set and dict iteration

[Aaron Brady]

Hello,

I observed an inconsistency in the behavior of 'set' and 'dict' iterators. It is "by design" according to the docs.

'''
http://docs.python.org/dev/library/stdtypes.html#dict-views

iter(dictview). Iterating views while adding or deleting entries in the dictionary may raise a RuntimeError or fail to iterate over all entries.
'''

The 'set' has the same behavior. Iteration might also complete successfully.

The inconsistency is, if we remove an element from a set and add another during iteration, the new element might appear later in the iteration, and might not, depending on the hash code; therefore comparing the size of the set between iterations isn't adequate. Example:

http://home.comcast.net/~castironpi-misc/clpy-0062%20set%20iterators.py

'''
# py: { 'ver': '3' }

set0= set( ( 1, 2 ) )

iter0= iter( set0 )
print( next( iter0 ) )

set0.add( 3 )
set0.remove( 2 )
print( next( iter0 ) )


print( )

set0= set( ( 6, 7 ) )

iter0= iter( set0 )
print( next( iter0 ) )

set0.add( 8 )
set0.remove( 7 )
print( next( iter0 ) )
'''

Output:

'''
1
3

6
Traceback (most recent call last):
File [...] line 22, in <module>
print( next( iter0 ) )
StopIteration
'''

Iteration should behave the same regardless of the contents of the set. Continuing iteration over sets and dicts after a modification isn't defined; it should unconditionally raise an error.

What's going on, is '8' is added before the position of the iterator due to hashing in the second part, but the size doesn't change, so the iterator reaches the end of the set after '7' is removed.

The inconsistency isn't easily solved. One possibility is to use a timestamp or other serial index in the object and iterators, and compare them on every iteration to determine if a modification has occurred.

Another possibility which the author prefers, is to maintain a secondary collection of the iterators of an object, and invalidate them upon modification. The applicable collection structure is a doubly-linked linked list, informally depicted:

http://home.comcast.net/~castironpi-misc/clpy-0062%20set%20iterators.png

Upon modification, the set traverses its iterators, setting an 'invalid' flag on each; and subsequent calls to any of them raise an 'IterationError'. Adding and removing iterators to and from the secondary list is performed in O( 1 ) time with no penalty.

The above example depicted a 'Set'. 'Dicts' have the same anomaly, but the solution is ambiguous, since dict values can be changed meaningfully without altering the structure of the object. In the author's opinion, the dict should not raise an 'IterationError' on value changes, only key changes like the set, but the argument isn't conclusive.

[Dave Angel]

On 08/16/2012 02:00 PM, Aaron Brady wrote:
> Hello,
>
> I observed an inconsistency in the behavior of 'set' and 'dict' iterators. It is "by design" according to the docs.
>
> '''
> http://docs.python.org/dev/library/stdtypes.html#dict-views
>
> iter(dictview). Iterating views while adding or deleting entries in the dictionary may raise a RuntimeError or fail to iterate over all entries.
> '''
>
> The 'set' has the same behavior. Iteration might also complete successfully.
>
> The inconsistency is, if we remove an element from a set and add another during iteration, the new element might appear later in the iteration, and might not, depending on the hash code; therefore comparing the size of the set between iterations isn't adequate. Example:

> <SNIP>

>
>
> Iteration should behave the same regardless of the contents of the set. Continuing iteration over sets and dicts after a modification isn't defined; it should unconditionally raise an error.

Why is it the iterator's job to protect against the user's bug? The
doc is clear enough. If you don't change the collection, you won't have
a problem.

> <SNIP more details>.

Everything else is implementation defined. Why should an implementation
be forced to have ANY extra data structure to detect a static bug in the
caller's code?



--

DaveA

[Paul Rubin]

Dave Angel <d...@davea.name> writes:
> Everything else is implementation defined. Why should an implementation
> be forced to have ANY extra data structure to detect a static bug in the
> caller's code?

For the same reason the interpreter checks for type errors at runtime
and raises TypeError, instead of letting the program go into the weeds.

[Ian Kelly]

On Thu, Aug 16, 2012 at 12:00 PM, Aaron Brady <casti...@gmail.com> wrote:
> The inconsistency is, if we remove an element from a set and add another during iteration, the new element might appear later in the iteration, and might not, depending on the hash code; therefore comparing the size of the set between iterations isn't adequate. Example:

It can be more than just the new element. For example, here the
entire set is repeated (Python 3.2):

>>> s = set(range(8, 13))
>>> it = iter(s)
>>> from itertools import islice
>>> list(islice(it, 5)) # avoid exhausting the iterator
[8, 9, 10, 11, 12]
>>> s.add(13)
>>> s.remove(13)
>>> list(it)
[8, 9, 10, 11, 12]

This occurs because the addition of the sixth item triggers a resize
of the underlying hash table, and the existing items, which were
originally in slots 0-4, are now in slots 8-12.

