etc.) and I found these additional classes to be a part of it. Additionally, I fail to see how to access these bundled exceptions in an easy manner like __cause__ and __context__. (As the PEP also referring to them). So, I removed the classes and added a regular attribute.
The reason I brought this up what the section “rejected ideas” didn’t showed anything in this direction (or I managed to missed that).
Sven
On 20. Apr 2021, at 22:05, Irit Katriel <iritkatriel@yahoo.com> wrote:
Hi Sven,
I don’t follow. What would the value of __group__ be and how would it work?
Irit
On 20 Apr 2021, at 20:44, srkunze@mail.de wrote: Hi Irit, reading this subthread specifically, I just got a wild idea and I couldn‘t find any related information in the PEP: Why not extending BaseException by __group__ among __cause__ and __context__? Would this reduce some of the added complexity and thus increase broader acceptance? Cheers, Sven
On Wed, Apr 21, 2021 at 11:50 AM srkunze@mail.de <srkunze@mail.de> wrote:
Removing two concepts and preserving semantics simplifies the matter for users. People need less to memorize and less to learn.
Or am I missing something here? Couldn’t we achieve our goal without these two new classes?
No, we can't. What you are proposing would make it very hard for users to understand at a glance if what you have in an innocently looking `except Exception` is correct or not. In my async/await code I'd have to always check the `__group__` attribute to make sure it's not an exception group in disguise. So while you're "simplifying" the proposal by removing a couple of types, you're complicating it in all other places. Besides, I don't think that adding the ExceptionGroup type is a controversial idea that needs any simplification. Yury
On Tue, Apr 20, 2021 at 2:15 PM srkunze@mail.de <srkunze@mail.de> wrote:
So, forgive me my relatively simple mental model about ExceptionGroup. I still try to create one for daily use.
As noted in the discussion, an EG provides a way to collect exceptions from different sources and raise them as a bundle. They have no apparent relation up until this point in time (for whatever reason they have been separate and for whatever reason they are bundled now). The result would be a tree graph in any case.
A usual datastructure for a tree is to store all child nodes at the parent node.
That was the idea behind the content of BaseException.__group__: it’s the list of child exceptions bundled at a specific point in time and raise as such a bundle. So all exceptions could become EGs with the additional semantics you‘ve described in the PEP.
Illustrative Example:
bundle_exc.__group__ [IOError(123), RuntimerError(‘issue somewhere’)]
I was wondering what of the PEP could be removed to make it simpler and more acceptable/less confusing (also looking at reactions from Twitter etc.) and I found these additional classes to be a part of it. Additionally, I fail to see how to access these bundled exceptions in an easy manner like __cause__ and __context__. (As the PEP also referring to them). So, I removed the classes and added a regular attribute.
This seems more confusing to me too. Instead of having a single ExceptionGroup class, you're suggesting that all exceptions should effectively become ExceptionGroups. I know what ExceptionGroup([ValueError]) means -- it means there was a ValueError, and it got put in a group, so it should probably be handled the same way as a ValueError. I have no idea what a KeyError([ValueError]) would mean. Is that a KeyError or a ValueError? Adding flexibility doesn't necessarily make things simpler :-) -n -- Nathaniel J. Smith -- https://vorpus.org
On Tue, Apr 20, 2021 at 3:15 AM Irit Katriel <iritkatriel@googlemail.com> wrote:
On Tue, Apr 20, 2021 at 2:48 AM Nathaniel Smith <njs@pobox.com> wrote:
The problem is that most of the time, even if you're using concurrency internally so multiple things *could* go wrong at once, only one thing actually *does* go wrong. So it's unfortunate if some existing code is prepared for a specific exception that it knows can be raised, that exact exception is raised... and the existing code fails to catch it because now it's wrapped in an EG.
Yes, this was discussed at length on this list. Raising an exception group is an API-breaking change. If a function starts raising exception groups its callers need to be prepared for that. Realistically we think exception groups will be raised by new APIs. We tried and were unable to define exception group semantics for except that would be reasonable and backwards compatible. That's why we added except*.
Sure. This was in my list of reasons why the backwards compatibility tradeoffs are forcing us into awkward compromises. I only elaborated on it b/c in your last email you said you didn't understand why this was a problem :-). And except* is definitely useful. But I think there are options for 'except' that haven't been considered fully. Saying that you have to make a new API every time you start using concurrency inside a function is extremely restrictive. The whole point of structured concurrency (and structured programming in general) is that function callers don't need to know about the control flow decisions inside a function. So right now, the EG proposal is like saying that if you every have a function that doesn't contain a 'for' loop, and then you want to add a 'for' loop, then you have to define a whole new API instead of modifying the existing one. I absolutely get why the proposal looks like that. I'm just making the point that we should make sure we've exhausted all our options before settling for that as a compromise.
It is easy enough to write a denormalize() function in traceback.py that constructs this from the current EG structure, if you need it (use the leaf_generator from the PEP). I'm not sure I see why we should trouble the interpreter with this.
In the current design, once an exception is wrapped in an EG, then it can never be unwrapped, because its traceback information is spread across the individual exception + the EG tree around it. This is confusing to users ("how do I check errno?"), and makes the design more complicated (the need for topology-preserving .split(), the inability to define a sensible EG.__iter__, ...). The advantage of making the denormalized form the native form is that now the leaf exceptions would be self-contained objects like they are now, so you don't need EG nesting at all, and users can write intuitive code like:
except OSError as *excs: remainder = [exc for exc in excs if exc.errno != ...] if remainder: raise ExceptionGroup(remainder)
We have this precise example in the PEP: match, rest = excs.split(lambda e: e.errno != ...)
