Charming Python #b25: Python Elegance, Python Warts

Sequences and comparisons


David Mertz, Ph.D.
Frog Prince, Gnosis Software
May, 2005

Since the "golden age" of Python 1.5.2--for a long time a stable and solid version--Python has greatly increased its number of syntactic features and built-in functions and types. Each of these additions has reasonable justification, in isolation, but taken as a whole, they make Python no longer a language that experienced programmers can pick up "in an afternoon." Moreover, some of the changes have pitfalls along with benefits.

Introduction

In this article, David discusses the non-obvious features and misfeatures that have been added to the last several Python versions; and weighs in on which are truly valuable, and which unnecessary complication. David hopes to provide a collection of valuable things to watch out for to all those programmers who use Python less than full time--either programmer in other languages, or people like scientists for whom programming is only a side task. Where some quandries are raised, solutions are suggested.

Incomparable Abominations

Between Python 2.0 and Python 2.1 something strange happened. Previously comparable objects started raising exceptions when compared. Specifically, complex numbers became incomparable to other numbers, including both other complex numbers and ints, floats, longs. Actually, the problem arose earlier than this with Unicode strings comparing to plain strings, but only in some edge cases.

To my mind, the changes are both grating just plain weird. Back in that golden age of 1.5.2, you knew the inequality operators would return a result regardless of which objects were compared. Sure the result would not necessarily be meaningful--a string is neither objectively less than nor greater than a float. But some consistent result would obtain.

After the change, some Pythonistas argued that a good behavior would be to disallow all inequality comparisons between objects of distinct types, at least unless we define custom comparison functions. My hunch is that this could get tricky in practice once you deal with custom classes and multiple inheritence. Moreover, not being able to compare among floats, ints and longs (or e.g. decimal) would be awkward. But maybe a sensible rule could be defined.

However, whatever rule that might be, it would be very different than what Python did historically. And what we are left with now is dramatically irregular comparison behavior in which even knowing the types of compared objects doesn't tell you whether they are comparable (and inequality is not transitive or closed):

Success of comparisons depends on both types and values

>>> map(type, (u1, s1, s2))
[<type 'unicode'>, <type 'str'>, <type 'str'>]
>>> u1 < s1
True
>>> s1 < s2
True
>>> u1 < s2
UnicodeDecodeError: 'ascii' codec can't decode byte 0xf0 in position 0:
    ordinal not in range(128)
>>> map(type, (n, j, u1))
[<type 'int'>, <type 'complex'>, <type 'unicode'>]
>>> n < u1
True
>>> j < u1
True
>>> n < j
TypeError: no ordering relation is defined for complex numbers

Adding insult to injury, complex numbers are now incomparable to most numeric values, but claim a definite inequality from most non-numeric values. I can get that holding theoretical purity in mind, 1+1j, for example, is neither more nor less than 2-3j, but then why have this:

Surprising comparison results

>>> 2-3j < 'spam'
True
>>> 4+0j < decimal.Decimal('3.14')
True
>>> 4+0j < 5+0j
TypeError: no ordering relation is defined for complex numbers

None of that is particularly "pure" theoretically.

A real wart: sorting heterogeneous collections

An argument is sometimes made that it is a programming mistake to try to compare items of incomensurable types. But Python is very happy to perform many such comparisons; and doing so meshes well with a philosophy of "duck typing" (it is not what an object is, but what it does). Python collections frequently group together objects on non-identical types with the hope of being able to do something similar with each such collected object. A frequent use case is that you often want to encode a bunch of disparate values for transmission over some protocol.

For most values of "do", inequality comparisons are not necessarily needed. However, one very common case where inequalities are implicitly extremely useful is in sorting collection, usually lists or list-like custom collections. Sometimes its meaningful to process a collection of things in a meaningfully ascending order (e.g. data values from smallest to largest). Other times, it is simply useful to create a stable order of multiple collections, particularly to run a sort of "list diff" on the two collections. That is, perhaps you want to take one type of action is an object is in both collections, but a different action if it is in only one collection. Constantly asking if x in otherlist runs into geometric big-O inefficiencies; marching in parallel between two stably-sorted lists is more efficient. For example:

Action depending on two-list membership

list1.sort()
list2.sort()
list2_xtra = []
list2_ndx = 0
for it1 in list1:
    it2 = list2[list2_ndx]
    while it1 < it2:
        list2_ndx += 1
        it2 = list2[list2_ndx]
        if it1 == it2:
            item_in_both(it1)
        elif it1 > it2:
            item_in_list1(it1)
        else:
            list2_xtra.appen(it2)
 for it2 in list2_xtra:
     item_in_list2(it2)

Sometimes as well, having "local sequences" of meaningful comparisons is nice, even where heterogeneous items also occur (e.g. process all the floating values "in order" even though they are not in any deep sense greater or less than the strings processed elsewhere).