[Ian Kelly]

On Thu, Aug 16, 2012 at 4:55 PM, Ian Kelly <ian.g...@gmail.com> wrote:
> On Thu, Aug 16, 2012 at 12:00 PM, Aaron Brady <casti...@gmail.com> wrote:

>> The inconsistency is, if we remove an element from a set and add another during iteration, the new element might appear later in the iteration, and might not, depending on the hash code; therefore comparing the size of the set between iterations isn't adequate. Example:
>

> It can be more than just the new element. For example, here the
> entire set is repeated (Python 3.2):
>
>>>> s = set(range(8, 13))
>>>> it = iter(s)
>>>> from itertools import islice
>>>> list(islice(it, 5)) # avoid exhausting the iterator
> [8, 9, 10, 11, 12]
>>>> s.add(13)
>>>> s.remove(13)
>>>> list(it)
> [8, 9, 10, 11, 12]
>
> This occurs because the addition of the sixth item triggers a resize
> of the underlying hash table, and the existing items, which were
> originally in slots 0-4, are now in slots 8-12.

Another curious example:

>>> s = set(range(8, 48, 8))
>>> s
{8, 16, 40, 24, 32}

>>> it = iter(s)
>>> from itertools import islice

>>> list(islice(it, 4))
[8, 16, 40, 24]
>>> s.add(48)
>>> s.remove(48)
>>> list(it)
[8, 16, 40, 24]

Hey, what happened to 32?

[Dave Angel]

There's an enormous difference between type errors, which affect the low
level dispatch, and checking for whether a dict has changed and may have
invalidated the iterator. If we were really going to keep track of what
iterators are tracking a given dict or set, why stop there? Why not
check if another process has changed a file we're iterating through? Or ...

[Ian Kelly]

On Thu, Aug 16, 2012 at 5:11 PM, Dave Angel <da...@dejaviewphoto.com> wrote:
> There's an enormous difference between type errors, which affect the low
> level dispatch, and checking for whether a dict has changed and may have
> invalidated the iterator. If we were really going to keep track of what
> iterators are tracking a given dict or set, why stop there? Why not
> check if another process has changed a file we're iterating through? Or ...

How does this affect low-level dispatch (Python 2.7)?

>>> class Foo(object):
... def bar(self):
... return self
...
>>> Foo().bar()
<__main__.Foo object at 0x00CBEAB0>
>>> Foo.bar(Foo())
<__main__.Foo object at 0x00CC9390>
>>> Foo.bar(object())

Traceback (most recent call last):

File "<stdin>", line 1, in <module>
TypeError: unbound method bar() must be called with Foo instance as
first argument (got object instance instead)

There is no low-level need for this TypeError -- it's purely a case of
not letting the developer shoot himself in the foot. Although to be
honest the interpreter doesn't give quite enough rope (to mix
metaphors) in this case, and I'm glad for the sake of duck typing that
they removed this particular error in Python 3.

With regard to key insertion and deletion while iterating over a dict
or set, though, there is just no good reason to be doing that
(especially as the result is very implementation-specific), and I
wouldn't mind a more complete low-level check against it as long as
it's not too expensive (which is not clearly the case with the current
suggestion at all).

[Paul Rubin]

Ian Kelly <ian.g...@gmail.com> writes:
> With regard to key insertion and deletion while iterating over a dict
> or set, though, there is just no good reason to be doing that
> (especially as the result is very implementation-specific), and I
> wouldn't mind a more complete low-level check against it as long as
> it's not too expensive (which is not clearly the case with the current
> suggestion at all).

One possible approach is to freeze the dictionary against modification
while any iterator is open on it. You could keep a count of active
iterators in the dict structure, adjusting it whenever an iterator is
created or closed/destroyed.

[Steven D'Aprano]

Which is why Python doesn't do it -- because it is (claimed to be)
excessively expensive for the benefit that you would get.

Not because it is a matter of principle that data integrity is
unimportant. Data integrity *is* important, but in the opinion of the
people who wrote these particular data structures, the effort required to
guarantee correct iteration in the face of mutation is too expensive for
the benefit.

Are they right? I don't know. I know that the list sort method goes to a
lot of trouble to prevent code from modifying lists while they are being
sorted. During the sort, the list temporarily appears to be empty to
anything which attempts to access it. So at least sometimes, the Python
developers spend effort to ensure data integrity.

Luckily, Python is open source. If anyone thinks that sets and dicts
should include more code protecting against mutation-during-iteration,
they are more than welcome to come up with a patch. Don't forget unit and
regression tests, and also a set of timing results which show that the
slow-down isn't excessive.