You use split() instead of iteration for that. split() preserves all __context__, __cause__ and __traceback__ information, on all leaf and non-leaf exceptions.
Well, yeah, I know, I'm the one who invented split() :-). My point was to compare these two options: in the flat EG example, most Python users could write and read that code without knowing anything except "there can be multiple exceptions now". It's all old, well-known constructs used in the obvious way. For the .split() version, you have to write a lambda (which is allowed to access parts of the exception object, but not all of it!), and use this idiosyncratic method that only makes sense if you know about EG tree structures. That's a lot more stuff that users have to understand and hold in their head. Or here's another example. Earlier you suggested:
If you are using concurrency internally and don't want to raise EGs externally, then surely you will catch EGs, select one of the exceptions to raise and throw away all the others
But with nested EGs, this is difficult, because you *can't* just pull out one of the exceptions to raise. The leaf exceptions aren't standalone objects; you need some obscure traceback manipulation to do this. I guarantee that users will get this wrong, even if we provide the tools, because the explanation about when and why you need the tools is complicated and people won't internalize it. With flat EGs, this is trivial: it's just `raise ExceptionGroup[the_selected_exception_index]`.
For display purposes, it is probably nicer to look at a normalized traceback where common parts are not repeated.
Yeah, I agree; display code would want to re-normalize before printing. But now it's only the display code that needs to care about figuring out shared parts of the traceback, rather than something that
has to be maintained as an invariant everywhere.
If you *do* want iteration, we show in the PEP how to write a leaf_generator() that gives you the leaf exceptions with their tracebacks (as a list of chunks). It is easy to copy the chunks into a single flat traceback. We didn't propose to add it to traceback.py yet because the use case is unclear but if people need it we're talking about 10-15 lines in traceback.py.
So for your suggestion:
Pros: 1. Those who want a denormalized traceback (if they exist) won't need to call traceback.denormalize().
Cons: 1. A significant change in the interpreter that will make it less efficient (both time and space).
In terms of interpreter complexity, the straightforward implementation would be pretty simple. Whenever the exception handling code goes to write back a traceback to an exception's __traceback__ attribute, check if the exception is an ExceptionGroup; if so, then loop over each exception in the group and push a clone of the traceback onto that exception's __traceback__. It takes code of course, everything does, but I'd be surprised if it was more than, say, 20 lines of C. I think that's a pretty good tradeoff to get a simpler user-level language semantics. Python is supposed to fit in your head :-). Time/space efficiency is interesting, because yeah, with the nested EG approach then common parts of the traceback are stored once, while the flat EG approach needs to store a separate copy for each exception. But: - The extra cost is 1 traceback object per common stack frame, per exception. So the worst case is if you have, like, a 1000 exceptions in a single group, and then that group travels together for 1000 stack frames. But... is that realistic? Very few programs recurse down 1000 levels and *then* spawn a bunch of tasks. And very few programs have 1000 tasks that all fail simultaneously with different exceptions. And even if that did happen, 1000*1000 = 1 million traceback objects take ~56 megabytes of RAM and ~90 ms to instantiate on my laptop, which is not a big deal. [2] - This is assuming the most simple/naive implementation, where we simply duplicate every traceback entry onto every exception. If it turns out to be unacceptable, there are lots of tricks we could use to speed it up, e.g. by updating __traceback__ lazily or letting traceback entry objects be shared between tracebacks. [2] Estimated based on 'sys.getsizeof(tb_object)' and '%timeit types.TracebackType(None, frame, 0, 0)'.
2. Display code will need to normalize the traceback, which is much more complicated than denormalizing because you need to discover the shared parts.
Am I missing something?
I think re-normalizing is very cheap and simple? It's just a common prefix problem. Given a tree-in-progress and a traceback, walk down each of them in parallel until you run out of matching entries, then insert a branch node and append the rest of the traceback.
It sounds like you want some way to enrich exceptions. This should be optional (we wouldn't want EG to require additional metadata for exceptions) so yeah, I agree it should sit on the leaf exception and not on the group. In that sense it's orthogonal to this PEP.
Well, the extra metadata would specifically be at "join" points in the traceback, which are a thing that this PEP is creating :-). And it's useful for every user of EGs, since by definition, an EG is multiplexing exceptions from multiple sources, so it's nice to tell the user which sources those are.
That said, you're right, if we want to handle this by defining a new kind of traceback entry that code like Trio/asyncio/hypothesis can manually attach to exceptions, then that could be written as a separate-but-complementary PEP. In my original design, instead of defining a new kind of traceback entry, I was storing this on the EG itself, so that's why I was thinking about it needing to be part of this PEP.
You can also create an ExceptionGroup subclass with whatever extra data you want to include.
Unfortunately, that doesn't work either, because this is data that should be included in traceback displays, similar to __cause__ and __context__... -n -- Nathaniel J. Smith -- https://vorpus.org
On Wed, Apr 21, 2021 at 11:22 PM Nathaniel Smith <njs@pobox.com> wrote:
Saying that you have to make a new API every time you start using concurrency inside a function is extremely restrictive.
