python-peps/pep-0654.rst

PEP: 654
Title: Exception Groups and except*
Author: Irit Katriel <iritkatriel@gmail.com>,
        Yury Selivanov <yury@edgedb.com>,
        Guido van Rossum <guido@python.org>
Status: Draft
Type: Standards Track
Content-Type: text/x-rst
Created: 22-Feb-2021
Post-History: 22-Feb-2021


Abstract
========

This document proposes language extensions that allow programs to raise and handle
multiple unrelated exceptions simultaneously:

* A new standard exception type, the ``ExceptionGroup``, which represents a
  group of unrelated exceptions being propagated together.

* A new syntax ``except*`` for handling ``ExceptionGroups``.

Motivation
==========

The interpreter is currently able to propagate at most one exception at a
time. The chaining features introduced in :pep:`3134` link together
exceptions that are related to each other as the cause or context, but
there are situations where multiple unrelated exceptions need to be propagated
together as the stack unwinds. Several real world use cases are listed below.

* **Concurrent errors**. Libraries for async concurrency provide APIs to invoke
  multiple tasks and return their results in aggregate. There isn't currently
  a good way for such libraries to handle situations where multiple tasks
  raise exceptions. The Python standard library's ``asyncio.gather()`` [1]_
  function provides two options: raise the first exception, or return the
  exceptions in the results list.  The Trio [2]_
  library has a ``MultiError`` exception type which it raises to report a
  collection of errors. Work on this PEP was initially motivated by the
  difficulties in handling ``MultiErrors`` [9]_, which are detailed in a design
  document for an improved version, ``MultiError2`` [3]_.
  That document demonstrates how difficult it is to create an effective API
  for reporting and handling multiple errors without the language changes we
  are proposing (see also the `Programming Without 'except \*'`_ section.)

  Implementing a better task spawning API in asyncio, inspired by Trio
  nurseries, was the main motivation for this PEP.  That work is currently
  blocked on Python not having native language level support for exception
  groups.

* **Multiple failures when retrying an operation.** The Python standard
  library's ``socket.create_connection`` function may attempt to connect to
  different addresses, and if all attempts fail it needs to report that to the
  user. It is an open issue how to aggregate these errors, particularly when
  they are different (see issue 29980 [4]_.)

* **Multiple user callbacks fail.** Python's ``atexit.register()`` function
  allows users to register functions that are called on system exit. If any of
  them raise exceptions, only the last one is reraised, but it would be better
  to reraised all of them together (see ``atexit`` documentation [5]_.)
  Similarly, the pytest library allows users to register finalizers which
  are executed at teardown. If more than one of these finalizers raises an
  exception, only the first is reported to the user. This can be improved with
  ``ExceptionGroups``, as explained in this issue by pytest developer Ran Benita
  (see pytest issue 8217 [6]_.)

* **Multiple errors in a complex calculation.** The Hypothesis library performs
  automatic bug reduction (simplifying code that demonstrates a bug). In the
  process it may find variations that generate different errors, and
  (optionally) reports all of them (see the Hypothesis documentation [7]_.)
  An ``ExceptionGroup`` mechanism as we are proposing here can resolve some of
  the difficulties with debugging that are mentioned in the link above, and
  which are due to the loss of context/cause information (communicated
  by Hypothesis Core Developer Zac Hatfield-Dodds).

* **Errors in wrapper code.** The Python standard library's
  ``tempfile.TemporaryDirectory`` context manager
  had an issue where an exception raised during cleanup in ``__exit__``
  effectively masked an exception that the user's code raised inside the context
  manager scope. While the user's exception was chained as the context of the
  cleanup error, it was not caught by the user's except clause
  (see issue 40857 [8]_.)

  The issue was resolved by making the cleanup code ignore errors, thus
  sidestepping the multiple exception problem. With the features we propose
  here, it would be possible for ``__exit__`` to raise an ``ExceptionGroup``
  containing its own errors along with the user's errors, and this would allow
  the user to catch their own exceptions by their types.


Rationale
=========

Grouping several exceptions together can be done without changes to the
language, simply by creating a container exception type.
Trio [2]_ is an example of a library that has made use of this technique in its
``MultiError`` [9]_ type. However, such an approach requires calling code to catch
the container exception type, and then to inspect it to determine the types of
errors that had occurred, extract the ones it wants to handle, and reraise the
rest.

Changes to the language are required in order to extend support for
``ExceptionGroups`` in the style of existing exception handling mechanisms. At
the very least we would like to be able to catch an ``ExceptionGroup`` only if
it contains an exception type that we choose to handle. Exceptions of
other types in the same ``ExceptionGroup`` need to be automatically reraised,
otherwise it is too easy for user code to inadvertently swallow exceptions
that it is not handling.

We considered whether it is possible to modify the semantics of ``except``
for this purpose, in a backwards-compatible manner, and found that it is not.
See the `Rejected Ideas`_ section for more on this.

The purpose of this PEP, then, is to add the ``except*`` syntax for handling
``ExceptionGroups`` in the interpreter, which in turn requires that
``ExceptionGroup`` is added as a builtin type. The semantics of handling
``ExceptionGroups`` are not backwards compatible with the current exception
handling semantics, so we are not proposing to modify the behavior of the
``except`` keyword but rather to add the new ``except*`` syntax.

