Abstract

    This PEP suggests a "sort by value" operation for dictionaries.
    The primary benefit would be in terms of "batteries included"
    support for a common Python idiom which, in its current form, is
    both difficult for beginners to understand and cumbersome for all
    to implement.

BDFL Pronouncement

    This PEP is rejected because the need for it has been largely
    fulfilled by Py2.4's sorted() builtin function:

        >>> sorted(d.iteritems(), key=itemgetter(1), reverse=True)
        [('b', 23), ('d', 17), ('c', 5), ('a', 2), ('e', 1)]

    or for just the keys:

        sorted(d, key=d.__getitem__, reverse=True)
        ['b', 'd', 'c', 'a', 'e']

    Also, Python 2.5's heapq.nlargest() function addresses the common use
    case of finding only a few of the highest valued items:

        >>> nlargest(2, d.iteritems(), itemgetter(1))
        [('b', 23), ('d', 17)]

Motivation

    A common use of dictionaries is to count occurrences by setting
    the value of d[key] to 1 on its first occurrence, then increment
    the value on each subsequent occurrence.  This can be done several
    different ways, but the get() method is the most succinct:

            d[key] = d.get(key, 0) + 1

    Once all occurrences have been counted, a common use of the
    resulting dictionary is to print the occurrences in
    occurrence-sorted order, often with the largest value first.

    This leads to a need to sort a dictionary's items by value.  The
    canonical method of doing so in Python is to first use d.items()
    to get a list of the dictionary's items, then invert the ordering
    of each item's tuple from (key, value) into (value, key), then
    sort the list; since Python sorts the list based on the first item
    of the tuple, the list of (inverted) items is therefore sorted by
    value.  If desired, the list can then be reversed, and the tuples
    can be re-inverted back to (key, value).  (However, in my
    experience, the inverted tuple ordering is fine for most purposes,
    e.g. printing out the list.)

    For example, given an occurrence count of:

        >>> d = {'a':2, 'b':23, 'c':5, 'd':17, 'e':1}

    we might do:

        >>> items = [(v, k) for k, v in d.items()]
        >>> items.sort()
        >>> items.reverse()             # so largest is first
        >>> items = [(k, v) for v, k in items]

    resulting in:

        >>> items
        [('b', 23), ('d', 17), ('c', 5), ('a', 2), ('e', 1)]

    which shows the list in by-value order, largest first.  (In this
    case, 'b' was found to have the most occurrences.)

    This works fine, but is "hard to use" in two aspects.  First,
    although this idiom is known to veteran Pythoneers, it is not at
    all obvious to newbies -- either in terms of its algorithm
    (inverting the ordering of item tuples) or its implementation
    (using list comprehensions -- which are an advanced Python
    feature.)  Second, it requires having to repeatedly type a lot of
    "grunge", resulting in both tedium and mistakes.

    We therefore would rather Python provide a method of sorting
    dictionaries by value which would be both easy for newbies to
    understand (or, better yet, not to _have to_ understand) and
    easier for all to use.

Rationale

    As Tim Peters has pointed out, this sort of thing brings on the
    problem of trying to be all things to all people.  Therefore, we
    will limit its scope to try to hit "the sweet spot".  Unusual
    cases (e.g. sorting via a custom comparison function) can, of
    course, be handled "manually" using present methods.

    Here are some simple possibilities:

    The items() method of dictionaries can be augmented with new
    parameters having default values that provide for full
    backwards-compatibility:

        (1) items(sort_by_values=0, reversed=0)

    or maybe just:

        (2) items(sort_by_values=0)

    since reversing a list is easy enough.

    Alternatively, items() could simply let us control the (key, value) 
    order:

        (3) items(values_first=0)

    Again, this is fully backwards-compatible.  It does less work than
    the others, but it at least eases the most complicated/tricky part
    of the sort-by-value problem: inverting the order of item tuples.
    Using this is very simple:

        items = d.items(1)
        items.sort()
        items.reverse()         # (if desired)

    The primary drawback of the preceding three approaches is the
    additional overhead for the parameter-less "items()" case, due to
    having to process default parameters.  (However, if one assumes
    that items() gets used primarily for creating sort-by-value lists,
    this is not really a drawback in practice.)

    Alternatively, we might add a new dictionary method which somehow
    embodies "sorting".  This approach offers two advantages.  First,
    it avoids adding overhead to the items() method.  Second, it is
    perhaps more accessible to newbies: when they go looking for a
    method for sorting dictionaries, they hopefully run into this one,
    and they will not have to understand the finer points of tuple
    inversion and list sorting to achieve sort-by-value.

    To allow the four basic possibilities of sorting by key/value and in 
    forward/reverse order, we could add this method:

        (4) sorted_items(by_value=0, reversed=0)

    I believe the most common case would actually be "by_value=1,
    reversed=1", but the defaults values given here might lead to
    fewer surprises by users: sorted_items() would be the same as
    items() followed by sort().

    Finally (as a last resort), we could use:

        (5) items_sorted_by_value(reversed=0)

Implementation

    The proposed dictionary methods would necessarily be implemented
    in C.  Presumably, the implementation would be fairly simple since
    it involves just adding a few calls to Python's existing
    machinery.

Concerns

    Aside from the run-time overhead already addressed in
    possibilities 1 through 3, concerns with this proposal probably
    will fall into the categories of "feature bloat" and/or "code
    bloat".  However, I believe that several of the suggestions made
    here will result in quite minimal bloat, resulting in a good
    tradeoff between bloat and "value added".

    Tim Peters has noted that implementing this in C might not be
    significantly faster than implementing it in Python today.
    However, the major benefits intended here are "accessibility" and
    "ease of use", not "speed".  Therefore, as long as it is not
    noticeably slower (in the case of plain items(), speed need not be
    a consideration.

References

    A related thread called "counting occurrences" appeared on
    comp.lang.python in August, 2001.  This included examples of
    approaches to systematizing the sort-by-value problem by
    implementing it as reusable Python functions and classes.

Copyright

    This document has been placed in the public domain.