You don't need a new API when you start using concurrency inside a function. You need a new API when you start raising exception groups from a function because that changes the function's external behaviour. -1 from me on the denormalized traceback idea. I've said what I had to say about it in my previous email. Cheers Irit
On Wed, Apr 21, 2021 at 3:26 PM Nathaniel Smith <njs@pobox.com> wrote:
On Tue, Apr 20, 2021 at 3:15 AM Irit Katriel <iritkatriel@googlemail.com> wrote:
On Tue, Apr 20, 2021 at 2:48 AM Nathaniel Smith <njs@pobox.com> wrote:
The problem is that most of the time, even if you're using concurrency internally so multiple things *could* go wrong at once, only one thing actually *does* go wrong. So it's unfortunate if some existing code is prepared for a specific exception that it knows can be raised, that exact exception is raised... and the existing code fails to catch it because now it's wrapped in an EG.
Yes, this was discussed at length on this list. Raising an exception group is an API-breaking change. If a function starts raising exception groups its callers need to be prepared for that. Realistically we think exception groups will be raised by new APIs. We tried and were unable to define exception group semantics for except that would be reasonable and backwards compatible. That's why we added except*.
Sure. This was in my list of reasons why the backwards compatibility tradeoffs are forcing us into awkward compromises. I only elaborated on it b/c in your last email you said you didn't understand why this was a problem :-). And except* is definitely useful. But I think there are options for 'except' that haven't been considered fully.
Do you have any suggestions, or are you just telling us to think harder? Because we've already thought as hard as we could and within all the constraints (backwards compatibility and otherwise) we just couldn't think of a better one.
Saying that you have to make a new API every time you start using concurrency inside a function is extremely restrictive. The whole point of structured concurrency (and structured programming in general) is that function callers don't need to know about the control flow decisions inside a function. So right now, the EG proposal is like saying that if you every have a function that doesn't contain a 'for' loop, and then you want to add a 'for' loop, then you have to define a whole new API instead of modifying the existing one.
That analogy doesn't fly. If "what exceptions can this raise" is not part of the function spec (so users who catch exceptions are already flying blind or guessing based on what they've seen happen in the past), then you can certainly switch to using EGs, and this *is* like adding a for-loop. But if the exceptions that a function raises *are* part of its spec, and users are catching them, then switching to concurrency (without catching and converting EGs in the function) *is* an API change, akin to taking a function that returns an int and changing it to (sometimes?) returning a list of ints.
I absolutely get why the proposal looks like that. I'm just making the point that we should make sure we've exhausted all our options before settling for that as a compromise.
But if you're not able to make constructive proposals I have to conclude that we *have* exhausted our options, and we should move on.
It is easy enough to write a denormalize() function in traceback.py that constructs this from the current EG structure, if you need it (use the leaf_generator from the PEP). I'm not sure I see why we should trouble the interpreter with this.
In the current design, once an exception is wrapped in an EG, then it can never be unwrapped, because its traceback information is spread across the individual exception + the EG tree around it. This is confusing to users ("how do I check errno?"), and makes the design more complicated (the need for topology-preserving .split(), the inability to define a sensible EG.__iter__, ...). The advantage of making the denormalized form the native form is that now the leaf exceptions would be self-contained objects like they are now, so you don't need EG nesting at all, and users can write intuitive code like:
except OSError as *excs: remainder = [exc for exc in excs if exc.errno != ...] if remainder: raise ExceptionGroup(remainder)
We have this precise example in the PEP: match, rest = excs.split(lambda e: e.errno != ...)
You use split() instead of iteration for that. split() preserves all __context__, __cause__ and __traceback__ information, on all leaf and non-leaf exceptions.
Well, yeah, I know, I'm the one who invented split() :-). My point was to compare these two options: in the flat EG example, most Python users could write and read that code without knowing anything except "there can be multiple exceptions now". It's all old, well-known constructs used in the obvious way.
For the .split() version, you have to write a lambda (which is allowed to access parts of the exception object, but not all of it!), and use this idiosyncratic method that only makes sense if you know about EG tree structures. That's a lot more stuff that users have to understand and hold in their head.
But it *is* a tree structure. The interior nodes of the tree have a meaning: they represent join points where exceptions were grouped together, and the selection process needs to traverse the tree.
Or here's another example. Earlier you suggested:
If you are using concurrency internally and don't want to raise EGs externally, then surely you will catch EGs, select one of the exceptions to raise and throw away all the others
But with nested EGs, this is difficult, because you *can't* just pull out one of the exceptions to raise. The leaf exceptions aren't standalone objects; you need some obscure traceback manipulation to do this. I guarantee that users will get this wrong, even if we provide the tools, because the explanation about when and why you need the tools is complicated and people won't internalize it.
With flat EGs, this is trivial: it's just `raise ExceptionGroup[the_selected_exception_index]`.
But it's generally wrong to raise just one of the exceptions anyway, so you want to make it easier to get a wrong result? I'd rather make this an option for asyncio.gather() or its TaskGroup equivalent, then the user doesn't have to do anything. gather() already has as its default behavior to raise the first exception it encounters, which seems good enough.
For display purposes, it is probably nicer to look at a normalized traceback where common parts are not repeated.