Our premise is that ``ExceptionGroups`` and ``except*`` will be used
selectively, only when they are needed. We do not expect them to become
the default mechanism for exception handling.  The decision to raise
``ExceptionGroup`` from a library needs to be considered carefully and
regarded as an API-breaking change. We expect that this will normally be
done by introducing a new API rather than modifying an existing one.


Specification
=============

ExceptionGroup
--------------

The new builtin exception type, ``ExceptionGroup`` is a subclass of
``BaseException``, so it is assignable to ``Exception.__cause__`` and
``Exception.__context__``, and can be raised and handled as any exception
with ``raise ExceptionGroup(...)`` and ``try: ... except ExceptionGroup: ...``.

Its constructor takes two positional-only parameters: a message string and a
sequence of the nested exceptions, for example:
``ExceptionGroup('issues', [ValueError('bad value'), TypeError('bad type')])``.

The ``ExceptionGroup`` class exposes these parameters in the fields ``message``
and ``errors``.  A nested exception can also be an ``ExceptionGroup`` so the
class represents a tree of exceptions, where the leaves are plain exceptions and
each internal node represent a time at which the program grouped some
unrelated exceptions into a new ``ExceptionGroup`` and raised them together.
The ``ExceptionGroup`` class is final, i.e., it cannot be subclassed.

The ``ExceptionGroup.subgroup(condition)`` method gives us a way to obtain an
``ExceptionGroup`` that has the same metadata (cause, context, traceback) as
the original group, and the same nested structure of ``ExceptionGroups``, but
contains only those exceptions for which the condition is true:

.. code-block::

   >>> eg = ExceptionGroup(
   ...     "one",
   ...     [
   ...         TypeError(1),
   ...         ExceptionGroup(
   ...             "two",
   ...              [TypeError(2), ValueError(3)]
   ...         ),
   ...         ExceptionGroup(
   ...              "three",
   ...               [OSError(4)]
   ...         )
   ...     ]
   ... )
   >>> traceback.print_exception(eg)
   ExceptionGroup: one
      ------------------------------------------------------------
      TypeError: 1
      ------------------------------------------------------------
      ExceptionGroup: two
        ------------------------------------------------------------
        TypeError: 2
        ------------------------------------------------------------
        ValueError: 3
      ------------------------------------------------------------
      ExceptionGroup: three
        ------------------------------------------------------------
        OSError: 4
   >>> type_errors = eg.subgroup(lambda e: isinstance(e, TypeError))
   >>> traceback.print_exception(type_errors)
   ExceptionGroup: one
      ------------------------------------------------------------
      TypeError: 1
      ------------------------------------------------------------
      ExceptionGroup: two
        ------------------------------------------------------------
        TypeError: 2
   >>>


Empty nested ``ExceptionGroups`` are omitted from the result, as in the
case of ``ExceptionGroup("three")`` in the example above.  If none of the
leaf exceptions match the condition, ``subgroup`` returns ``None`` rather
than an empty ``ExceptionGroup``. The original ``eg``
is unchanged by ``subgroup``, but the value returned is not necessarily a full
new copy. Leaf exceptions are not copied, nor are ``ExceptionGroups`` which are
fully contained in the result. When it is necessary to partition an
``ExceptionGroup`` because the condition holds for some, but not all of its
contained exceptions, a new ``ExceptionGroup`` is created but the ``__cause__``,
``__context__`` and ``__traceback__`` fields are copied by reference, so are
shared with the original ``eg``.

If both the subgroup and its complement are needed, the
``ExceptionGroup.split(condition)`` method can be used:

.. code-block::

   >>> type_errors, other_errors = eg.split(lambda e: isinstance(e, TypeError))
   >>> traceback.print_exception(type_errors)
   ExceptionGroup: one
      ------------------------------------------------------------
      TypeError: 1
      ------------------------------------------------------------
      ExceptionGroup: two
        ------------------------------------------------------------
        TypeError: 2
   >>> traceback.print_exception(other_errors)
   ExceptionGroup: one
      ------------------------------------------------------------
      ExceptionGroup: two
        ------------------------------------------------------------
        ValueError: 3
      ------------------------------------------------------------
      ExceptionGroup: three
        ------------------------------------------------------------
        OSError: 4
   >>>


If a split is trivial (one side is empty), then None is returned for the
other side:

.. code-block::

   >>> other_errors.split(lambda e: isinstance(e, SyntaxError))
   (None, ExceptionGroup('one', [
     ExceptionGroup('two', [
       ValueError(3)
     ]),
     ExceptionGroup('three', [
       OSError(4)])]))

Since splitting by exception type is a very common use case, ``subgroup`` and
``split`` can take an exception type or tuple of exception types and treat it
as a shorthand for matching that type: ``eg.split(T)`` divides ``eg`` into the
subgroup of leaf exceptions that match the type ``T``, and the subgroup of those
that do not (using the same check as ``except`` for a match).

The Traceback of an ``ExceptionGroup``
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

For regular exceptions, the traceback represents a simple path of frames,
from the frame in which the exception was raised to the frame in which it
was caught or, if it hasn't been caught yet, the frame that the program's
execution is currently in. The list is constructed by the interpreter, which
appends any frame from which it exits to the traceback of the 'current
exception' if one exists. To support efficient appends, the links in a
traceback's list of frames are from the oldest to the newest frame. Appending
a new frame is then simply a matter of inserting a new head to the linked
list referenced from the exception's ``__traceback__`` field. Crucially, the
traceback's frame list is immutable in the sense that frames only need to be
added at the head, and never need to be removed.