--
Steven

[Paul Rubin]

Steven D'Aprano <steve+comp....@pearwood.info> writes:
> Luckily, Python is open source. If anyone thinks that sets and dicts
> should include more code protecting against mutation-during-iteration,
> they are more than welcome to come up with a patch. Don't forget unit and
> regression tests, and also a set of timing results which show that the
> slow-down isn't excessive.

It could be a debugging option, in which case even a fairly significant
slowdown is acceptable.

[Ulrich Eckhardt]

What if there is an iterator left over from a loop that was terminated
early? That could block access to the sequence even though nothing is
/really/ iterating over it.

I personally prefer a reliable error, at least when __debug__ is set.
Someone suggested a timestamp or a list of active iterators, which both
sound reasonable. The two should be O(1) and O(#iterators) in complexity
on all mutating operations and O(1) on iteration, so they should be
acceptable. With a C implementation it probably boils down to very few
cycles (checking a pointer/incrementing an integer). I can't say if this
is feasible without compromising performance though, at the very least
it requires an additional member in all dicts and iterators.

Uli

[Aaron Brady]

Good examples. The former occurs without the 'islice' as well.

s= set( range( 8, 13 ) )
it= iter( s )
print( [ next( it ) for _ in range( 5 ) ] )
s.add( 13 )
s.remove( 13 )
print( [ next( it ) for _ in range( 5 ) ] )

Output:

[Aaron Brady]

Good examples. The former occurs without the 'islice' as well.

s= set( range( 8, 13 ) )
it= iter( s )
print( [ next( it ) for _ in range( 5 ) ] )
s.add( 13 )
s.remove( 13 )
print( [ next( it ) for _ in range( 5 ) ] )

Output:

[Aaron Brady]

Good point. Your approach is another consistent solution.

The difference is in where the exception is raised. Invalidating the iterators raises an exception when they're called. Locking the set/dict raises an exception on 'add' and 'remove' calls.

The latter also forces additional statements to delete iterators before they leave scope in some cases. We wouldn't be able to make modifications in a 'for' suite, even if followed by a 'break', which could be a problem for existing code.

[Aaron Brady]

Another possibility is to use the 'gc.get_referrers' mechanism to obtain the iterators.


import gc
a= set( ( 0, 1, 2 ) )
b= iter( a )
c= iter( a )
d= iter( a )
print( gc.get_referrers( a ) )

Output:

[<set_iterator object at 0x00C0B9E0>, <set_iterator object at 0x00C0BA08>, <set_iterator object at 0x00C0BA30>, [others] ]

This approach wouldn't be as time-efficient as a dedicated secondary structure, due to the other objects which refer to the set, including variable namespaces.

[Aaron Brady]

I contribute a patch some time ago. It wasn't accepted. However this thread seems to show a moderately more favorable sentiment than that one.

Is there a problem with hacking on the Beta? Or should I wait for the Release? Does anyone want to help me with the changes? Perhaps P. Rubin could contribute the variation he suggested as well.

[Chris Angelico]

On Sat, Aug 18, 2012 at 4:37 AM, Aaron Brady <casti...@gmail.com> wrote:
> Is there a problem with hacking on the Beta?

Nope. Hack on the beta, then when the release arrives, rebase your
work onto it. I doubt that anything of this nature will be changed
between now and then.

ChrisA

[Aaron Brady]

Thanks Chris, your post was encouraging.

I have a question about involving the 'tp_clear' field of the types.

http://docs.python.org/dev/c-api/typeobj.html#PyTypeObject.tp_clear

'''
...The tuple type does not implement a tp_clear function, because it’s possible to prove that no reference cycle can be composed entirely of tuples.
'''

I didn't follow the reasoning in the proof; the premise is necessary but IMHO not obviously sufficient. Nevertheless, the earlier diagram contains an overt homogeneous reference cycle.

Reposting: http://home.comcast.net/~castironpi-misc/clpy-0062%20set%20iterators.png

In my estimate, the 'tp_traverse' and 'tp_clear' fields of the set doesn't need to visit the auxiliary collection; the same fields of the iterators don't need to visit the primary set or other iterators; and references in the linked list don't need to be included in the iterators' reference counts.

Can someone who is more familiar with the cycle detector and cycle breaker, help prove or disprove the above?

[Aaron Brady]

On Friday, August 17, 2012 4:57:41 PM UTC-5, Chris Angelico wrote:

[MRAB]

In simple terms, when you create an immutable object it can contain
only references to pre-existing objects, but in order to create a cycle
you need to make an object refer to another which is created later, so
it's not possible to create a cycle out of immutable objects.