Sorting falling down

Of course, the code above to perform a "list diff" is subject to blowing up, almost randomly. For example, here is a collection of small lists one might encounter for list1 and list2. Try to guess which are sortable:

Hodgepodge of sortables and unsortables

['x','y','z', 1],
['x','y','z', 1j],
['x','y','z', 1j, 1],       # Adding an element makes it unsortable
[0j, 1j, 2j],               # An obvious "natural" order
[0j, 1, 2],
[0, 1, 2],                  # Notice that 0==0j --> True
[chr(120), chr(240)],
[chr(120), chr(240), 'x'],
[chr(120), chr(240), u'x'], # Notice u'x'=='x' --> True
[u'a', 'b', chr(240)],
[chr(240), u'a', 'b']       # Same items, different initial order

I wrote a short program to try sorting each list:

% python compare.py
(0)  ['x', 'y', 'z', 1] --> [1, 'x', 'y', 'z']
(1)  ['x', 'y', 'z', 1j] --> [1j, 'x', 'y', 'z']
(2)  ['x', 'y', 'z', 1j, 1] --> exceptions.TypeError
(3)  [0j, 1j, 2j] --> exceptions.TypeError
(4)  [0j, 1, 2] --> exceptions.TypeError
(5)  [0, 1, 2] --> [0, 1, 2]
(6)  ['x', '\xf0'] --> ['x', '\xf0']
(7)  ['x', '\xf0', 'x'] --> ['x', 'x', '\xf0']
(8)  ['x', '\xf0', u'x'] --> exceptions.UnicodeDecodeError
(9)  [u'a', 'b', '\xf0'] --> [u'a', 'b', '\xf0']
(10) ['\xf0', u'a', 'b'] --> exceptions.UnicodeDecodeError

Some of these results more-or-less follow from the prior caveats. But look at (9) and (10) which contain exactly the same objects in different orders: the failure depends not only on the type and values of the items in the list, but on the specific implementation of the list.sort() method!

Fixing comparisons

Along the path away from 1.5.2, Python grew a very useful datatype: sets, first as a standard module then as a built-in (the module still contains some extras). For a lot of the problem I describe above, simply using sets rather than lists lets you easily answer the question of what items are in one collection, or the other, or in both, all without having to roll your own "list diff" code. For example:

Sets and set operations

>>> set1 = set([1j, u'2', 3, 4.0])
>>> set2 = set([4, 3, 2, 1])
>>> set1 | set2
set([3, 1, 2, 1j, 4.0, u'2'])
>>> set1 & set2
set([3, 4])

I discovered something rather odd in composing the above example. Set operations apparently use equality rather than identity. Perhaps there is some sense to this, but it strikes me as peculiar that the union of the two sets contains the float 4.0 while their intersection contains the int 4. Or more specifically, what exact value gets included is order-sensitive, despite the set-theoretic symmetry of the union and intersection operations:

Odd type results in sets

>>> set2 & set1
set([3, 4.0])
>>> set([3, 4.0, 4, 4+0j])
set([3, 4.0])

Still, at as a first pass, sets are a wonderful datatype. Nonetheless, it is worth keeping custom comparisons in mind as a workaround. Prior to Python 2.4, it was possible to implement a custom cmp() function to pass to list.sort(). That would work to implement comparisons for otherwise incomparable objects; the problem with the cmp argument is that it calls the function on every comparison: Python's call overhead is relatively high, but worse than this, computed values wind up being computed multiple times.

The efficient solution to cmp inefficiency is to use a Schwartzian sort instead: decorate each item, sort, then undecorate. Unfortunately, that requires a bit of custom code, rather than a simple call to list.sort(). Python 2.4 finds a good combination solution using the key argument. This argument just takes a function that returns a decorated object, and does the Schwartzian sort mechanics "behind the scenes". Keeping in mind the fact that complex numbers are incomparable to even each other, while unicode objects only have problems comparing to some strings, we can use:

A stable and universal sort key

def stablesort(o):
    # Use as: mylist.sort(key=stablesort)
    if type(o) is complex:
        return (type(o), o.real, o.imag)
    else:
        return (type(o), o)

Mind you, the order of elements might not be strictly what you expect: it is not identical to an undecorated sort, even where the undecorated sort succeeds. In particular, elements like different numeric types are no longer intermixed, but separated into different parts of the sorted result. But at least it is stable, and will succeed on almost any list (if you try, you can still get a custom object to make the sort blow up).

Generators As Not-quite-sequences

Over several versions, Python has hugely enhanced its "laziness". For several versions, we have had generators defined with the yield statement in a function body. But along the way we also got the itertools modules to combine and create various types of iterators. We have the iter() built-in function to turn many sequence-like objects into iterators. With Python 2.4, we got generator expressions; and with 2.5 we will get enhanced generators that make writing coroutines easier. Moreover, more and more Python objects have become iterators or iterator-like; for example, what used to require the .xreadlines() method or before that the xreadlines module, is now simply the default behavior of open() to read files.