We do not need to make any changes to this data structure. The ``__traceback__``
field of the ``ExceptionGroup`` instance represents the path that the contained
exceptions travelled through together after being joined into the
``ExceptionGroup``, and the same field on each of the nested exceptions
represents the path through which this exception arrived at the frame of the
merge.

What we do need to change is any code that interprets and displays tracebacks,
because it now needs to continue into tracebacks of nested exceptions, as
in the following example:

.. code-block::

   >>> def f(v):
   ...     try:
   ...         raise ValueError(v)
   ...     except ValueError as e:
   ...         return e
   ...
   >>> try:
   ...     raise ExceptionGroup("one", [f(1)])
   ... except ExceptionGroup as e:
   ...     eg1 = e
   ...
   >>> try:
   ...     raise ExceptionGroup("two", [f(2), eg1])
   ... except ExceptionGroup as e:
   ...     eg2 = e
   ...
   >>> import traceback
   >>> traceback.print_exception(eg2)
   Traceback (most recent call last):
     File "<stdin>", line 2, in <module>
   ExceptionGroup: two
      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 3, in f
      ValueError: 2
      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 2, in <module>
      ExceptionGroup: one
        ------------------------------------------------------------
        Traceback (most recent call last):
         File "<stdin>", line 3, in f
        ValueError: 1
   >>>

Handling ``ExceptionGroups``
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

We expect that when programs catch and handle ``ExceptionGroups``, they will
typically either query to check if it has leaf exceptions for which some
condition holds (using ``subgroup`` or ``split``) or format the exception
(using the ``traceback`` module's methods).

It is unlikely to be useful to inspect the individual leaf exceptions. To see
why, suppose that an application caught an ``ExceptionGroup`` raised in an
``asyncio.gather()`` call. At this stage, the context for each specific
exception is lost. Any recovery for this exception should have been performed
before it was grouped with other exceptions into the ``ExceptionGroup`` [10]_.
Furthermore, the application is likely to react in the same way to any number
of instances of a certain exception type, so it is more likely that we will
want to know whether ``eg.subgroup(T)`` is None or not, than we are to be
intersted in the number of ``Ts`` in ``eg``.

If it does turn out to be necessary for an applicaiton to iterate over the
individual exceptions of an ``ExceptionGroup`` ``eg``, this can be done by
calling ``traverse(eg)``, where ``traverse`` is defined as follows:

.. code-block::

    def traverse(exc, tbs=None):
        if tbs is None:
            tbs = []

        tbs.append(exc.__traceback__)
        if isinstance(exc, ExceptionGroup):
            for e in exc.errors:
                traverse(e, tbs)
        else:
            # exc is a leaf exception and its traceback
            # is the concatenation of the traceback in tbs
            process_leaf(exc, tbs)
        tbs.pop()


except*
-------

We are proposing to introduce a new variant of the ``try..except`` syntax to
simplify working with exception groups. The ``*`` symbol indicates that multiple
exceptions can be handled by each ``except*`` clause:

.. code-block::

   try:
       ...
   except *SpamError:
       ...
   except *FooError as e:
       ...
   except *(BarError, BazError) as e:
       ...

In a traditional ``try-except`` statement there is only one exception to handle,
so the body of at most one ``except`` clause executes; the first one that matches
the exception. With the new syntax, an ``except*`` clause can match a subgroup
of the ``ExceptionGroup`` that was raised, while the remaining part is matched
by following ``except*`` clauses. In other words, a single ``ExceptionGroup`` can
cause several ``except*`` clauses to execute, but each such clause executes at
most once (for all matching exceptions from the group) and each exception is
either handled by exactly one clause (the first one that matches its type)
or is reraised at the end.

For example, suppose that the body of the ``try`` block above raises
``eg = ExceptionGroup('msg', [FooError(1), FooError(2), BazError()])``.
The ``except*`` clauses are evaluated in order by calling ``split`` on the
``unhandled`` ``ExceptionGroup``, which is initially equal to ``eg`` and then shrinks
as exceptions are matched and extracted from it.  In the first ``except*`` clause,
``unhandled.split(SpamError)`` returns ``(None, unhandled)`` so the body of this
block is not executed and ``unhandled`` is unchanged. For the second block,
``unhandled.split(FooError)`` returns a non-trivial split ``(match, rest)`` with
``match = ExceptionGroup('msg', [FooError(1), FooError(2)])``
and ``rest = ExceptionGroup('msg', [BazError()])``. The body of this ``except*``
block is executed, with the value of ``e`` and ``sys.exc_info()`` set to ``match``.
Then, ``unhandled`` is set to ``rest``.
Finally, the third block matches the remaining exception so it is executed
with ``e`` and ``sys.exc_info()`` set to ``ExceptionGroup('msg', [BazError()])``.


