PEP 479 – Change StopIteration handling inside generators

PEP: 479
Title: Change StopIteration handling inside generators
Author: Chris Angelico <rosuav at gmail.com>, Guido van Rossum <guido at python.org>
Status: Final
Type: Standards Track
Created: 15-Nov-2014
Python-Version: 3.5
Post-History: 15-Nov-2014, 19-Nov-2014, 5-Dec-2014

Abstract

This PEP proposes a change to generators: when StopIteration is raised inside a generator, it is replaced with RuntimeError. (More precisely, this happens when the exception is about to bubble out of the generator’s stack frame.) Because the change is backwards incompatible, the feature is initially introduced using a __future__ statement.

Acceptance

This PEP was accepted by the BDFL on November 22. Because of the exceptionally short period from first draft to acceptance, the main objections brought up after acceptance were carefully considered and have been reflected in the “Alternate proposals” section below. However, none of the discussion changed the BDFL’s mind and the PEP’s acceptance is now final. (Suggestions for clarifying edits are still welcome – unlike IETF RFCs, the text of a PEP is not cast in stone after its acceptance, although the core design/plan/specification should not change after acceptance.)

Rationale

The interaction of generators and StopIteration is currently somewhat surprising, and can conceal obscure bugs. An unexpected exception should not result in subtly altered behaviour, but should cause a noisy and easily-debugged traceback. Currently, StopIteration raised accidentally inside a generator function will be interpreted as the end of the iteration by the loop construct driving the generator.

The main goal of the proposal is to ease debugging in the situation where an unguarded next() call (perhaps several stack frames deep) raises StopIteration and causes the iteration controlled by the generator to terminate silently. (Whereas, when some other exception is raised, a traceback is printed pinpointing the cause of the problem.)

This is particularly pernicious in combination with the yield from construct of PEP 380 1, as it breaks the abstraction that a subgenerator may be factored out of a generator. That PEP notes this limitation, but notes that “use cases for these [are] rare to non- existent”. Unfortunately while intentional use is rare, it is easy to stumble on these cases by accident:

import contextlib

@contextlib.contextmanager
def transaction():
    print('begin')
    try:
        yield from do_it()
    except:
        print('rollback')
        raise
    else:
        print('commit')

def do_it():
    print('Refactored initial setup')
    yield # Body of with-statement is executed here
    print('Refactored finalization of successful transaction')

def gene():
    for i in range(2):
        with transaction():
            yield i
            # return
            raise StopIteration  # This is wrong
        print('Should not be reached')

for i in gene():
    print('main: i =', i)

Here factoring out do_it into a subgenerator has introduced a subtle bug: if the wrapped block raises StopIteration, under the current behavior this exception will be swallowed by the context manager; and, worse, the finalization is silently skipped! Similarly problematic behavior occurs when an asyncio coroutine raises StopIteration, causing it to terminate silently, or when next is used to take the first result from an iterator that unexpectedly turns out to be empty, for example:

# using the same context manager as above
import pathlib

with transaction():
    print('commit file {}'.format(
        # I can never remember what the README extension is
        next(pathlib.Path('/some/dir').glob('README*'))))

In both cases, the refactoring abstraction of yield from breaks in the presence of bugs in client code.

Additionally, the proposal reduces the difference between list comprehensions and generator expressions, preventing surprises such as the one that started this discussion 2. Henceforth, the following statements will produce the same result if either produces a result at all:

a = list(F(x) for x in xs if P(x))
a = [F(x) for x in xs if P(x)]

With the current state of affairs, it is possible to write a function F(x) or a predicate P(x) that causes the first form to produce a (truncated) result, while the second form raises an exception (namely, StopIteration). With the proposed change, both forms will raise an exception at this point (albeit RuntimeError in the first case and StopIteration in the second).

Finally, the proposal also clears up the confusion about how to terminate a generator: the proper way is return, not raise StopIteration.

As an added bonus, the above changes bring generator functions much more in line with regular functions. If you wish to take a piece of code presented as a generator and turn it into something else, you can usually do this fairly simply, by replacing every yield with a call to print() or list.append(); however, if there are any bare next() calls in the code, you have to be aware of them. If the code was originally written without relying on StopIteration terminating the function, the transformation would be that much easier.