However, using Python's C API it _is_ possible to create such a cycle,
by mutating an otherwise-immutable tuple (see PyTuple_SetItem and
PyTuple_SET_ITEM).

[Aaron Brady]

Are there any precedents for storing uncounted references to PyObject's?

One apparent problematic case is creating an iterator to a set, then adding it to the set. However the operation is a modification, and causes the iterator to be removed from the secondary list before the set is examined for collection.

Otherwise, the iterator keeps a counted reference to the set, but the set does not keep a counted reference to the iterator, so the iterator will always be freed first. Therefore, the set's secondary list will be empty when the set is freed.

Concurrent addition and deletion of iterators should be disabled, and the iterators should remove themselves from the set's secondary list before they decrement their references to the set.

Please refresh the earlier diagram; counted references are distinguished separately. Reposting: http://home.comcast.net/~castironpi-misc/clpy-0062%20set%20iterators.png

[Aaron Brady]

On Saturday, August 18, 2012 5:14:05 PM UTC-5, MRAB wrote:

[Aaron Brady]

The patch for the above is only 40-60 lines. However it introduces two new concepts.

The first is a "linked list", a classic dynamic data structure, first developed in 1955, cf. http://en.wikipedia.org/wiki/Linked_list . Linked lists are absent in Python, including the standard library and CPython implementation, beyond the weak reference mechanism and garbage collector. The "collections.deque" structure shares some of the linked list interface but uses arrays.

The second is "uncounted references". The uncounted references are references to "set iterators" exclusively, exist only internally to "set" objects, and are invisible to the rest of the program. The reason for the exception is that iterators are unique in the Python Data Model; iterators consist of a single immutable reference, unlike both immutable types such as strings and numbers, as well as container types. Counted references could be used instead, but would be consistently wasted work for the garbage collector, though the benefit to programmers' peace of mind could be significant.

Please share your opinion! Do you agree that the internal list resolves the inconsistency? Do you agree with the strategy? Do you agree that uncounted references are justified to introduce, or are counted references preferable?

[Aaron Brady]

On Saturday, August 18, 2012 9:28:32 PM UTC-5, Aaron Brady wrote:

[Steven D'Aprano]

On Thu, 23 Aug 2012 09:49:41 -0700, Aaron Brady wrote:

[...]

> The patch for the above is only 40-60 lines. However it introduces two
> new concepts.
>
> The first is a "linked list", a classic dynamic data structure, first
> developed in 1955, cf. http://en.wikipedia.org/wiki/Linked_list .
> Linked lists are absent in Python

They certainly are not. There's merely no named "linked list" class.

Linked lists are used by collections.ChainMap, tracebacks, xml.dom,
Abstract Syntax Trees, and probably many other places. (Well, technically
some of these are trees rather than lists.) You can trivially create a
linked list:

x = [a, [b, [c, [d, [e, None]]]]]

is equivalent to a singly-linked list with five nodes. Only less
efficient.

> The second is "uncounted references". The uncounted references are
> references to "set iterators" exclusively, exist only internally to
> "set" objects, and are invisible to the rest of the program. The reason
> for the exception is that iterators are unique in the Python Data Model;
> iterators consist of a single immutable reference, unlike both immutable
> types such as strings and numbers, as well as container types. Counted
> references could be used instead, but would be consistently wasted work
> for the garbage collector, though the benefit to programmers' peace of
> mind could be significant.

The usual way to implement "uncounted references" is by using weakrefs.
Why invent yet another form of weakref?



--
Steven

[Aaron Brady]

Hello S. D'Aprano. Thanks for your support as always.

The semantics of the second collection are equivalent to a WeakSet. The space and time consumption of a WeakSet are higher in comparison to a linked list. However, so long as we iterated over it using the C API instead of creating an iterator, it would be consistent. If we dynamically create the WeakSet on demand and free it when empty, the space consumption would be lower. Typical use cases don't involve creating thousands of iterators, or rapidly creating and destroying them, so the performance impact might not be severe.

Regarding the bare weakrefs, if the iterator's destructor hasn't been called, then the pointer is still valid. If it has been called, then it's not present in the list. Unlike Python classes, the destructors of C extension classes are guaranteed to be called. Therefore there are no points during exection at which a node needs to check whether a reference to its neighbor is valid.

Are your concerns based on data integrity, future maintainability, or what?

[Ian Kelly]

On Thu, Aug 23, 2012 at 10:49 AM, Aaron Brady <casti...@gmail.com> wrote:
> The patch for the above is only 40-60 lines. However it introduces two new concepts.

Is there a link to the patch?