Exceptions are matched using a subclass check. For example:

.. code-block::

   try:
       low_level_os_operation()
   except *OSerror as eg:
       for e in eg.errors:
           print(type(e).__name__)

could output:

.. code-block::

   BlockingIOError
   ConnectionRefusedError
   OSError
   InterruptedError
   BlockingIOError

The order of ``except*`` clauses is significant just like with the regular
``try..except``:

.. code-block::

   >>> try:
   ...     raise ExceptionGroup("problem", [BlockingIOError()])
   ... except *OSError as e:   # Would catch the error
   ...     print(repr(e))
   ... except *BlockingIOError: # Would never run
   ...     print('never')
   ...
   ExceptionGroup('problem', [BlockingIOError()])

Recursive Matching
~~~~~~~~~~~~~~~~~~

The matching of ``except*`` clauses against an ``ExceptionGroup`` is performed
recursively, using the ``ExceptionGroup.split()`` method:

.. code-block::

   >>> try:
   ...     raise ExceptionGroup(
   ...         "eg",
   ...         [
   ...             ValueError('a'),
   ...             TypeError('b'),
   ...             ExceptionGroup(
   ...                 "nested",
   ...                 [TypeError('c'), KeyError('d')])
   ...         ]
   ...     )
   ... except *TypeError as e1:
   ...     print(f'e1 = {e1!r}')
   ... except *Exception as e2:
   ...     print(f'e2 = {e2!r}')
   ...
   e1 = ExceptionGroup('eg', [TypeError('b'), ExceptionGroup('nested', [TypeError('c')])])
   e2 = ExceptionGroup('eg', [ValueError('a'), ExceptionGroup('nested', [KeyError('d')])])
   >>>

Unmatched Exceptions
~~~~~~~~~~~~~~~~~~~~

If not all exceptions in an ``ExceptionGroup`` were matched by the ``except*``
clauses, the remaining part of the ``ExceptionGroup`` is propagated on:

.. code-block::

   >>> try:
   ...     try:
   ...         raise ExceptionGroup(
   ...             "msg", [
   ...                  ValueError('a'), TypeError('b'),
   ...                  TypeError('c'), KeyError('e')
   ...             ]
   ...         )
   ...     except *ValueError as e:
   ...         print(f'got some ValueErrors: {e!r}')
   ...     except *TypeError as e:
   ...         print(f'got some TypeErrors: {e!r}')
   ... except ExceptionGroup as e:
   ...     print(f'propagated: {e!r}')
   ...
   got some ValueErrors: ExceptionGroup('msg', [ValueError('a')])
   got some TypeErrors: ExceptionGroup('msg', [TypeError('b'), TypeError('c')])
   propagated: ExceptionGroup('msg', [KeyError('e')])
   >>>


Naked Exceptions
~~~~~~~~~~~~~~~~

If the exception raised inside the ``try`` body is not of type ``ExceptionGroup``,
we call it a ``naked`` exception. If its type matches one of the ``except*``
clauses, it is caught and wrapped by an ``ExceptionGroup`` with an empty message
string. This is to make the type of ``e`` consistent and statically known:

.. code-block::

   >>> try:
   ...     raise BlockingIOError
   ... except *OSError as e:
   ...     print(repr(e))
   ...
   ExceptionGroup('', [BlockingIOError()])

However, if a naked exception is not caught, it propagates in its original
naked form:

.. code-block::

   >>> try:
   ...     try:
   ...         raise ValueError(12)
   ...     except *TypeError as e:
   ...         print('never')
   ... except ValueError as e:
   ...     print(f'caught ValueError: {e!r}')
   ...
   caught ValueError: ValueError(12)
   >>>

Raising exceptions in an ``except*`` block
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

In a traditional ``except`` block, there are two ways to raise exceptions:
``raise e`` to explicitly raise an exception object ``e``, or naked ``raise`` to
reraise the 'current exception'. When ``e`` is the current exception, the two
forms are not equivalent because a reraise does not add the current frame to
the stack:

.. code-block::

   def foo():                           | def foo():
       try:                             |     try:
           1 / 0                        |         1 / 0
       except ZeroDivisionError as e:   |     except ZeroDivisionError:
           raise e                      |         raise
                                        |
   foo()                                | foo()
                                        |
   Traceback (most recent call last):   | Traceback (most recent call last):
     File "/Users/guido/a.py", line 7   |   File "/Users/guido/b.py", line 7
      foo()                             |     foo()
     File "/Users/guido/a.py", line 5   |   File "/Users/guido/b.py", line 3
      raise e                           |     1/0
     File "/Users/guido/a.py", line 3   | ZeroDivisionError: division by zero
      1/0                               |
