349

Dictionaries are ordered in Python 3.6 (under the CPython implementation at least) unlike in previous incarnations. This seems like a substantial change, but it's only a short paragraph in the documentation. It is described as a CPython implementation detail rather than a language feature, but also implies this may become standard in the future.

How does the new dictionary implementation perform better than the older one while preserving element order?

Here is the text from the documentation:

dict() now uses a “compact” representation pioneered by PyPy. The memory usage of the new dict() is between 20% and 25% smaller compared to Python 3.5. PEP 468 (Preserving the order of **kwargs in a function.) is implemented by this. The order-preserving aspect of this new implementation is considered an implementation detail and should not be relied upon (this may change in the future, but it is desired to have this new dict implementation in the language for a few releases before changing the language spec to mandate order-preserving semantics for all current and future Python implementations; this also helps preserve backwards-compatibility with older versions of the language where random iteration order is still in effect, e.g. Python 3.5). (Contributed by INADA Naoki in issue 27350. Idea originally suggested by Raymond Hettinger.)

Update December 2017: dicts retaining insertion order is guaranteed for Python 3.7

    
improve this question
  • 2
    See this thread on Python-Dev mailing-list : mail.python.org/pipermail/python-dev/2016-September/146327.htmlif you haven't seen it ; it's basically a discussion around these subjects. – mgc Oct 11 '16 at 15:11
  • 5
    Notice that a long time ago (2003), Perl implementers decided to make hash tables (equivalent to Python dictionaries) not only explicitly unordered, but randomized for security reasons (perldoc.perl.org/perlsec.html#Algorithmic-Complexity-Attacks). So I would definitely not count on this "feature", because if experience of others may be a guide, it's probably deemed to be reversed at some point... – wazooxOct 19 '16 at 13:50
  • 1
    If kwargs are now supposed to be ordered (which is nice idea) and kwargs are dict, not OrderedDict, then I guess one could assume that dict keys will stay ordered in the future version of Python, despite the documentation says otherwise. – Dmitriy Sintsov Jan 12 '17 at 12:32
  • 4
    @DmitriySintsov No, don't make that assumption. This was an issue brought up during the writing of the PEP that defines order preserving feature of **kwargs and as such the wording used is diplomatic: **kwargs in a function signature is now guaranteed to be an insertion-order-preserving mapping. They've used the term mapping in order to not force any other implementations to make the dict ordered (and use an OrderedDictinternally) and as a way to signal that this isn't supposed to depend on the fact that the dict is not ordered. – Jim Fasarakis Hilliard Feb 4 '17 at 17:18
  • 6
    A good video explanation from Raymond Hettinger – Alex Jul 22 '17 at 16:38
382

Are dictionaries ordered in Python 3.6+?

They are insertion ordered[1]. As of Python 3.6, for the CPython implementation of Python, dictionaries remember the order of items inserted. This is considered an implementation detail in Python 3.6; you need to use OrderedDict if you want insertion ordering that's guaranteed across other implementations of Python (and other ordered behavior[1]).

As of Python 3.7, this is no longer an implementation detail and instead becomes a language feature. From a python-dev message by GvR:

Make it so. "Dict keeps insertion order" is the ruling. Thanks!

This simply means that you can depend on it. Other implementations of Python must also offer an insertion ordered dictionary if they wish to be a conforming implementation of Python 3.7.

How does the Python 3.6 dictionary implementation perform better[2] than the older one while preserving element order?

Essentially, by keeping two arrays.

  • The first array, dk_entries, holds the entries (of type PyDictKeyEntry) for the dictionary in the order that they were inserted. Preserving order is achieved by this being an append only array where new items are always inserted at the end (insertion order).

  • The second, dk_indices, holds the indices for the dk_entries array (that is, values that indicate the position of the corresponding entry in dk_entries). This array acts as the hash table. When a key is hashed it leads to one of the indices stored in dk_indices and the corresponding entry is fetched by indexing dk_entries. Since only indices are kept, the type of this array depends on the overall size of the dictionary (ranging from type int8_t(1 byte) to int32_t/int64_t (4/8 bytes) on 32/64 bit builds)

In the previous implementation, a sparse array of type PyDictKeyEntry and size dk_size had to be allocated; unfortunately, it also resulted in a lot of empty space since that array was not allowed to be more than 2/3 * dk_size full for performance reasons. (and the empty space still had PyDictKeyEntry size!).

