Does Python have an ordered set?

The answer is no, but you can use the simple dict from the Python standard library with just keys (and values as None) for the same purpose. As of Python 3.7 (and CPython 3.6), dict is guaranteed to preserve order.

Here's an example of how to use dict as an ordered set to filter out duplicate items while preserving order, thereby emulating an ordered set. Use the dict class method fromkeys() to create a dict, then simply ask for the keys() back.

>>> keywords = ['foo', 'bar', 'bar', 'foo', 'baz', 'foo']

>>> list(dict.fromkeys(keywords))
['foo', 'bar', 'baz']

That said, for backward compatibility and/or readability, you may wish to use collections.OrderedDict.

Update: This answer is obsolete as of Python 3.7. See jrc's answer above for a better solution. Will keep this answer here only for historical reasons.

An ordered set is functionally a special case of an ordered dictionary.

The keys of a dictionary are unique. Thus, if one disregards the values in an ordered dictionary (e.g. by assigning them None), then one has essentially an ordered set.

As of Python 3.1 and 2.7 there is collections.OrderedDict. The following is an example implementation of an OrderedSet. (Note that only few methods need to be defined or overridden: collections.OrderedDict and collections.MutableSet do the heavy lifting.)

import collections

class OrderedSet(collections.OrderedDict, collections.MutableSet):

    def update(self, *args, **kwargs):
        if kwargs:
            raise TypeError("update() takes no keyword arguments")

        for s in args:
            for e in s:
                 self.add(e)

    def add(self, elem):
        self[elem] = None

    def discard(self, elem):
        self.pop(elem, None)

    def __le__(self, other):
        return all(e in other for e in self)

    def __lt__(self, other):
        return self <= other and self != other

    def __ge__(self, other):
        return all(e in self for e in other)

    def __gt__(self, other):
        return self >= other and self != other

    def __repr__(self):
        return 'OrderedSet([%s])' % (', '.join(map(repr, self.keys())))

    def __str__(self):
        return '{%s}' % (', '.join(map(repr, self.keys())))
    
    difference = property(lambda self: self.__sub__)
    difference_update = property(lambda self: self.__isub__)
    intersection = property(lambda self: self.__and__)
    intersection_update = property(lambda self: self.__iand__)
    issubset = property(lambda self: self.__le__)
    issuperset = property(lambda self: self.__ge__)
    symmetric_difference = property(lambda self: self.__xor__)
    symmetric_difference_update = property(lambda self: self.__ixor__)
    union = property(lambda self: self.__or__)

Implementations on PyPI

While others have pointed out that there is no built-in implementation of an insertion-order preserving set in Python (yet), I am feeling that this question is missing an answer which states what there is to be found on PyPI.

There are the packages:

• ordered-set (Python based)
• collections-extended
• boltons (under setutils.IndexedSet, Python-based)
• oset (last updated in 2012)

Some of these implementations are based on the recipe posted by Raymond Hettinger to ActiveState which is also mentioned in other answers here.

Some differences

• ordered-set (version 1.1)
• advantage: O(1) for lookups by index (e.g. my_set[5])
• oset (version 0.1.3)
• advantage: O(1) for remove(item)
• disadvantage: apparently O(n) for lookups by index

Both implementations have O(1) for add(item) and __contains__(item) (item in my_set).

I can do you one better than an OrderedSet: boltons has a pure-Python IndexedSet type that is not only an ordered set, but also supports indexing (as with lists).

Simply pip install boltons (or copy setutils.py into your codebase), import the IndexedSet and:

>>> from boltons.setutils import IndexedSet
>>> x = IndexedSet(list(range(4)) + list(range(8)))
>>> x
IndexedSet([0, 1, 2, 3, 4, 5, 6, 7])
>>> x - set(range(2))
IndexedSet([2, 3, 4, 5, 6, 7])
>>> x[-1]
7
>>> fcr = IndexedSet('freecreditreport.com')
>>> ''.join(fcr[:fcr.index('.')])
'frecditpo'

Everything is unique and retained in order. Full disclosure: I wrote the IndexedSet, but that also means you can bug me if there are any issues.

If you're using the ordered set to maintain a sorted order, consider using a sorted set implementation from PyPI. The sortedcontainers module provides a SortedSet for just this purpose. Some benefits: pure-Python, fast-as-C implementations, 100% unit test coverage, hours of stress testing.

Installing from PyPI is easy with pip:

pip install sortedcontainers

Note that if you can't pip install, simply pull down the sortedlist.py and sortedset.py files from the open-source repository.

Once installed you can simply:

from sortedcontainers import SortedSet
help(SortedSet)

The sortedcontainers module also maintains a performance comparison with several alternative implementations.

