Python Enhancement Proposals

PEP 455 – Adding a key-transforming dictionary to collections

Adding a key-transforming dictionary to collections
Antoine Pitrou <solipsis at>
Raymond Hettinger
Standards Track



This PEP proposes a new data structure for the collections module, called “TransformDict” in this PEP. This structure is a mutable mapping which transforms the key using a given function when doing a lookup, but retains the original key when reading.


See the rationale at and for an earlier partial review, see .


Numerous specialized versions of this pattern exist. The most common is a case-insensitive case-preserving dict, i.e. a dict-like container which matches keys in a case-insensitive fashion but retains the original casing. It is a very common need in network programming, as many protocols feature some arrays of “key / value” properties in their messages, where the keys are textual strings whose case is specified to be ignored on receipt but by either specification or custom is to be preserved or non-trivially canonicalized when retransmitted.

Another common request is an identity dict, where keys are matched according to their respective id()s instead of normal matching.

Both are instances of a more general pattern, where a given transformation function is applied to keys when looking them up: that function being str.lower or str.casefold in the former example and the built-in id function in the latter.

(It could be said that the pattern projects keys from the user-visible set onto the internal lookup set.)


TransformDict is a MutableMapping implementation: it faithfully implements the well-known API of mutable mappings, like dict itself and other dict-like classes in the standard library. Therefore, this PEP won’t rehash the semantics of most TransformDict methods.

The transformation function needn’t be bijective, it can be strictly surjective as in the case-insensitive example (in other words, different keys can lookup the same value):

>>> d = TransformDict(str.casefold)
>>> d['SomeKey'] = 5
>>> d['somekey']
>>> d['SOMEKEY']

TransformDict retains the first key used when creating an entry:

>>> d = TransformDict(str.casefold)
>>> d['SomeKey'] = 1
>>> d['somekey'] = 2
>>> list(d.items())
[('SomeKey', 2)]

The original keys needn’t be hashable, as long as the transformation function returns a hashable one:

>>> d = TransformDict(id)
>>> l = [None]
>>> d[l] = 5
>>> l in d


As shown in the examples above, creating a TransformDict requires passing the key transformation function as the first argument (much like creating a defaultdict requires passing the factory function as first argument).

The constructor also takes other optional arguments which can be used to initialize the TransformDict with certain key-value pairs. Those optional arguments are the same as in the dict and defaultdict constructors:

>>> d = TransformDict(str.casefold, [('Foo', 1)], Bar=2)
>>> sorted(d.items())
[('Bar', 2), ('Foo', 1)]

Getting the original key

TransformDict also features a lookup method returning the stored key together with the corresponding value:

>>> d = TransformDict(str.casefold, {'Foo': 1})
>>> d.getitem('FOO')
('Foo', 1)
>>> d.getitem('bar')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
KeyError: 'bar'

The method name getitem() follows the standard popitem() method on mutable mappings.

Getting the transformation function

TransformDict has a simple read-only property transform_func which gives back the transformation function.

Alternative proposals and questions

Retaining the last original key

Most python-dev respondents found retaining the first user-supplied key more intuitive than retaining the last. Also, it matches the dict object’s own behaviour when using different but equal keys:

>>> d = {}
>>> d[1] = 'hello'
>>> d[1.0] = 'world'
>>> d
{1: 'world'}

Furthermore, explicitly retaining the last key in a first-key-retaining scheme is still possible using the following approach:

d.pop(key, None)
d[key] = value

while the converse (retaining the first key in a last-key-retaining scheme) doesn’t look possible without rewriting part of the container’s code.

Using an encoder / decoder pair

Using a function pair isn’t necessary, since the original key is retained by the container. Moreover, an encoder / decoder pair would require the transformation to be bijective, which prevents important use cases like case-insensitive matching.

Providing a transformation function for values

Dictionary values are not used for lookup, their semantics are totally irrelevant to the container’s operation. Therefore, there is no point in having both an “original” and a “transformed” value: the transformed value wouldn’t be used for anything.

Providing a specialized container, not generic

It was asked why we would provide the generic TransformDict construct rather than a specialized case-insensitive dict variant. The answer is that it’s nearly as cheap (code-wise and performance-wise) to provide the generic construct, and it can fill more use cases.

Even case-insensitive dicts can actually elicit different transformation functions: str.lower, str.casefold or in some cases bytes.lower when working with text encoded in an ASCII-compatible encoding.

Other constructor patterns

Two other constructor patterns were proposed by Serhiy Storchaka:

  • A type factory scheme:
    d = TransformDict(str.casefold)(Foo=1)
  • A subclassing scheme:
    class CaseInsensitiveDict(TransformDict):
        __transform__ = str.casefold
    d = CaseInsensitiveDict(Foo=1)

While both approaches can be defended, they don’t follow established practices in the standard library, and therefore were rejected.


A patch for the collections module is tracked on the bug tracker at

Existing work

Case-insensitive dicts are a popular request:

Identity dicts have been requested too:

Several modules in the standard library use identity lookups for object memoization, for example pickle, json, copy, cProfile, doctest and _threading_local.

Other languages

C# / .Net

.Net has a generic Dictionary class where you can specify a custom IEqualityComparer:

Using it is the recommended way to write case-insensitive dictionaries:


Java has a specialized CaseInsensitiveMap:

It also has a separate IdentityHashMap:


The C++ Standard Template Library features an unordered_map with customizable hash and equality functions:


Last modified: 2017-11-11 19:28:55 GMT