> The first is a "linked list", a classic dynamic data structure, first developed in 1955, cf. http://en.wikipedia.org/wiki/Linked_list . Linked lists are absent in Python, including the standard library and CPython implementation, beyond the weak reference mechanism and garbage collector. The "collections.deque" structure shares some of the linked list interface but uses arrays.
>
> The second is "uncounted references". The uncounted references are references to "set iterators" exclusively, exist only internally to "set" objects, and are invisible to the rest of the program. The reason for the exception is that iterators are unique in the Python Data Model; iterators consist of a single immutable reference, unlike both immutable types such as strings and numbers, as well as container types. Counted references could be used instead, but would be consistently wasted work for the garbage collector, though the benefit to programmers' peace of mind could be significant.
>
> Please share your opinion! Do you agree that the internal list resolves the inconsistency? Do you agree with the strategy? Do you agree that uncounted references are justified to introduce, or are counted references preferable?

This feature is a hard sell as it is; I think that adding uncounted
references into the mix is only going to make that worse. May I
suggest an alternate approach? Internally tag each set or dict with a
"version", which is just a C int. Every time the hash table is
modified, increment the version. When an iterator is created, store
the current version on the iterator. When the iterator is advanced,
check that the iterator version matches the dict/set version. If
they're not equal, raise an error.

This should add less overhead than the linked list without any
concerns about reference counting. It does introduce a small bug in
that an error condition could be "missed", if the version is
incremented a multiple of 2**32 or 2**64 times between iterations --
but how often is that really likely to occur? Bearing in mind that
this error is meant for debugging and not production error handling,
you could even make the version a single byte and I'd still be fine
with that.

Cheers,
Ian

[Aaron Brady]

On Monday, August 27, 2012 2:17:45 PM UTC-5, Ian wrote:
> On Thu, Aug 23, 2012 at 10:49 AM, Aaron Brady <casti...@gmail.com> wrote:
>
> > The patch for the above is only 40-60 lines. However it introduces two new concepts.
>
>
>
> Is there a link to the patch?

Please see below. It grew somewhat during development.

> > The first is a "linked list".
SNIP.

> > The second is "uncounted references". The uncounted references are references to "set iterators" exclusively, exist only internally to "set" objects, and are invisible to the rest of the program. The reason for the exception is that iterators are unique in the Python Data Model; iterators consist of a single immutable reference, unlike both immutable types such as strings and numbers, as well as container types. Counted references could be used instead, but would be consistently wasted work for the garbage collector, though the benefit to programmers' peace of mind could be significant.
>
> >
>
> > Please share your opinion! Do you agree that the internal list resolves the inconsistency? Do you agree with the strategy? Do you agree that uncounted references are justified to introduce, or are counted references preferable?
>
>
>
> This feature is a hard sell as it is; I think that adding uncounted
>
> references into the mix is only going to make that worse. May I
>
> suggest an alternate approach? Internally tag each set or dict with a
>
> "version", which is just a C int. Every time the hash table is
>
> modified, increment the version. When an iterator is created, store
>
> the current version on the iterator. When the iterator is advanced,
>
> check that the iterator version matches the dict/set version. If
>
> they're not equal, raise an error.
>
>
>
> This should add less overhead than the linked list without any
>
> concerns about reference counting. It does introduce a small bug in
>
> that an error condition could be "missed", if the version is
>
> incremented a multiple of 2**32 or 2**64 times between iterations --
>
> but how often is that really likely to occur? Bearing in mind that
>
> this error is meant for debugging and not production error handling,
>
> you could even make the version a single byte and I'd still be fine
>
> with that.
>
>
>
> Cheers,
>
> Ian

Hi Ian,

We could use a Python long object for the version index to prevent overflow. Combined with P. Rubin's idea to count the number of open iterators, most use cases still wouldn't exceed a single word comparison; we could reset the counter when there weren't any. Using the linked list collection, modification operations are expensive in rare cases. Using the version index, iteration is expensive in rare cases.

I was more interested in the linked list for conceptual reasons, so I developed it further. Changelog, diff file, test suite, and links are below. The devs should be aware that a competing patch might be developed. I would be pleased to hear what everybody thinks of it!

Linked list with uncounted references implementation for Win32.

Added:
- 'set_clear_ex' and 'set_clear_internal_ex' methods, differ in invalidation and conditional invalidation behavior and return type.. The 'set.clear()' method and 'tp_clear' type field both called the same method.
- 'set_invalidate_iter_linked' method. Iterate over the iterators of a set, mark them invalid, and clear the list.
- 'setiter_unlink_internal' method. Remove the iterator from the set's linked list of iterators.
- 'IterationError', global.
- New fields:
-- PySetObject: setiterobject *iter_linked. Pointer to the first element of the linked list of the iterators of the set.
-- setiterobject: setiterobject *linked_pred, *linked_succ. Predecessor and successor nodes in the linked list of iterators of the same set.
-- setiterobject: char ob_valid. Validation status of the iterator.
- Result is compared with original in 'set_intersection_update' and '_multi' to determine whether to invalidate the list of iterators. Asymptotic running time is unchanged.
- Pending: add 'tp_clear' field to 'PySetIter_Type'?
- Test script included, 'large numbers' test pending.