   ZeroDivisionError: division by zero  |


This holds for ``ExceptionGroups`` as well, but the situation is now more complex
because there can be exceptions raised and reraised from multiple ``except*``
clauses, as well as unhandled exceptions that need to propagate.
The interpreter needs to combine all those exceptions into a result, and
raise that.

The reraised exceptions and the unhandled exceptions are subgroups of the
original ``ExceptionGroup``, and share its metadata (cause, context, traceback).
On the other hand, each of the explicitly raised exceptions has its own
metadata - the traceback contains the line from which it was raised, its
cause is whatever it may have been explicitly chained to, and its context is the
value of ``sys.exc_info()`` in the ``except*`` clause of the raise.

In the aggregated ``ExceptionGroup``, the reraised and unhandled exceptions have
the same relative structure as in the original exception, as if they were split
off together in one ``subgroup`` call. For example, in the snippet below the
inner ``try-except*`` block raises an ``ExceptionGroup`` that contains all
``ValueErrors`` and ``TypeErrors`` merged back into the same shape they had in
the original ``ExceptionGroup``:

.. code-block::

   >>> try:
   ...     try:
   ...         raise ExceptionGroup(
   ...             "eg",
   ...             [
   ...                 ValueError(1),
   ...                 TypeError(2),
   ...                 OSError(3),
   ...                 ExceptionGroup(
   ...                     "nested",
   ...                     [OSError(4), TypeError(5), ValueError(6)])
   ...             ]
   ...         )
   ...     except *ValueError as e:
   ...         print(f'*ValueError: {e!r}')
   ...         raise
   ...     except *OSError as e:
   ...         print(f'*OSError: {e!r}')
   ... except ExceptionGroup as e:
   ...     print(repr(e))
   ...
   *ValueError: ExceptionGroup('eg', [ValueError(1), ExceptionGroup('nested', [ValueError(6)])])
   *OSError: ExceptionGroup('eg', [OSError(3), ExceptionGroup('nested', [OSError(4)])])
   ExceptionGroup('eg', [ValueError(1), TypeError(2), ExceptionGroup('nested', [TypeError(5), ValueError(6)])])
   >>>


When exceptions are raised explicitly, they are independent of the original
exception group, and cannot be merged with it (they have their own cause,
context and traceback). Instead, they are combined into a new ``ExceptionGroup``,
which also contains the reraised/unhandled subgroup described above.

In the following example, the ``ValueErrors`` were raised so they are in their
own ``ExceptionGroup``, while the ``OSErrors`` were reraised so they were
merged with the unhandled ``TypeErrors``.

.. code-block::

   >>> try:
   ...     try:
   ...         raise ExceptionGroup(
   ...             "eg",
   ...             [
   ...                 ValueError(1),
   ...                 TypeError(2),
   ...                 OSError(3),
   ...                 ExceptionGroup(
   ...                     "nested",
   ...                     [OSError(4), TypeError(5), ValueError(6)])
   ...             ]
   ...         )
   ...     except *ValueError as e:
   ...         print(f'*ValueError: {e!r}')
   ...         raise e
   ...     except *OSError as e:
   ...         print(f'*OSError: {e!r}')
   ...         raise
   ... except ExceptionGroup as e:
   ...     traceback.print_exception(e)
   ...
   *ValueError: ExceptionGroup('eg', [ValueError(1), ExceptionGroup('nested', [ValueError(6)])])
   *OSError: ExceptionGroup('eg', [OSError(3), ExceptionGroup('nested', [OSError(4)])])
   Traceback (most recent call last):
     File "<stdin>", line 3, in <module>
   ExceptionGroup
      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 12, in <module>
       File "<stdin>", line 3, in <module>
      ExceptionGroup: eg
        ------------------------------------------------------------
        ValueError: 1
        ------------------------------------------------------------
        ExceptionGroup: nested
          ------------------------------------------------------------
          ValueError: 6
      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 3, in <module>
      ExceptionGroup: eg
        ------------------------------------------------------------
        TypeError: 2
        ------------------------------------------------------------
        OSError: 3
        ------------------------------------------------------------
        ExceptionGroup: nested
          ------------------------------------------------------------
          OSError: 4
          ------------------------------------------------------------
          TypeError: 5
   >>>


Chaining
~~~~~~~~