This is not the case now since only the required entries are stored (those that have been inserted) and a sparse array of type intX_t (X depending on dict size) 2/3 * dk_sizes full is kept. The empty space changed from type PyDictKeyEntry to intX_t.

So, obviously, creating a sparse array of type PyDictKeyEntry is much more memory demanding than a sparse array for storing ints.

You can see the full conversation on Python-Dev regarding this feature if interested, it is a good read.

In the original proposal made by Raymond Hettinger, a visualization of the data structures used can be seen which captures the gist of the idea.

For example, the dictionary:

d = {'timmy': 'red', 'barry': 'green', 'guido': 'blue'}

is currently stored as:

entries = [['--', '--', '--'],
           [-8522787127447073495, 'barry', 'green'],
           ['--', '--', '--'],
           ['--', '--', '--'],
           ['--', '--', '--'],
           [-9092791511155847987, 'timmy', 'red'],
           ['--', '--', '--'],
           [-6480567542315338377, 'guido', 'blue']]

Instead, the data should be organized as follows:

indices =  [None, 1, None, None, None, 0, None, 2]
entries =  [[-9092791511155847987, 'timmy', 'red'],
            [-8522787127447073495, 'barry', 'green'],
            [-6480567542315338377, 'guido', 'blue']]

As you can visually now see, in the original proposal, a lot of space is essentially empty to reduce collisions and make look-ups faster. With the new approach, you reduce the memory required by moving the sparseness where it's really required, in the indices.

[1]: I say "insertion ordered" and not "ordered" since, with the existence of OrderedDict, "ordered" suggests further behavior that the dict object doesn't provide. OrderedDicts are reversible, provide order sensitive methods and, mainly, provide an order-sensive equality tests (==, !=). dicts currently don't offer any of those behaviors/methods.

[2]: The new dictionary implementations performs better memory wise by being designed more compactly; that's the main benefit here. Speed wise, the difference isn't so drastic, there's places where the new dict might introduce slight regressions (key-lookups, for example) while in others (iteration and resizing come to mind) a performance boost should be present.

Overall, the performance of the dictionary, especially in real-life situations, improves due to the compactness introduced.

improve this answer
  • 7
    So, what happens when an item is removed? is the entries list resized? or is a blank space kept? or is it compressed from time to time? – njzk2 Oct 11 '16 at 19:19
  • 10
    @njzk2 When an item is removed, the corresponding index is replaced by DKIX_DUMMY with a value of -2and the entry in the entry array replaced by NULL, when inserting is performed the new values are appended to the entries array, Haven't been able to discern yet, but pretty sure when the indices fills up beyond the 2/3 threshold resizing is performed. This can lead to shrinking instead of growing if many DUMMY entries exist. – Jim Fasarakis Hilliard Oct 11 '16 at 20:03 
  • 3
    @Chris_Rands Nope, the only actual regression I've seen is on the tracker in a message by Victor. Other than that microbenchmark, I've seen no other issue/message indicating a serious speed difference in real-life work loads. There's places where the new dict might introduce slight regressions (key-lookups, for example) while in others (iteration and resizing come to mind) a performance boost would be present. – Jim Fasarakis HilliardMar 14 '17 at 13:26
  • 2
    Correction on the resizing part: Dictionaries don't resize when you delete items, they re-calculate when you re-insert. So, if a dict is created with d = {i:i for i in range(100)} and you .pop all items w/o inserting, the size won't change. When you add to it again, d[1] = 1, the appropriate size is calculated and the dict resizes. – Jim Fasarakis Hilliard Aug 2 '17 at 19:47 
  • 4
    @Chris_Rands I'm pretty sure it is staying. The thing is, and the reason why I changed my answer to remove blanket statements about 'dict being ordered', dicts aren't ordered in the sense OrderedDicts are. The notable issue is equality. dicts have order insensitive ==, OrderedDicts have order sensitive ones. Dumping OrderedDicts and changing dicts to now have order sensitive comparisons could lead to a lot of breakage in old code. I'm guessing the only thing that might change about OrderedDicts is its implementation. – Jim Fasarakis Hilliard Apr 10 '18 at 16:57