Yeah, I agree; display code would want to re-normalize before printing. But now it's only the display code that needs to care about figuring out shared parts of the traceback, rather than something that
has to be maintained as an invariant everywhere.
If you *do* want iteration, we show in the PEP how to write a leaf_generator() that gives you the leaf exceptions with their tracebacks (as a list of chunks). It is easy to copy the chunks into a single flat traceback. We didn't propose to add it to traceback.py yet because the use case is unclear but if people need it we're talking about 10-15 lines in traceback.py.
So for your suggestion:
Pros: 1. Those who want a denormalized traceback (if they exist) won't need to call traceback.denormalize().
Cons: 1. A significant change in the interpreter that will make it less efficient (both time and space).
In terms of interpreter complexity, the straightforward implementation would be pretty simple. Whenever the exception handling code goes to write back a traceback to an exception's __traceback__ attribute, check if the exception is an ExceptionGroup; if so, then loop over each exception in the group and push a clone of the traceback onto that exception's __traceback__. It takes code of course, everything does, but I'd be surprised if it was more than, say, 20 lines of C. I think that's a pretty good tradeoff to get a simpler user-level language semantics. Python is supposed to fit in your head :-).
Time/space efficiency is interesting, because yeah, with the nested EG approach then common parts of the traceback are stored once, while the flat EG approach needs to store a separate copy for each exception. But:
- The extra cost is 1 traceback object per common stack frame, per exception. So the worst case is if you have, like, a 1000 exceptions in a single group, and then that group travels together for 1000 stack frames. But... is that realistic? Very few programs recurse down 1000 levels and *then* spawn a bunch of tasks. And very few programs have 1000 tasks that all fail simultaneously with different exceptions. And even if that did happen, 1000*1000 = 1 million traceback objects take ~56 megabytes of RAM and ~90 ms to instantiate on my laptop, which is not a big deal. [2]
- This is assuming the most simple/naive implementation, where we simply duplicate every traceback entry onto every exception. If it turns out to be unacceptable, there are lots of tricks we could use to speed it up, e.g. by updating __traceback__ lazily or letting traceback entry objects be shared between tracebacks.
[2] Estimated based on 'sys.getsizeof(tb_object)' and '%timeit types.TracebackType(None, frame, 0, 0)'.
The real cost here is that we would need a new "TracebackGroup" concept, since the internal data structures and APIs keep the traceback chain and the exception object separated until the exception is caught. In our early design stages we actually explored this and the complexity of the data structures was painful. We eventually realized that we didn't need this concept at all, and the result is much clearer, despite what you seem to think.
2. Display code will need to normalize the traceback, which is much more complicated than denormalizing because you need to discover the shared parts.
Am I missing something?
I think re-normalizing is very cheap and simple? It's just a common prefix problem. Given a tree-in-progress and a traceback, walk down each of them in parallel until you run out of matching entries, then insert a branch node and append the rest of the traceback.
It sounds like you want some way to enrich exceptions. This should be
optional (we wouldn't want EG to require additional metadata for exceptions) so yeah, I agree it should sit on the leaf exception and not on the group. In that sense it's orthogonal to this PEP.
Well, the extra metadata would specifically be at "join" points in the traceback, which are a thing that this PEP is creating :-). And it's useful for every user of EGs, since by definition, an EG is multiplexing exceptions from multiple sources, so it's nice to tell the user which sources those are.
That said, you're right, if we want to handle this by defining a new kind of traceback entry that code like Trio/asyncio/hypothesis can manually attach to exceptions, then that could be written as a separate-but-complementary PEP. In my original design, instead of defining a new kind of traceback entry, I was storing this on the EG itself, so that's why I was thinking about it needing to be part of this PEP.
You can also create an ExceptionGroup subclass with whatever extra data you want to include.
Unfortunately, that doesn't work either, because this is data that should be included in traceback displays, similar to __cause__ and __context__...
Fortunately in the flat design there wouldn't be a need for such metadata anyway... -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-change-the-world/>
On Wed, Apr 21, 2021 at 4:50 PM Guido van Rossum <guido@python.org> wrote:
On Wed, Apr 21, 2021 at 3:26 PM Nathaniel Smith <njs@pobox.com> wrote:
Sure. This was in my list of reasons why the backwards compatibility tradeoffs are forcing us into awkward compromises. I only elaborated on it b/c in your last email you said you didn't understand why this was a problem :-). And except* is definitely useful. But I think there are options for 'except' that haven't been considered fully.
Do you have any suggestions, or are you just telling us to think harder? Because we've already thought as hard as we could and within all the constraints (backwards compatibility and otherwise) we just couldn't think of a better one.
The main possibility that I don't think we've examined fully is to make 'except' blocks fire multiple times when there are multiple exceptions. We ruled it out early b/c it's incompatible with nested EGs, but if flat EGs are better anyway, then the balance shifts around and it might land somewhere different. it's a tricky discussion though, b/c both the current proposal and the alternative have very complex implications and downsides. So we probably shouldn't get too distracted by that until after the flat vs nested discussion has settled down more. I'm not trying to filibuster here -- I really want some form of EGs to land. I think python has the potential to be the most elegant and accessible language around for writing concurrent programs, and EGs are a key part of that. I don't want to fight about anything; I just want to work together to make sure we have a full picture of our options, so we can be confident we're making the best choice.