6 files changed: { setobject.h, setobject.c, exceptions.c, pyerrors.h, python3.def, python33stub.def }. Test script 'set_iterator_test.py' new. Linked list interface and pseudocode 'patch_pseudocode.txt'.
Zip file:
http://home.comcast.net/~castironpi-misc/clpy-0062-set_iterator_patch.zip
Diff file of 3.3.0b2:
http://home.comcast.net/~castironpi-misc/clpy-0062-set_iterator_diff.txt

[Aaron Brady]

On Monday, August 27, 2012 2:17:45 PM UTC-5, Ian wrote:

> On Thu, Aug 23, 2012 at 10:49 AM, Aaron Brady <casti...@gmail.com> wrote:
>
> > The patch for the above is only 40-60 lines. However it introduces two new concepts.
>
>
>
> Is there a link to the patch?

Please see below. It grew somewhat during development.

> > The first is a "linked list".
SNIP.

> > The second is "uncounted references". The uncounted references are references to "set iterators" exclusively, exist only internally to "set" objects, and are invisible to the rest of the program. The reason for the exception is that iterators are unique in the Python Data Model; iterators consist of a single immutable reference, unlike both immutable types such as strings and numbers, as well as container types. Counted references could be used instead, but would be consistently wasted work for the garbage collector, though the benefit to programmers' peace of mind could be significant.
>
> >
>
> > Please share your opinion! Do you agree that the internal list resolves the inconsistency? Do you agree with the strategy? Do you agree that uncounted references are justified to introduce, or are counted references preferable?
>
>
>
> This feature is a hard sell as it is; I think that adding uncounted
>
> references into the mix is only going to make that worse. May I
>
> suggest an alternate approach? Internally tag each set or dict with a
>
> "version", which is just a C int. Every time the hash table is
>
> modified, increment the version. When an iterator is created, store
>
> the current version on the iterator. When the iterator is advanced,
>
> check that the iterator version matches the dict/set version. If
>
> they're not equal, raise an error.
>
>
>
> This should add less overhead than the linked list without any
>
> concerns about reference counting. It does introduce a small bug in
>
> that an error condition could be "missed", if the version is
>
> incremented a multiple of 2**32 or 2**64 times between iterations --
>
> but how often is that really likely to occur? Bearing in mind that
>
> this error is meant for debugging and not production error handling,
>
> you could even make the version a single byte and I'd still be fine
>
> with that.
>
>
>
> Cheers,
>
> Ian

[Ian Kelly]

On Sun, Sep 2, 2012 at 11:43 AM, Aaron Brady <casti...@gmail.com> wrote:
> We could use a Python long object for the version index to prevent overflow. Combined with P. Rubin's idea to count the number of open iterators, most use cases still wouldn't exceed a single word comparison; we could reset the counter when there weren't any.

We could use a Python long; I just don't think the extra overhead is
justified in a data structure that is already highly optimized for
speed. Incrementing and testing a C int is *much* faster than doing
the same with a Python long.

[Aaron Brady]

I think the technique would require two python longs and a bool in the set, and a python long in the iterator.

One long counts the number of existing (open) iterators. Another counts the version. The bool keeps track of whether an iterator has been created since the last modification, in which case the next modification requires incrementing the version and resetting the flag.

[Aaron Brady]

On Monday, September 3, 2012 2:30:24 PM UTC-5, Ian wrote:

[Dave Angel]

I think you're over-engineering the problem. it's a bug if an iterator
is used after some change is made to the set it's iterating over. We
don't need to catch every possible instance of the bug, that's what
testing is for. The point is to "probably" detect it, and for that, all
we need is a counter in the set and a counter in the open iterator.
Whenever changing the set, increment its count. And whenever iterating,
check the two counters. if they don't agree, throw an exception, and
destroy the iterator. i suppose that could be done with a flag, but it
could just as easily be done by zeroing the pointer to the set.

I'd figure a byte or two would be good enough for the counts, but a C
uint would be somewhat faster, and wouldn't wrap as quickly.

--

DaveA

[Serhiy Storchaka]

On 27.08.12 22:17, Ian Kelly wrote:
> May I
> suggest an alternate approach? Internally tag each set or dict with a
> "version", which is just a C int. Every time the hash table is
> modified, increment the version. When an iterator is created, store
> the current version on the iterator. When the iterator is advanced,
> check that the iterator version matches the dict/set version. If
> they're not equal, raise an error.