Explicitly raised ``ExceptionGroups`` are chained as with any exceptions. The
following example shows how part of ``ExceptionGroup`` "one" became the
context for ``ExceptionGroup`` "two", while the other part was combined with
it into the new ``ExceptionGroup``.

.. code-block::

   >>> try:
   ...     try:
   ...         raise ExceptionGroup("one", [ValueError('a'), TypeError('b')])
   ...     except *ValueError:
   ...         raise ExceptionGroup("two", [KeyError('x'), KeyError('y')])
   ... except BaseException as e:
   ...     traceback.print_exception(e)
   ...
   Traceback (most recent call last):
     File "<stdin>", line 3, in <module>
   ExceptionGroup
      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 3, in <module>
      ExceptionGroup: one
        ------------------------------------------------------------
        ValueError: a

      During handling of the above exception, another exception occurred:

      Traceback (most recent call last):
       File "<stdin>", line 5, in <module>
      ExceptionGroup: two
        ------------------------------------------------------------
        KeyError: 'x'
        ------------------------------------------------------------
        KeyError: 'y'

      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 3, in <module>
      ExceptionGroup: one
        ------------------------------------------------------------
        TypeError: b


Raising New Exceptions
~~~~~~~~~~~~~~~~~~~~~~

In the previous examples the explicit raises were of the exceptions that
were caught, so for completion we show a new exception being raised, with
chaining:

.. code-block::

   >>> try:
   ...     try:
   ...         raise TypeError('bad type')
   ...     except *TypeError as e:
   ...         raise ValueError('bad value') from e
   ... except ExceptionGroup as e:
   ...     traceback.print_exception(e)
   ...
   Traceback (most recent call last):
     File "<stdin>", line 3, in <module>
   ExceptionGroup
      ------------------------------------------------------------
      ExceptionGroup
        ------------------------------------------------------------
        Traceback (most recent call last):
         File "<stdin>", line 3, in <module>
        TypeError: bad type

      The above exception was the direct cause of the following exception:

      Traceback (most recent call last):
       File "<stdin>", line 5, in <module>
      ValueError: bad value
   >>>


Note that exceptions raised in one ``except*`` clause are not eligible to match
other clauses from the same ``try`` statement:

.. code-block::

   >>> try:
   ...     try:
   ...         raise TypeError(1)
   ...     except *TypeError:
   ...         raise ValueError(2)  # <- not caught in the next clause
   ...     except *ValueError:
   ...         print('never')
   ... except ExceptionGroup as e:
   ...     traceback.print_exception(e)
   ...
   Traceback (most recent call last):
     File "<stdin>", line 3, in <module>
   ExceptionGroup
      ------------------------------------------------------------
      ExceptionGroup
        ------------------------------------------------------------
        Traceback (most recent call last):
         File "<stdin>", line 3, in <module>
        TypeError: 1

      During handling of the above exception, another exception occurred:

      Traceback (most recent call last):
       File "<stdin>", line 5, in <module>
      ValueError: 2


Raising a new instance of a naked exception does not cause this exception to
be wrapped by an ``ExceptionGroup``. Rather, the exception is raised as is, and
if it needs to be combined with other propagated exceptions, it becomes a
direct child of the new ``ExceptionGroup`` created for that:


.. code-block::

   >>> try:
   ...     try:
   ...         raise ExceptionGroup("eg", [ValueError('a')])
   ...     except *ValueError:
   ...         raise KeyError('x')
   ... except BaseException as e:
   ...     traceback.print_exception(e)
   ...
   Traceback (most recent call last):
     File "<stdin>", line 3, in <module>
   ExceptionGroup
      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 3, in <module>
      ExceptionGroup: eg
        ------------------------------------------------------------
        ValueError: a

      During handling of the above exception, another exception occurred:

      Traceback (most recent call last):
       File "<stdin>", line 5, in <module>
      KeyError: 'x'
   >>>
   >>> try:
   ...     try:
   ...         raise ExceptionGroup("eg", [ValueError('a'), TypeError('b')])
   ...     except *ValueError:
   ...         raise KeyError('x')
   ... except BaseException as e:
   ...     traceback.print_exception(e)
   ...
   Traceback (most recent call last):
     File "<stdin>", line 3, in <module>
   ExceptionGroup
      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 3, in <module>
      ExceptionGroup: eg
        ------------------------------------------------------------
        ValueError: a

      During handling of the above exception, another exception occurred:

      Traceback (most recent call last):
       File "<stdin>", line 5, in <module>
      KeyError: 'x'

      ------------------------------------------------------------
      Traceback (most recent call last):
       File "<stdin>", line 3, in <module>
      ExceptionGroup: eg
        ------------------------------------------------------------
        TypeError: b
   >>>


Finally, as an example of how the proposed API can help us work effectively
with ``ExceptionGroups``, the following code ignores all ``EPIPE`` OS errors,
while letting all other exceptions propagate.

.. code-block::

   try:
       low_level_os_operation()
   except *OSerror as errors:
       raise errors.subgroup(lambda e: e.errno != errno.EPIPE) from None


Caught Exception Objects
~~~~~~~~~~~~~~~~~~~~~~~~

It is important to point out that the ``ExceptionGroup`` bound to ``e`` is an
ephemeral object. Raising it via ``raise`` or ``raise e`` will not cause changes
to the overall shape of the ``ExceptionGroup``.  Any modifications to it will
likely be lost:

.. code-block::

   >>> eg = ExceptionGroup("eg", [TypeError(12)])
   >>> eg.foo = 'foo'
   >>> try:
   ...     raise eg
   ... except *TypeError as e:
   ...     e.foo = 'bar'
   ... #   ^----------- ``e`` is an ephemeral object that might get
   >>> #                      destroyed after the ``except*`` clause.
   >>> eg.foo
   'foo'