The real cost here is that we would need a new "TracebackGroup" concept, since the internal data structures and APIs keep the traceback chain and the exception object separated until the exception is caught. In our early design stages we actually explored this and the complexity of the data structures was painful. We eventually realized that we didn't need this concept at all, and the result is much clearer, despite what you seem to think.
I'm not talking about TracebackGroups (at least, I think I'm not?). I think it can be done with exactly our current data structures, nothing new. - When an EG is raised, build the traceback for just that EG while it's unwinding. This means if any C code peeks at exc_info while it's in flight, it'll only see the current branch of the traceback tree, but that seems fine. - When the exception is caught and we go to write back the traceback to its __traceback__ attribute, instead "peek through" the EG and append the built-up traceback entries onto each of the constituent exceptions. You could get cleverer for efficiency, but that basic concept seems pretty simple and viable to me. What am I missing? -n -- Nathaniel J. Smith -- https://vorpus.org
On Fri, Apr 23, 2021 at 6:25 PM Nathaniel Smith <njs@pobox.com> wrote:
On Wed, Apr 21, 2021 at 4:50 PM Guido van Rossum <guido@python.org> wrote:
On Wed, Apr 21, 2021 at 3:26 PM Nathaniel Smith <njs@pobox.com> wrote:
Sure. This was in my list of reasons why the backwards compatibility tradeoffs are forcing us into awkward compromises. I only elaborated on it b/c in your last email you said you didn't understand why this was a problem :-). And except* is definitely useful. But I think there are options for 'except' that haven't been considered fully.
Do you have any suggestions, or are you just telling us to think harder? Because we've already thought as hard as we could and within all the constraints (backwards compatibility and otherwise) we just couldn't think of a better one.
The main possibility that I don't think we've examined fully is to make 'except' blocks fire multiple times when there are multiple exceptions.
Vanilla except blocks? Not sure if I'm misunderstanding, but that could cause some problems. Consider: trn.begin() try: ... except BaseException: trn.rollback() raise else: trn.commit() What happens if multiple exceptions get raised? Will the transaction be rolled back more than once? What gets reraised? If I'm completely misunderstanding you here, my apologies. ChrisA
On Fri, Apr 23, 2021 at 9:22 AM Nathaniel Smith <njs@pobox.com> wrote:
On Wed, Apr 21, 2021 at 4:50 PM Guido van Rossum <guido@python.org> wrote:
On Wed, Apr 21, 2021 at 3:26 PM Nathaniel Smith <njs@pobox.com> wrote:
Sure. This was in my list of reasons why the backwards compatibility tradeoffs are forcing us into awkward compromises. I only elaborated on it b/c in your last email you said you didn't understand why this was a problem :-). And except* is definitely useful. But I think there are options for 'except' that haven't been considered fully.
Do you have any suggestions, or are you just telling us to think harder? Because we've already thought as hard as we could and within all the constraints (backwards compatibility and otherwise) we just couldn't think of a better one.
The main possibility that I don't think we've examined fully is to make 'except' blocks fire multiple times when there are multiple exceptions. We ruled it out early b/c it's incompatible with nested EGs, but if flat EGs are better anyway, then the balance shifts around and it might land somewhere different. it's a tricky discussion though, b/c both the current proposal and the alternative have very complex implications and downsides. So we probably shouldn't get too distracted by that until after the flat vs nested discussion has settled down more.
I'm not trying to filibuster here -- I really want some form of EGs to land.
I'm very glad to hear that. It's been hard to know where you stand, because you didn't even decline our invitation in October to work together on this, and several later invitations to look at the implementation and try it with Trio -- you just didn't reply. The only responses I saw were public, on this list, taking us right back to the drawing board (for example - the suggestion you mention in the previous paragraph as not-sufficiently-explored, actually appears in the rejected ideas section of the PEP, and the reason for rejection is not that it's incompatible with nesting). So thank you for clarifying that you are, ultimately, supportive of our efforts. We do realize that we didn't adequately credit you for the contributions of your 2018 work to this PEP, and have now added an acknowledgements section for that purpose. Apologies for that. I'm confused about the flattening suggestion - above you talk about "flat EG", but below about tracebacks. It's not clear to me whether you want EG to be flat (ie no nesting of EGs) or just the traceback to be flat (but you can still have a nested EG). I also don't know what problem you are trying to solve with this. Fortunately we're not at the whiteboard stage anymore, we have a fully working implementation and you can use it to demonstrate any problems you see with what we did. Irit
I think python has the potential to be the most elegant and accessible language around for writing concurrent programs, and EGs are a key part of that. I don't want to fight about anything; I just want to work together to make sure we have a full picture of our options, so we can be confident we're making the best choice.
The real cost here is that we would need a new "TracebackGroup" concept, since the internal data structures and APIs keep the traceback chain and the exception object separated until the exception is caught. In our early design stages we actually explored this and the complexity of the data structures was painful. We eventually realized that we didn't need this concept at all, and the result is much clearer, despite what you seem to think.
I'm not talking about TracebackGroups (at least, I think I'm not?). I think it can be done with exactly our current data structures, nothing new.
- When an EG is raised, build the traceback for just that EG while it's unwinding. This means if any C code peeks at exc_info while it's in flight, it'll only see the current branch of the traceback tree, but that seems fine. - When the exception is caught and we go to write back the traceback to its __traceback__ attribute, instead "peek through" the EG and append the built-up traceback entries onto each of the constituent exceptions.