For the comment that asked about Python's bag data type, there's alternatively a SortedList data type which can be used to efficiently implement a bag.

As other answers mention, as for python 3.7+, the dict is ordered by definition. Instead of subclassing OrderedDict we can subclass abc.collections.MutableSet or typing.MutableSet using the dict's keys to store our values.

import typing

T = typing.TypeVar("T")


class OrderedSet(typing.MutableSet[T]):
    """A set that preserves insertion order by internally using a dict."""

    def __init__(self, iterable: typing.Iterable[T]):
        self._d = dict.fromkeys(iterable)

    def add(self, value: T) -> None:
        self._d[value] = None

    def discard(self, value: T) -> None:
        self._d.pop(value, None)

    def __contains__(self, value: object) -> bool:
        return self._d.__contains__(value)

    def __len__(self) -> int:
        return self._d.__len__()

    def __iter__(self) -> typing.Iterator[T]:
        return self._d.__iter__()

    def __str__(self):
        return f"{{{', '.join(str(i) for i in self)}}}"

    def __repr__(self):
        return f"<OrderedSet {self}>"

Then just:

x = OrderedSet([1, 2, -1, "bar"])
x.add(0)
assert list(x) == [1, 2, -1, "bar", 0]

I added this code, with some tests, in a small library, so anyone can just pip install it.

In case you're already using pandas in your code, its Index object behaves pretty like an ordered set, as shown in this article.

Examples from the article:

indA = pd.Index([1, 3, 5, 7, 9])
indB = pd.Index([2, 3, 5, 7, 11])

indA & indB  # intersection
indA | indB  # union
indA - indB  # difference
indA ^ indB  # symmetric difference

A little late to the game, but I've written a class setlist as part of collections-extended that fully implements both Sequence and Set

>>> from collections_extended import setlist
>>> sl = setlist('abracadabra')
>>> sl
setlist(('a', 'b', 'r', 'c', 'd'))
>>> sl[3]
'c'
>>> sl[-1]
'd'
>>> 'r' in sl  # testing for inclusion is fast
True
>>> sl.index('d')  # so is finding the index of an element
4
>>> sl.insert(1, 'd')  # inserting an element already in raises a ValueError
ValueError
>>> sl.index('d')
4

GitHub: https://github.com/mlenzen/collections-extended

Documentation: http://collections-extended.lenzm.net/en/latest/

PyPI: https://pypi.python.org/pypi/collections-extended

There's no OrderedSet in official library. I make an exhaustive cheatsheet of all the data structure for your reference.

DataStructure = {
    'Collections': {
        'Map': [
            ('dict', 'OrderDict', 'defaultdict'),
            ('chainmap', 'types.MappingProxyType')
        ],
        'Set': [('set', 'frozenset'), {'multiset': 'collection.Counter'}]
    },
    'Sequence': {
        'Basic': ['list', 'tuple', 'iterator']
    },
    'Algorithm': {
        'Priority': ['heapq', 'queue.PriorityQueue'],
        'Queue': ['queue.Queue', 'multiprocessing.Queue'],
        'Stack': ['collection.deque', 'queue.LifeQueue']
        },
    'text_sequence': ['str', 'byte', 'bytearray']
}

As others have said, OrderedDict is a superset of an ordered set in terms of functionality, but if you need a set for interacting with an API and don't need it to be mutable, OrderedDict.keys() is actually an implementation abc.collections.Set:

import random
from collections import OrderedDict, abc

a = list(range(0, 100))
random.shuffle(a)

# True
a == list(OrderedDict((i, 0) for i in a).keys())

# True
isinstance(OrderedDict().keys(), abc.Set)   

The caveats are immutability and having to build up the set like a dict, but it's simple and only uses built-ins.

The ParallelRegression package provides a setList( ) ordered set class that is more method-complete than the options based on the ActiveState recipe. It supports all methods available for lists and most if not all methods available for sets.

There is a pip library that does this:

pip install ordered-set

Then you can use it:

from ordered_set import OrderedSet

Just use pd.unique from pandas - does exactly what you need!

>>> import pandas as pd
>>> pd.unique([3, 1, 4, 5, 2, 2])
array([3, 1, 4, 5, 2])

This answer is for completeness sake. If the length of your set is small, and your code is single-threaded, a list could work just fine as it's implicitly ordered.

if not new_item in my_list:
    my_list.append(new_item)

If using this approach:

• To append or remove an item, first check for presence as in the code above.
• To compare equality, use set(my_list).

Checking for presence in a list of course has an O(n) complexity, but this could be acceptable for a small list, especially if high performance isn't required.