Oh, I'm surprised that this strategy is not used yet. I was sure that it
is used.

[Aaron Brady]

Hi D. Angel,

The serial index constantly reminds me of upper limits. I have the same problem with PHP arrays, though it's not a problem with the language itself.

The linked list doesn't have a counter, it invalidates iterators when a modification is made, therefore it's the "correct" structure in my interpretation. But it does seem more precarious comparatively, IMHO.

Both strategies solve the problem I posed originally, they both involve trade-offs, and it's too late to include either in 3.3.0.

[Aaron Brady]

On Monday, September 3, 2012 3:28:28 PM UTC-5, Dave Angel wrote:

[Steven D'Aprano]

I think that is over-engineered and could be the difference between
having the patch accepted and having it rejected. After all, many people
will argue that the existing solution to the problem is good enough.

Dicts are extremely common, and your patch increases both the memory
usage of every dict, and the overhead of every write operation
(__setitem__, __delitem__, update). Only a very few dicts will actually
need this overhead, for the rest it is waste. It is important to keep
that waste to a minimum or risk having the patch rejected.

An unsigned C int can count up to 4,294,967,295. I propose that you say
that is enough iterators for anyone, and use a single, simple, version
counter in the dict and the iterator. If somebody exceeds that many
iterators to a single dict or set, and the version field overflows by
exactly 2**32 versions, the results are no worse than they are now. You
won't be introducing any new bugs.

Complicating the solution is, in my opinion, unnecessary. Why should
every set and dict carry the cost of incrementing TWO Python longs and a
flag when just a single C int is sufficient for all realistic use-cases?

The opportunity for failure is extremely narrow:

- I must have an iterator over a dict or set;
- and I must have paused the iteration in some way;
- and then I must create exactly 2**32 other iterators to the
same dict or set;
- and at some point modify the dict or set
- and then restart the first iterator

at which point some items returned by the iterator *may* be duplicated or
skipped (depends on the nature of the modifications).


--
Steven

[Dave Angel]

On 09/03/2012 09:26 PM, Steven D'Aprano wrote:
> On Mon, 03 Sep 2012 13:04:23 -0700, Aaron Brady wrote:
>

>> <snip>

>>
>> I think the technique would require two python longs and a bool in the
>> set, and a python long in the iterator.
>>
>> One long counts the number of existing (open) iterators. Another counts
>> the version. The bool keeps track of whether an iterator has been
>> created since the last modification, in which case the next modification
>> requires incrementing the version and resetting the flag.
> I think that is over-engineered and could be the difference between
> having the patch accepted and having it rejected. After all, many people
> will argue that the existing solution to the problem is good enough.
>
> Dicts are extremely common, and your patch increases both the memory
> usage of every dict, and the overhead of every write operation
> (__setitem__, __delitem__, update). Only a very few dicts will actually
> need this overhead, for the rest it is waste. It is important to keep
> that waste to a minimum or risk having the patch rejected.
>
> An unsigned C int can count up to 4,294,967,295. I propose that you say
> that is enough iterators for anyone, and use a single, simple, version
> counter in the dict and the iterator. If somebody exceeds that many
> iterators to a single dict or set,

I think you have the count confused. it has to be a count of how many
changes have been made to the dict or set, not how many iterators exist.

> and the version field overflows by
> exactly 2**32 versions, the results are no worse than they are now. You
> won't be introducing any new bugs.
>
> Complicating the solution is, in my opinion, unnecessary. Why should
> every set and dict carry the cost of incrementing TWO Python longs and a
> flag when just a single C int is sufficient for all realistic use-cases?
>
> The opportunity for failure is extremely narrow:
>
> - I must have an iterator over a dict or set;
> - and I must have paused the iteration in some way;
> - and then I must create exactly 2**32 other iterators to the
> same dict or set;
> - and at some point modify the dict or set
> - and then restart the first iterator
>
> at which point some items returned by the iterator *may* be duplicated or
> skipped (depends on the nature of the modifications).
>
>

I agree with almost your entire point, exact that the place where it
would fail to detect a bug is when somebody has modified the dict
exactly 2**32 times while an iterator is paused. See my own response,
of 4:27pm. (my time).

--

DaveA

[Steven D'Aprano]

On Mon, 03 Sep 2012 21:50:57 -0400, Dave Angel wrote:

> On 09/03/2012 09:26 PM, Steven D'Aprano wrote:

>> An unsigned C int can count up to 4,294,967,295. I propose that you say
>> that is enough iterators for anyone, and use a single, simple, version
>> counter in the dict and the iterator. If somebody exceeds that many
>> iterators to a single dict or set,
>
> I think you have the count confused. it has to be a count of how many
> changes have been made to the dict or set, not how many iterators exist.