You could get cleverer for efficiency, but that basic concept seems pretty simple and viable to me. What am I missing?
-n
-- Nathaniel J. Smith -- https://vorpus.org
On Fri, Apr 23, 2021 at 2:45 AM Chris Angelico <rosuav@gmail.com> wrote:
On Fri, Apr 23, 2021 at 6:25 PM Nathaniel Smith <njs@pobox.com> wrote:
The main possibility that I don't think we've examined fully is to make 'except' blocks fire multiple times when there are multiple exceptions.
Vanilla except blocks? Not sure if I'm misunderstanding, but that could cause some problems. Consider:
trn.begin() try: ... except BaseException: trn.rollback() raise else: trn.commit()
What happens if multiple exceptions get raised? Will the transaction be rolled back more than once? What gets reraised?
If I'm completely misunderstanding you here, my apologies.
Yeah, you've understood correctly, and you see why I wrote "both the current proposal and the alternative have very complex implications and downsides" :-) A lot depends on the details, too. One possible design is "in general, vanilla except blocks will trigger multiple times, but as a special case, except: and except BaseException: will only fire once, receiving the whole ExceptionGroup as a single exception". (Rationale for making a special case: if you're saying "catch anything, I don't care what", then you obviously didn't intend to do anything in particular with those exceptions, plus this is the only case where the types work.) In that design, your example becomes correct again. Some other interesting cases: # Filters out all OSError's that match the condition, while letting all other OSError's continue try: ... except OSError as exc: if exc.errno != errno.EWHATEVER: raise # If this gets an ExceptionGroup([ValueError, KeyboardInterrupt]), then it logs the ValueError and # then exits with the selected error code try: ... except Exception as exc: logger.log_exception(exc) except KeyboardInterrupt: sys.exit(17) OTOH you can still come up with cases that aren't handled correctly. If you have old 'except' code mixed with new code raising ExceptionGroup, then there's no way to make that do the right thing in *all* cases. By definition, when the 'except' code was written, the author wasn't thinking about ExceptionGroups! So the question is which combination of semantics ends up doing the least harm on average across all sorts of different real-world cases... -n -- Nathaniel J. Smith -- https://vorpus.org
On Mon, Apr 26, 2021 at 2:27 PM Nathaniel Smith <njs@pobox.com> wrote:
Yeah, you've understood correctly, and you see why I wrote "both the current proposal and the alternative have very complex implications and downsides" :-)
[chomp lots of very helpful summarizing]
Gotcha, thanks! ChrisA
On Fri, Apr 23, 2021 at 4:08 AM Irit Katriel <iritkatriel@googlemail.com> wrote:
On Fri, Apr 23, 2021 at 9:22 AM Nathaniel Smith <njs@pobox.com> wrote:
I'm not trying to filibuster here -- I really want some form of EGs to land.
I'm very glad to hear that. It's been hard to know where you stand, because you didn't even decline our invitation in October to work together on this, and several later invitations to look at the implementation and try it with Trio -- you just didn't reply. The only responses I saw were public, on this list, taking us right back to the drawing board (for example - the suggestion you mention in the previous paragraph as not-sufficiently-explored, actually appears in the rejected ideas section of the PEP, and the reason for rejection is not that it's incompatible with nesting). So thank you for clarifying that you are, ultimately, supportive of our efforts.
Yes, I apologize again for the radio silence there -- I had real world stuff that left me with no cope for open-source. I, uh, was feeling guilty about not getting back to you the whole time, if that helps? Probably that doesn't help. My memory is that in our initial discussions, I suggested having each 'except Blah as exc' clause be executed once, receiving an ExceptionGroup containing all the Blah exceptions. Guido pointed out that this broke typing -- 'exc' would not longer have type 'Blah' -- and I was like... okay yeah that's a fatal flaw, never mind. And I never seriously raised the 'execute a single clause multiple times' option, because of the issue where in the nested design, taking individual exceptions out of an ExceptionGroup breaks tracebacks. Looking at the relevant section of the PEP again [1], it notes the same fatal flaw with my first suggestion, and then says that the multiple-except-executions option should be rejected because users have written code like 'except SomeError: ...' with the expectation that the 'except' clause would run exactly once. That's definitely true, and it's a downside of the multiple-except-executions approach, but I don't think it's convincing enough to rule this out on its own. The problem is, *all* our options for how 'except' should interact with ExceptionGroups will somehow break previous expectations. Concretely: imagine you have a pre-existing 'except SomeError', and some new code inside the 'try' block raises some number of 'SomeError's wrapped in an ExceptionGroup. There are three options: - Execute the 'except' block multiple times. This breaks the expectation that it should be executed at most once. - Execute the 'except' block exactly once. But if there are multiple SomeError's, this require they be grouped and delivered as a single exception, which breaks typing. - Execute the 'except' block zero times. This is what the current PEP chooses, and breaks the expectation that 'except SomeError' should catch 'SomeError'. So we have to pick our poison. [1] https://www.python.org/dev/peps/pep-0654/#extend-except-to-handle-exception-...
We do realize that we didn't adequately credit you for the contributions of your 2018 work to this PEP, and have now added an acknowledgements section for that purpose. Apologies for that.
Oh, that didn't bother me at all, but thanks :-). And I'm sorry if I was denying you credit for things that you had actually independently re-invented.
I'm confused about the flattening suggestion - above you talk about "flat EG", but below about tracebacks. It's not clear to me whether you want EG to be flat (ie no nesting of EGs) or just the traceback to be flat (but you can still have a nested EG).
Hmm, I was thinking about making both of them flat, so no nested EGs. In all my designs, the only reason I ever had nesting was because I couldn't figure out any other way to make the tracebacks work. Do you have some other motivation for wanting nesting? If so that would be interesting, because it might point to why we're talking past each other and help us understand the problem better...
I also don't know what problem you are trying to solve with this.
I'm not saying that there's some fatal problem with the current PEP. (In my first message I explicitly said that it would be an improvement over the status quo :-).) But I think that nesting will be really counterintuitive/confusing for users in some ways. And concurrency APIs will be offputting if they force you to use a different special form of 'except' all the time. Basically the 'flat' version might be a lot more ergonomic, and that's important for a language like Python. -n -- Nathaniel J. Smith -- https://vorpus.org
On Wed, Apr 28, 2021 at 8:53 PM Nathaniel Smith <njs@pobox.com> wrote:
Looking at the relevant section of the PEP again [1], it notes the same fatal flaw with my first suggestion, and then says that the multiple-except-executions option should be rejected because users have written code like 'except SomeError: ...' with the expectation that the 'except' clause would run exactly once. That's definitely true, and it's a downside of the multiple-except-executions approach, but I don't think it's convincing enough to rule this out on its own. The problem is, *all* our options for how 'except' should interact with ExceptionGroups will somehow break previous expectations.
Well, this is where we respectfully disagree. The case of changing how the regular try..except works in a backwards incompatible way is a very convincing blocker to us.
Concretely: imagine you have a pre-existing 'except SomeError', and some new code inside the 'try' block raises some number of 'SomeError's wrapped in an ExceptionGroup. There are three options:
- Execute the 'except' block multiple times. This breaks the expectation that it should be executed at most once. - Execute the 'except' block exactly once. But if there are multiple SomeError's, this require they be grouped and delivered as a single exception, which breaks typing. - Execute the 'except' block zero times. This is what the current PEP chooses, and breaks the expectation that 'except SomeError' should catch 'SomeError'.
So we have to pick our poison.
We did. The PEP talks at length as to why this isn't a problem.
I'm confused about the flattening suggestion - above you talk about "flat EG", but below about tracebacks. It's not clear to me whether you want EG to be flat (ie no nesting of EGs) or just the traceback to be flat (but you can still have a nested EG).
Hmm, I was thinking about making both of them flat, so no nested EGs. In all my designs, the only reason I ever had nesting was because I couldn't figure out any other way to make the tracebacks work. Do you have some other motivation for wanting nesting? If so that would be interesting, because it might point to why we're talking past each other and help us understand the problem better...
I also don't know what problem you are trying to solve with this.
I'm not saying that there's some fatal problem with the current PEP. (In my first message I explicitly said that it would be an improvement over the status quo :-).) But I think that nesting will be really counterintuitive/confusing for users in some ways. And concurrency APIs will be offputting if they force you to use a different special form of 'except' all the time. Basically the 'flat' version might be a lot more ergonomic, and that's important for a language like Python.
You keep saying that your idea of flat EGs is "ergonomic" and "important for a language like Python". The problem is that after all these emails I still have absolutely no idea about: - what exactly are you trying to propose? - what specific problem do you want to address? (that the current PEP, in your opinion, does not address) - why do you think that what you propose would be more ergonomic or simple? - what "flat EGs" or "flat tracebacks" even mean; I can't figure out not only the high level API, I don't even understand what you're talking about at the data structures level. Nathaniel, at this point it's clear that this thread somehow does not help us understand what you want. Could you please just write your own PEP clearly outlining your proposal, its upsides and downsides? Without a PEP from you this thread is just a distraction. Yury
Nathaniel Smith (in a conversation with Irit) wrote:
... suggested having each 'except Blah as exc' clause be executed once, receiving an ExceptionGroup containing all the Blah exceptions. Guido pointed out that this broke typing -- 'exc' would not longer have type 'Blah' -- and I was like... okay yeah that's a fatal flaw, never mind.
The fact that "Blah as exc" could return either a single Blah or a collection of Blahs would be annoying, but not fatal -- that seems well within the level of ambiguity typing can deal with.
never seriously raised the 'execute a single clause multiple times' option, because of the issue where in the nested design, taking individual exceptions out of an ExceptionGroup breaks tracebacks.
This also seems like something that isn't hard to work around. The obvious way would involve shared (parts of) tracebacks, but even that isn't required.
multiple-except-executions option should be rejected because users have written code like 'except SomeError: ...' with the expectation that the 'except' clause would run exactly once.
That is indeed a problem, because some error handlers would do bad things if rerun.
The problem is, *all* our options for how 'except' should interact with ExceptionGroups will somehow break previous expectations.
agreed.
Concretely: imagine you have a pre-existing 'except SomeError', and some new code inside the 'try' block raises some number of 'SomeError's wrapped in an ExceptionGroup. There are three options: - Execute the 'except' block multiple times. This breaks the expectation that it should be executed at most once.
Do we know how bad this really is? A second rollback or cancel or endConnection won't do what is expected, but they normally won't do much harm either. A second logError does the right thing.