Background information

When a generator frame is (re)started as a result of a __next__() (or send() or throw()) call, one of three outcomes can occur:

A yield point is reached, and the yielded value is returned.
The frame is returned from; StopIteration is raised.
An exception is raised, which bubbles out.

In the latter two cases the frame is abandoned (and the generator object’s gi_frame attribute is set to None).

Proposal

If a StopIteration is about to bubble out of a generator frame, it is replaced with RuntimeError, which causes the next() call (which invoked the generator) to fail, passing that exception out. From then on it’s just like any old exception. 3

This affects the third outcome listed above, without altering any other effects. Furthermore, it only affects this outcome when the exception raised is StopIteration (or a subclass thereof).

Note that the proposed replacement happens at the point where the exception is about to bubble out of the frame, i.e. after any except or finally blocks that could affect it have been exited. The StopIteration raised by returning from the frame is not affected (the point being that StopIteration means that the generator terminated “normally”, i.e. it did not raise an exception).

A subtle issue is what will happen if the caller, having caught the RuntimeError, calls the generator object’s __next__() method again. The answer is that from this point on it will raise StopIteration – the behavior is the same as when any other exception was raised by the generator.

Another logical consequence of the proposal: if someone uses g.throw(StopIteration) to throw a StopIteration exception into a generator, if the generator doesn’t catch it (which it could do using a try/except around the yield), it will be transformed into RuntimeError.

During the transition phase, the new feature must be enabled per-module using:

from __future__ import generator_stop

Any generator function constructed under the influence of this directive will have the REPLACE_STOPITERATION flag set on its code object, and generators with the flag set will behave according to this proposal. Once the feature becomes standard, the flag may be dropped; code should not inspect generators for it.

A proof-of-concept patch has been created to facilitate testing. 4

Consequences for existing code

This change will affect existing code that depends on StopIteration bubbling up. The pure Python reference implementation of groupby 5 currently has comments “Exit on StopIteration” where it is expected that the exception will propagate and then be handled. This will be unusual, but not unknown, and such constructs will fail. Other examples abound, e.g. 6, 7.

(Nick Coghlan comments: “””If you wanted to factor out a helper function that terminated the generator you’d have to do “return yield from helper()” rather than just “helper()”.”””)

There are also examples of generator expressions floating around that rely on a StopIteration raised by the expression, the target or the predicate (rather than by the __next__() call implied in the for loop proper).

Writing backwards and forwards compatible code

With the exception of hacks that raise StopIteration to exit a generator expression, it is easy to write code that works equally well under older Python versions as under the new semantics.

This is done by enclosing those places in the generator body where a StopIteration is expected (e.g. bare next() calls or in some cases helper functions that are expected to raise StopIteration) in a try/except construct that returns when StopIteration is raised. The try/except construct should appear directly in the generator function; doing this in a helper function that is not itself a generator does not work. If raise StopIteration occurs directly in a generator, simply replace it with return.

Examples of breakage

Generators which explicitly raise StopIteration can generally be changed to simply return instead. This will be compatible with all existing Python versions, and will not be affected by __future__. Here are some illustrations from the standard library.

Lib/ipaddress.py:

if other == self:
    raise StopIteration

Becomes:

if other == self:
    return

In some cases, this can be combined with yield from to simplify the code, such as Lib/difflib.py:

if context is None:
    while True:
        yield next(line_pair_iterator)

Becomes:

if context is None:
    yield from line_pair_iterator
    return

(The return is necessary for a strictly-equivalent translation, though in this particular file, there is no further code, and the return can be omitted.) For compatibility with pre-3.3 versions of Python, this could be written with an explicit for loop:

if context is None:
    for line in line_pair_iterator:
        yield line
    return

More complicated iteration patterns will need explicit try/except constructs. For example, a hypothetical parser like this:

def parser(f):
    while True:
        data = next(f)
        while True:
            line = next(f)
            if line == "- end -": break
            data += line
        yield data

would need to be rewritten as:

def parser(f):
    while True:
        try:
            data = next(f)
            while True:
                line = next(f)
                if line == "- end -": break
                data += line
            yield data
        except StopIteration:
            return

or possibly:

def parser(f):
    for data in f:
        while True:
            line = next(f)
            if line == "- end -": break
            data += line
        yield data

The latter form obscures the iteration by purporting to iterate over the file with a for loop, but then also fetches more data from the same iterator during the loop body. It does, however, clearly differentiate between a “normal” termination (StopIteration instead of the initial line) and an “abnormal” termination (failing to find the end marker in the inner loop, which will now raise RuntimeError).