Forbidden Combinations
~~~~~~~~~~~~~~~~~~~~~~

It is not possible to use both traditional ``except`` blocks and the new
``except*`` clauses in the same ``try`` statement. The following is a
``SyntaxError``:

.. code-block::

   try:
       ...
   except ValueError:
       pass
   except *CancelledError:  # <- SyntaxError:
       pass                 #    combining ``except`` and ``except*``
                            #    is prohibited

It is possible to catch the ``ExceptionGroup`` type with ``except``, but not
with ``except*`` because the latter is ambiguous:

.. code-block::

   try:
       ...
   except ExceptionGroup:  # <- This works
       pass

   try:
       ...
   except *ExceptionGroup:  # <- Runtime error
       pass

   try:
       ...
   except *(TypeError, ExceptionGroup):  # <- Runtime error
       pass


An empty "match anything" ``except*`` block is not supported as its meaning may
be confusing:

.. code-block::

   try:
       ...
   except*:   # <- SyntaxError
       pass


``continue``, ``break``, and ``return`` are disallowed in ``except*`` clauses,
causing a ``SyntaxError``. This is because the exceptions in an
``ExceptionGroup`` are assumed to be independent, and the presence or absence
of one of them should not impact handling of the others, as could happen if we
allow an ``except*`` clause to change the way control flows through other
clauses.


Backwards Compatibility
=======================

Backwards compatibility was a requirement of our design, and the changes we
propose in this PEP will not break any existing code:

* The addition of a new builtin exception type ``ExceptionGroup`` does not impact
  existing programs. The way that existing exceptions are handled and displayed
  does not change in any way.

* The behaviour of ``except`` is unchanged so existing code will continue to work.
  Programs will only be impacted by the changes proposed in this PEP once they
  begin to use ``ExceptionGroups`` and ``except*``.


Once programs begin to use these features, there will be migration issues to
consider:

* An ``except Exception:`` clause will not catch ``ExceptionGroup`` because it
  is derived from ``BaseException``. Any such clause will need to be replaced
  by ``except (Exception, ExceptionGroup):`` or ``except *Exception:``.

* Similarly, any ``except T:`` clause that wraps code which is now potentially
  raising ``ExceptionGroup`` needs to become ``except *T:``, and its body may
  need to be updated.

* Libraries that need to support older Python versions will not be able to use
  ``except*`` or raise ``ExceptionGroups``.


How to Teach This
=================

``ExceptionGroups`` and ``except*`` will be documented as part of the language
standard. Libraries that raise ``ExceptionGroups`` such as ``asyncio`` will need
to specify this in their documentation and clarify which API calls need to be
wrapped with ``try-except*`` rather than ``try-except``.


Reference Implementation
========================

We developed these concepts (and the examples for this PEP) with
the help of the reference implementation [11]_.

It has the builtin ``ExceptionGroup`` along with the changes to the traceback
formatting code, in addition to the grammar, compiler and interpreter changes
required to support ``except*``.

Two opcodes were added: one implements the exception type match check via
``ExceptionGroup.split()``, and the other is used at the end of a ``try-except``
construct to merge all unhandled, raised and reraised exceptions (if any).
The raised/reraised exceptions are collected in a list on the runtime stack.
For this purpose, the body of each ``except*`` clause is wrapped in a traditional
``try-except`` which captures any exceptions raised. Both raised and reraised
exceptions are collected in the same list. When the time comes to merge them
into a result, the raised and reraised exceptions are distinguished by comparing
their metadata fields (context, cause, traceback) with those of the originally
raised exception. As mentioned above, the reraised exceptions have the same
metadata as the original, while the raised ones do not.

Rejected Ideas
==============

Make ExceptionGroup Iterable
----------------------------

We considered making ``ExceptionGroups`` iterable, so that ``list(eg)`` would
produce a flattened list of the leaf exceptions contained in the group.
We decided that this would not be a sound API, because the metadata
(cause, context and traceback) of the individual exceptions in a group is
incomplete and this could create problems.

Furthermore, as we explained in the `Handling ExceptionGroups`_ section, we
find it unlikely that iteration over leaf exceptions will have many use cases.
We did, however, provide there the code for a traversal algorithm that
correctly constructs each leaf exception's metadata. If it does turn out to
be useful in practice, we can add that utility to the standard library.

Traceback Representation
------------------------

We considered options for adapting the traceback data structure to represent
trees, but it became apparent that a traceback tree is not meaningful once
separated from the exceptions it refers to. While a simple-path traceback can
be attached to any exception by a ``with_traceback()`` call, it is hard to
imagine a case where it makes sense to assign a traceback tree to an exception
group.  Furthermore, a useful display of the traceback includes information
about the nested exceptions. For these reasons we decided that it is best to
leave the traceback mechanism as it is and modify the traceback display code.

Extend ``except`` to handle ``ExceptionGroups``
-----------------------------------------------

We considered extending the semantics of ``except`` to handle
``ExceptionGroups``, instead of introducing ``except*``. There were two
backwards compatibility concerns with this. The first is the type of the
caught exception. Consider this example:

.. code-block::

    try:
        . . .
    except OSError as err:
        if err.errno != ENOENT:
            raise

If the value assigned to err is an ``ExceptionGroup`` containing all of
the ``OSErrors`` that were raised, then the attribute access ``err.errno``
no longer works. So we would need to execute the body of the ``except``
clause multiple times, once for each exception in the group. However, this
too is a potentially breaking change because at the moment we write ``except``
clauses with the knowledge that they are only executed once. If there is
a non-idempotent operation there, such as releasing a resource, the
repetition could be harmful.

A new ``except`` alternative
----------------------------

We considered introducing a new keyword (such as ``catch``) which can be used
to handle both naked exceptions and ``ExceptionGroups``. Its semantics would
be the same as those of ``except*`` when catching an ``ExceptionGroup``, but
it would not wrap a naked exception to create an ``ExceptionGroup``. This
would have been part of a long term plan to replace ``except`` by ``catch``,
but we decided that deprecating ``except`` in favour of an enhanced keyword
would be too confusing for users at this time, so it is more appropriate
to introduce the ``except*`` syntax for ``ExceptionGroups`` while ``except``
continues to be used for simple exceptions.

Applying an ``except*`` clause on one exception at a time
---------------------------------------------------------

We explained above why it is unsafe to execute an ``except`` clause in existing