- Execute the 'except' block exactly once. But if there are multiple SomeError's, this require they be grouped and delivered as a single exception, which breaks typing.
I'm not worried about typing, but I am worried about silently dropping all but the "first" SomeError. On the other hand, in practice were they all (including the "first") dropped before, in favor of some sort of superseding asyncCancelledError?
- Execute the 'except' block zero times. This is what the current PEP chooses, and breaks the expectation that 'except SomeError' should catch 'SomeError'.
To me, this seems the worst of the three options, but ... maybe it is effectively the status quo often enough that it becomes the least bad?
I'm confused about the flattening suggestion - above you talk about "flat EG", but below about tracebacks. It's not clear to me whether you want EG to be flat (ie no nesting of EGs) or just the traceback to be flat (but you can still have a nested EG). Hmm, I was thinking about making both of them flat, so no nested EGs. In all my designs, the only reason I ever had nesting was because I couldn't figure out any other way to make the tracebacks work. Do you have some other motivation for wanting nesting? If so that would be interesting, because it might point to why we're talking past each other and help us understand the problem better...
When I view this as strictly a typing problem, it matters which exceptions got joined at which junction; the shape of the tree is part of the type. When I view this as a production support programmer, I really don't care about that ... I only care what triggered each SomeException so that I can account for all the bad data instead of just one piece. Having to dig through multiple layers of grouping to get to each original SomeException separately is just an annoyance that *will* cause juniors to sometimes miss part of the bad data. ("Open the door and check for problems" is a lot easier to learn and remember than a recursive "Open the door and check for problems or other doors, and then repeat on each door you find.") -jJ
On Thu, Apr 29, 2021 at 9:14 AM Yury Selivanov <yselivanov.ml@gmail.com> wrote:
Nathaniel, at this point it's clear that this thread somehow does not help us understand what you want. Could you please just write your own PEP clearly outlining your proposal, its upsides and downsides? Without a PEP from you this thread is just a distraction.
If that's the best way to move forward, then ok. My main thing is just that I don't want to make this some antagonistic me-vs-you thing. After all, we all want the best design to be chosen, and none of us know what that is yet, so there's no need for conflict :-). Irit, Yury, would you be interested in co-authoring a PEP for the "flat EG" approach? Basically trying to set down the best possible version of each approach, so that we can put them next to each other? -n -- Nathaniel J. Smith -- https://vorpus.org
On Thu, May 6, 2021 at 2:17 AM Nathaniel Smith <njs@pobox.com> wrote:
On Thu, Apr 29, 2021 at 9:14 AM Yury Selivanov <yselivanov.ml@gmail.com> wrote:
Nathaniel, at this point it's clear that this thread somehow does not help us understand what you want. Could you please just write your own PEP clearly outlining your proposal, its upsides and downsides? Without a PEP from you this thread is just a distraction.
If that's the best way to move forward, then ok. My main thing is just that I don't want to make this some antagonistic me-vs-you thing. After all, we all want the best design to be chosen, and none of us know what that is yet, so there's no need for conflict :-).
Irit, Yury, would you be interested in co-authoring a PEP for the "flat EG" approach? Basically trying to set down the best possible version of each approach, so that we can put them next to each other?
Uh, probably this is obvious, but I mean "co-authoring with me". I'm not suggesting you two go off to do it without me :-) -- Nathaniel J. Smith -- https://vorpus.org
Hi Nathaniel, Philosophy of Science 101: nobody knows what TheBestSolution is and nobody ever will. That said, when we decided to submit PEP 654, that was a point in time when we came to believe that we found the right solution. We made significant changes later following feedback we agreed with, but in the case of "flat EG" I see several disadvantages and don't quite understand what the win is. I appreciate the invitation, but I don't want to work on that idea (nor am I the right person to push it forward). However, I think it's fantastic if you and others do pursue alternative options (independently of my biases). You will either find something better or validate PEP 654, and in either case the world wins. Feel free to adapt our implementation if that helps. You have a year before we have our next chance to commit to PEP 654. Irit On Thu, May 6, 2021 at 10:18 AM Nathaniel Smith <njs@pobox.com> wrote:
On Thu, Apr 29, 2021 at 9:14 AM Yury Selivanov <yselivanov.ml@gmail.com> wrote:
Nathaniel, at this point it's clear that this thread somehow does not help us understand what you want. Could you please just write your own PEP clearly outlining your proposal, its upsides and downsides? Without a PEP from you this thread is just a distraction.
If that's the best way to move forward, then ok. My main thing is just that I don't want to make this some antagonistic me-vs-you thing. After all, we all want the best design to be chosen, and none of us know what that is yet, so there's no need for conflict :-).
Irit, Yury, would you be interested in co-authoring a PEP for the "flat EG" approach? Basically trying to set down the best possible version of each approach, so that we can put them next to each other?
-n
-- Nathaniel J. Smith -- https://vorpus.org
participants (14)
-
Chris Angelico
-
Chris Jerdonek
-
Greg Ewing
-
Guido van Rossum
-
Irit Katriel
-
Irit Katriel
-
Jim J. Jewett
-
Nathaniel Smith
-
Nick Coghlan
-
Paul Sokolovsky
-
srkunze@mail.de
-
Terry Reedy
-
Thomas Wouters
-
Yury Selivanov