This effect of StopIteration has been used to cut a generator expression short, creating a form of takewhile:

def stop():
    raise StopIteration
print(list(x for x in range(10) if x < 5 or stop()))
# prints [0, 1, 2, 3, 4]

Under the current proposal, this form of non-local flow control is not supported, and would have to be rewritten in statement form:

def gen():
    for x in range(10):
        if x >= 5: return
        yield x
print(list(gen()))
# prints [0, 1, 2, 3, 4]

While this is a small loss of functionality, it is functionality that often comes at the cost of readability, and just as lambda has restrictions compared to def, so does a generator expression have restrictions compared to a generator function. In many cases, the transformation to full generator function will be trivially easy, and may improve structural clarity.

Explanation of generators, iterators, and StopIteration

The proposal does not change the relationship between generators and iterators: a generator object is still an iterator, and not all iterators are generators. Generators have additional methods that iterators don’t have, like send and throw. All this is unchanged. Nothing changes for generator users – only authors of generator functions may have to learn something new. (This includes authors of generator expressions that depend on early termination of the iteration by a StopIteration raised in a condition.)

An iterator is an object with a __next__ method. Like many other special methods, it may either return a value, or raise a specific exception - in this case, StopIteration - to signal that it has no value to return. In this, it is similar to __getattr__ (can raise AttributeError), __getitem__ (can raise KeyError), and so on. A helper function for an iterator can be written to follow the same protocol; for example:

def helper(x, y):
    if x > y: return 1 / (x - y)
    raise StopIteration

def __next__(self):
    if self.a: return helper(self.b, self.c)
    return helper(self.d, self.e)

Both forms of signalling are carried through: a returned value is returned, an exception bubbles up. The helper is written to match the protocol of the calling function.

A generator function is one which contains a yield expression. Each time it is (re)started, it may either yield a value, or return (including “falling off the end”). A helper function for a generator can also be written, but it must also follow generator protocol:

def helper(x, y):
    if x > y: yield 1 / (x - y)

def gen(self):
    if self.a: return (yield from helper(self.b, self.c))
    return (yield from helper(self.d, self.e))

In both cases, any unexpected exception will bubble up. Due to the nature of generators and iterators, an unexpected StopIteration inside a generator will be converted into RuntimeError, but beyond that, all exceptions will propagate normally.

Transition plan

Python 3.5: Enable new semantics under __future__ import; silent deprecation warning if StopIteration bubbles out of a generator not under __future__ import.
Python 3.6: Non-silent deprecation warning.
Python 3.7: Enable new semantics everywhere.

Alternate proposals

Raising something other than RuntimeError

Rather than the generic RuntimeError, it might make sense to raise a new exception type UnexpectedStopIteration. This has the downside of implicitly encouraging that it be caught; the correct action is to catch the original StopIteration, not the chained exception.

Supplying a specific exception to raise on return

Nick Coghlan suggested a means of providing a specific StopIteration instance to the generator; if any other instance of StopIteration is raised, it is an error, but if that particular one is raised, the generator has properly completed. This subproposal has been withdrawn in favour of better options, but is retained for reference.

Making return-triggered StopIterations obvious

For certain situations, a simpler and fully backward-compatible solution may be sufficient: when a generator returns, instead of raising StopIteration, it raises a specific subclass of StopIteration (GeneratorReturn) which can then be detected. If it is not that subclass, it is an escaping exception rather than a return statement.