Oops, yes you are absolutely right. It's a version number, not a count of
iterators.


--
Steven

[Aaron Brady]

Hello. We have a number of proposed solutions so far.

1) Collection of iterators
a) Linked list
i) Uncounted references
ii) Counted references
iii) Weak references
b) Weak set
2) Serial index / timestamp
a) No overflow - Python longs
b) Overflow - C ints / shorts / chars
c) Reset index if no iterators left
3) Iterator count
- Raise exception on set modifications, not iteration

Note, "2b" still leaves the possibility of missing a case and letting an error pass silently, as the current behavior does. The rest catch the error 100% of the time.

Anyway, I plan to develop the above patch for the 'dict' class. Would anyone like to take over or help me do it?

[Thomas Rachel]

Am 19.08.2012 00:14 schrieb MRAB:

>> Can someone who is more familiar with the cycle detector and cycle
>> breaker, help prove or disprove the above?
>>
> In simple terms, when you create an immutable object it can contain
> only references to pre-existing objects, but in order to create a cycle
> you need to make an object refer to another which is created later, so
> it's not possible to create a cycle out of immutable objects.

Yes, but if I add a list in-between, I can create a refcycle:

a = []
b = (a,)
a.append(b)

So b is a tuple consisting of one list which in turn contains b.

It is not a direct cycle, but an indirect one.

Or would that be detected via the list?


Thomas

[Hans Mulder]

It's a cycle and it contains an immutable object, but it's not
a cycle consisting of immutable objects only.

As MRAB was arguing: it is not possbible to create a cycle
consisting of immutable objects only (unless you do unspeakable
things at the C level).

-- HansM

[MRAB]

The quote was:

'''
...The tuple type does not implement a tp_clear function, because it�s

possible to prove that no reference cycle can be composed entirely of
tuples.
'''

Note: "composed entirely of tuples".

Or, in general, composed entirely of immutables.

Lists are not immutable, therefore the proof does not apply.

[Aaron Brady]

On Thursday, August 23, 2012 1:11:14 PM UTC-5, Steven D'Aprano wrote:
> On Thu, 23 Aug 2012 09:49:41 -0700, Aaron Brady wrote:
>
>
>
> [...]
>
> > The patch for the above is only 40-60 lines. However it introduces two
>
> > new concepts.
>
> >
>
> > The first is a "linked list", a classic dynamic data structure, first
>
> > developed in 1955, cf. http://en.wikipedia.org/wiki/Linked_list .
>
> > Linked lists are absent in Python
>
>
>
> They certainly are not. There's merely no named "linked list" class.
>
>
>
> Linked lists are used by collections.ChainMap, tracebacks, xml.dom,
>
> Abstract Syntax Trees, and probably many other places. (Well, technically
>
> some of these are trees rather than lists.) You can trivially create a
>
> linked list:
>
>
>
> x = [a, [b, [c, [d, [e, None]]]]]
>
>
>
> is equivalent to a singly-linked list with five nodes. Only less
>
> efficient.
>

[snip.]
>
> --
>
> Steven

That's not totally true. Your formulation is equivalent to a single-linked list, but a double-linked list can't be represented as an expression because it contains reference cycles.

Single and double-linked lists have the same requirements for inserting a new node. For instance:

[a, [b, [c, [d, [e, None]]]]]

[a, [a2, [b, [c, [d, [e, None]]]]]]

However, to remove a node, the single-linked list requires iterating over the entire list to find the predecessor of the target, whereas the double-linked list already contains that information.

[a, [b, [c, [d, [e, None]]]]]

[a, [b, [c, [e, None]]]]

The difference might be moot in some applications, such as if we only remove nodes when we're already iterating.

Memory constraints were more severe 50 years ago when the structure was developed. The difference between one pointer and two in a structure could have a big impact in repetition. The single-linked list admits more easily of recursive algorithms as well.

[88888 Dihedral]

Paul Rubin於 2012年8月17日星期五UTC+8上午9時01分39秒寫道：
> Ian Kelly <ian.g...@gmail.com> writes:
>
> > With regard to key insertion and deletion while iterating over a dict
>
> > or set, though, there is just no good reason to be doing that
>
> > (especially as the result is very implementation-specific), and I
>
> > wouldn't mind a more complete low-level check against it as long as
>
> > it's not too expensive (which is not clearly the case with the current
>
> > suggestion at all).
>
>
>
> One possible approach is to freeze the dictionary against modification
>
> while any iterator is open on it. You could keep a count of active
>
> iterators in the dict structure, adjusting it whenever an iterator is
>
> created or closed/destroyed.

If there is only one iterator of a frozen dictionary,
then nothing is saved.

But if there are manny iterators based on the same frozen dictionary,
this approach saves a lot.