Similarly, looping through a dict lazily used to require the .iterkeys() method; now it is just the default for key in dct behavior. Functions like xrange() are a bit "special" in being generator-like, but neither quite a real iterator (no .next() method), nor a realized list like range() returns. However, enumerate() returns a true generator, and usually does what you had earlier wanted xrange() for. And itertools.count() is another lazy call that does almost the same thing as xrange(), but as a full-fledged iterator.

Python is strongly moving towards lazily constructing sequence-like objects; and overall this is an excellent direction. Lazy pseudo-sequences both save memory space and speeds up operations (especially when dealing with very large sequence-like "things").

The problem is that Python still has a schizoaffective condition when it comes to deciding what the differences and similarities between "hard" sequences and iterators are. The troublesome part of this is that it really violates Python's idea of "duck typing": the ability to use a given object for a purpose, just as long as it has the right behaviors, but not necessarily any inheritance or type restriction. The various things that are iterators or iterator-like sometimes act sequence-like, but other times do not; conversely, sequences often act iterator-like, but not always. Outside of those steeped in Python arcana, what does what is not obvious.

Divergences

The main point of similarity is that everything that is sequence- or iterator-like lets you loop over it, whether using a for loop, a list comprehension, or a generator comprehension. Past that, divergences occur. The most important of these differences is that sequences can be indexed, and directly sliced, while iterators cannot. In fact, indexing into a sequence is probably the most common thing you ever do with a sequence--why on earth does it fall down so badly on iterators? For example:

Sequence-like and iterator-like things

>>> r = range(10)
>>> i = iter(r)
>>> x = xrange(10)
>>> g = itertools.takewhile(lambda n: n<10, itertools.count())
#...etc...

For all of these, you can use for n in thing. In fact, if you "concretize" any of them with list(thing) you wind up with exactly the same result. But if you wish to obtain a specific item--or a slice of a few items--you need to start caring about the exact type of thing. E.g.:

When indexing succeeds and fails

>>> r[4]
4
>>> i[4]
TypeError: unindexable object

With enough contortions, you can get an item for every type of sequence/iterator. One way is to loop until you get there. Another hackish combination might be something like:

Contortions to obtain an index

>>> thing, temp = itertools.tee(thing)
>>> zip(temp, '.'*5)[-1][0]
4

The pre-call to itertools.tee() preserves the original iterator. For a slice, you might use the itertools.islice() function, wrapped up in contortions.

Contortions to obtain a slice

>>> r[4:9:2]
[4, 6, 8]
>>> list(itertools.islice(r,4,9,2))  # works for iterators
[4, 6, 8]

A class wrapper

You might combine these techniques into a class wrapper for convenience, using some magic methods:

Making iterators indexable

>>> class Indexable(object):
...     def __init__(self, it):
...         self.it = it
...     def __getitem__(self, x):
...         self.it, temp = itertools.tee(self.it)
...         if type(x) is slice:
...             return list(itertools.islice(self.it, x.start, x.stop, x.step))
...         else:
...             return zip(temp, range(x+1))[-1][0]
...     def __iter__(self):
...         self.it, temp = itertools.tee(self.it)
...         return temp
...
>>> integers = Indexable(itertools.count())
>>> integers[4]
4
>>> integers[4:9:2]
[4, 6, 8]

So with some effort, you can coax an object to behave like both a sequence and an iterator. But this much effort should really not be necessary; indexing and slicing should "just work" whether a concrete sequence or a iterator is involved.

Notice that the Indexable class wrapper is still not as flexible as might be desirable. The main problem is that we create a new copy of the iterator every time--a better approach would be to cache the head of the sequence when we slice it, then use that cached head for future access of elements already examined. Of course, there is a tradeoff between memory used and the speed penalty of running through the iterator. Nonetheless, the best thing would be if Python itself would do all of this "behind the scenes"--the behavior might be fine-tuned somehow by "power users", but average programmers should not have to think about any of this.

Resources

Andrew Kuchling wrote a pretty well known page about "Python Warts" (though it hasn't changed in a couple years):

http://www.amk.ca/python/writing/warts.html

Frank McIngvale, my coauthor on Gnosis Utilities, has written an excellent discussion of issues with Python and Unicode, including his motivation for including some enhanced Unicode handling facilities in Gnosis Utilities. Frank's essay is titled, charmingly, "All About Python and Unicode... and even more about Unicode:"

http://boodebr.org/python/pyunicode/index.html

Incidentally, to see what is in the latest Gnosis Utilities--currently including Frank McIngvale's Unicode tools added in version 1.2.0--take a look at:

http://www.gnosis.cx/download/Gnosis_Utils.ANNOUNCE

About The Author

Picture of Author David Mertz almost enjoys problems because of the solutions they enable. David may be reached at mertz@gnosis.cx; his life pored over athttp://gnosis.cx/publish/. Check out David's book Text Processing in Python (http://gnosis.cx/TPiP/).