The inspiration for this alternative proposal was Nick’s observation 8 that if an asyncio coroutine 9 accidentally raises StopIteration, it currently terminates silently, which may present a hard-to-debug mystery to the developer. The main proposal turns such accidents into clearly distinguishable RuntimeError exceptions, but if that is rejected, this alternate proposal would enable asyncio to distinguish between a return statement and an accidentally-raised StopIteration exception.

Of the three outcomes listed above, two change:

If a yield point is reached, the value, obviously, would still be returned.
If the frame is returned from, GeneratorReturn (rather than StopIteration) is raised.
If an instance of GeneratorReturn would be raised, instead an instance of StopIteration would be raised. Any other exception bubbles up normally.

In the third case, the StopIteration would have the value of the original GeneratorReturn, and would reference the original exception in its __cause__. If uncaught, this would clearly show the chaining of exceptions.

This alternative does not affect the discrepancy between generator expressions and list comprehensions, but allows generator-aware code (such as the contextlib and asyncio modules) to reliably differentiate between the second and third outcomes listed above.

However, once code exists that depends on this distinction between GeneratorReturn and StopIteration, a generator that invokes another generator and relies on the latter’s StopIteration to bubble out would still be potentially wrong, depending on the use made of the distinction between the two exception types.

Converting the exception inside next()

Mark Shannon suggested 10 that the problem could be solved in next() rather than at the boundary of generator functions. By having next() catch StopIteration and raise instead ValueError, all unexpected StopIteration bubbling would be prevented; however, the backward-incompatibility concerns are far more serious than for the current proposal, as every next() call now needs to be rewritten to guard against ValueError instead of StopIteration - not to mention that there is no way to write one block of code which reliably works on multiple versions of Python. (Using a dedicated exception type, perhaps subclassing ValueError, would help this; however, all code would still need to be rewritten.)

Note that calling next(it, default) catches StopIteration and substitutes the given default value; this feature is often useful to avoid a try/except block.

Sub-proposal: decorator to explicitly request current behaviour

Nick Coghlan suggested 11 that the situations where the current behaviour is desired could be supported by means of a decorator:

from itertools import allow_implicit_stop

@allow_implicit_stop
def my_generator():
    ...
    yield next(it)
    ...

Which would be semantically equivalent to:

def my_generator():
    try:
        ...
        yield next(it)
        ...
    except StopIteration
        return

but be faster, as it could be implemented by simply permitting the StopIteration to bubble up directly.

Single-source Python 2/3 code would also benefit in a 3.7+ world, since libraries like six and python-future could just define their own version of “allow_implicit_stop” that referred to the new builtin in 3.5+, and was implemented as an identity function in other versions.

However, due to the implementation complexities required, the ongoing compatibility issues created, the subtlety of the decorator’s effect, and the fact that it would encourage the “quick-fix” solution of just slapping the decorator onto all generators instead of properly fixing the code in question, this sub-proposal has been rejected. 12

Criticism

Unofficial and apocryphal statistics suggest that this is seldom, if ever, a problem. 13 Code does exist which relies on the current behaviour (e.g. 3, 6, 7), and there is the concern that this would be unnecessary code churn to achieve little or no gain.

Steven D’Aprano started an informal survey on comp.lang.python 14; at the time of writing only two responses have been received: one was in favor of changing list comprehensions to match generator expressions (!), the other was in favor of this PEP’s main proposal.

The existing model has been compared to the perfectly-acceptable issues inherent to every other case where an exception has special meaning. For instance, an unexpected KeyError inside a __getitem__ method will be interpreted as failure, rather than permitted to bubble up. However, there is a difference. Special methods use return to indicate normality, and raise to signal abnormality; generators yield to indicate data, and return to signal the abnormal state. This makes explicitly raising StopIteration entirely redundant, and potentially surprising. If other special methods had dedicated keywords to distinguish between their return paths, they too could turn unexpected exceptions into RuntimeError; the fact that they cannot should not preclude generators from doing so.

Why not fix all next() methods?

When implementing a regular __next__() method, the only way to indicate the end of the iteration is to raise StopIteration. So catching StopIteration here and converting it to RuntimeError would defeat the purpose. This is a reminder of the special status of generator functions: in a generator function, raising StopIteration is redundant since the iteration can be terminated by a simple return.