code more than once. We considered doing this in the new ``except*`` clauses, 
where the backwards compatibility considerations do not exist.
The idea is to always execute an ``except*`` clause on a single exception,
possibly executing the same clause multiple times when it matches multiple
exceptions. We decided instead to execute each ``except*`` clause at most once,
giving it an ``ExceptionGroup`` that contains all matching exceptions. The
reason for this decision was the observation that when a program needs to know
the particular context of an exception it is handling, the exception is
handled before it is grouped and raised together with other exceptions.

For example, ``KeyError`` is an exception that typically relates to a certain
operation. Any recovery code would be local to the place where the error
occurred, and would use the traditional ``except``:

.. code-block::

   try:
       dct[key]
   except KeyError:
       # handle the exception

It is unlikely that asyncio users would want to do something like this:

.. code-block::

   try:
       async with asyncio.TaskGroup() as g:
           g.create_task(task1); g.create_task(task2)
   except *KeyError:
       # handling KeyError here is meaningless, there's
       # no context to do anything with it but to log it.

When a program handles a collection of exceptions that were aggregated into
an exception group, it would not typically attempt to recover from any
particular failed operation, but will rather use the types of the errors to
determine how they should impact the program's control flow or what logging
or cleanup is required. This decision is likely to be the same whether the group
contains a single or multiple instances of something like a ``KeyboardInterrupt``
or ``asyncio.CancelledError``.  Therefore, it is more convenient to handle all
exceptions matching an ``except*`` at once.  If it does turn out to be necessary,
the handler can inpect the ``ExceptionGroup`` and process the individual
exceptions in it.

Not matching naked exceptions in ``except*``
--------------------------------------------

We considered the option of making ``except *T`` match only ``ExceptionGroups``
that contain ``Ts``, but not naked ``Ts``. To see why we thought this would
not be a desirable feature, return to the distinction in the previous paragraph
between operation errors and control flow exceptions. If we don't know whether
we should expect naked exceptions or ``ExceptionGroups`` from the body of a
``try`` block,  then we're not in the position of handling operation errors.
Rather, we are likely calling a fairly generic function and will be handling
errors to make control flow decisions. We are likely to do the same thing
whether we catch a naked exception of type ``T`` or an ``ExceptionGroup``
with one or more ``Ts``. Therefore, the burden of having to explicitly handle
both is not likely to have semantic benefit.

If it does turn out to be necessary to make the distinction, it is always
possible to nest in the ``try-except*`` clause an additional ``try-except`` clause
which intercepts and handles a naked exception before the ``except*`` clause
has a change to wrap it in an  ``ExceptionGroup``. In this case the overhead
of specifying both is not additional burden - we really do need to write a
separate code block to handle each case:

.. code-block::

   try:
       try:
           ...
       except SomeError:
           # handle the naked exception
   except *SomeError:
       # handle the ExceptionGroup


Allow mixing ``except:`` and ``except*:`` in the same ``try``
-------------------------------------------------------------

This option was rejected because it adds complexity without adding useful
semantics. Presumably the intention would be that an ``except T:`` block handles
only naked exceptions of type ``T``, while ``except *T:`` handles ``T`` in
``ExceptionGroups``. We already discussed above why this is unlikely
to be useful in practice, and if it is needed then the nested ``try-except``
block can be used instead to achieve the same result.

``try*`` instead of ``except*``
-------------------------------

Since either all or none of the clauses of a ``try`` construct are ``except*``,
we considered changing the syntax of the ``try`` instead of all the ``except*``
clauses. We rejected this because it would be less obvious. The fact that we
are handling ``ExceptionGroups`` of ``T`` rather than only naked ``Ts`` should be
specified in the same place where we state ``T``.


Programming Without 'except \*'
===============================

Consider the following simple example of the ``except *`` syntax (pretending
Trio natively supported this proposal):

.. code-block::

   try:
       async with trio.open_nursery() as nursery:
           # Make two concurrent calls to child()
           nursery.start_soon(child)
           nursery.start_soon(child)
   except *ValueError:
       pass

Here is how this code would look in Python 3.9:

.. code-block::

   def handle_ValueError(exc):
       if isinstance(exc, ValueError):
           return None
       else:
           return exc   # reraise exc

   with MultiError.catch(handle_ValueError):
       async with trio.open_nursery() as nursery:
           # Make two concurrent calls to child()
           nursery.start_soon(child)
           nursery.start_soon(child)

This example clearly demonstrates how unintuitive and cumbersome handling
of multiple errors is in current Python.  The exception handling logic has
to be in a separate closure and is fairly low level, requiring the writer to
have non-trivial understanding of both Python exceptions mechanics and the
Trio APIs.  Instead of using the ``try..except`` block we have to use a
``with`` block.  We need to explicitly reraise exceptions we are not handling.
Handling more exception types or implementing more complex
exception handling logic will only further complicate the code to the point
of it being unreadable.


See Also
========

* An analysis of how exception groups will likely be used in asyncio
  programs: [10]_.

* The issue where the ``except*`` concept was first formalized: [12]_.

* ``MultiError2`` design document: [3]_.

* Reporting Multiple Errors in the Hypothesis library: [7]_.


References
==========

.. [1] https://docs.python.org/3/library/asyncio-task.html#asyncio.gather

.. [2] https://trio.readthedocs.io/en/stable/

.. [3] https://github.com/python-trio/trio/issues/611

.. [4] https://bugs.python.org/issue29980

.. [5] https://docs.python.org/3/library/atexit.html#atexit.register

.. [6] https://github.com/pytest-dev/pytest/issues/8217

.. [7] https://hypothesis.readthedocs.io/en/latest/settings.html#hypothesis.settings.report_multiple_bugs

.. [8] https://bugs.python.org/issue40857

.. [9] https://trio.readthedocs.io/en/stable/reference-core.html#trio.MultiError

.. [10] https://github.com/python/exceptiongroups/issues/3#issuecomment-716203284

.. [11] https://github.com/iritkatriel/cpython/tree/exceptionGroup-stage5

.. [12] https://github.com/python/exceptiongroups/issues/4

.. [13] https://trio.readthedocs.io/en/stable/reference-core.html#nurseries-and-spawning

Copyright
=========

This document is placed in the public domain or under the
CC0-1.0-Universal license, whichever is more permissive.


..
   Local Variables:
   mode: indented-text
   indent-tabs-mode: nil
   sentence-end-double-space: t
   fill-column: 70
   coding: utf-8
   End: