bidict strives to be as performant as possible while being faithful to its purpose. The need for speed is balanced with the responsibility to protect users from shooting themselves in the foot.

In general, accomplishing some task using bidict should have about the same performance as keeping two inverse dicts in sync manually. The test suite includes benchmarks for common workloads to catch any performance regressions.

If you spot a case where bidict’s performance could be improved, please don’t hesitate to file an issue or submit a pull request.

bidict Avoids Reference Cycles

A careful reader might notice the following…

>>> fwd = bidict(one=1)
>>> inv = fwd.inverse
>>> inv.inverse is fwd

…and worry that a bidict and its inverse create a reference cycle. If this were true, in CPython this would mean that the memory for a bidict could not be immediately reclaimed when you retained no more references to it, but rather would have to wait for the next gargage collection to kick in before it could be reclaimed.

However, bidicts use a weakref.ref to store the inverse reference in one direction, avoiding the strong reference cycle. As a result, when you no longer retain any references to a bidict you create, you can be sure that its refcount in CPython drops to zero, and that its memory will therefore be reclaimed immediately.


In PyPy this is not an issue, as PyPy doesn’t use reference counts. The memory for unreferenced objects in PyPy is only reclaimed when GC kicks in, which is unpredictable.


  • It’s intentional that the term “inverse” is used rather than “reverse”.

    Consider a collection of (k, v) pairs. Taking the reverse of the collection can only be done if it is ordered, and (as you’d expect) reverses the order of the pairs in the collection. But each original (k, v) pair remains in the resulting collection.

    By contrast, taking the inverse of such a collection neither requires the collection to be ordered nor guarantees any ordering in the result, but rather just replaces every (k, v) pair with the inverse pair (v, k).

  • “keys” and “values” could perhaps more properly be called “primary keys” and “secondary keys” (as in a database), or even “forward keys” and “inverse keys”, respectively. bidict sticks with the terms “keys” and “values” for the sake of familiarity and to avoid potential confusion, but technically values are also keys themselves.

    Concretely, this allows bidicts to return a set-like (dict_keys) object for values(), rather than a non-set-like dict_values object.

Missing bidicts in the Standard Library

The Python standard library actually contains some examples where bidicts could be used for fun and profit (depending on your ideas of fun and profit):

  • The logging module contains a private _levelToName dict which maps integer levels like 10 to their string names like DEBUG. If I had a nickel for every time I wanted that exposed in a bidirectional map (and as a public attribute, no less), I bet I could afford some better turns of phrase.

  • The dis module maintains a mapping from opnames to opcodes dis.opmap and a separate list of opnames indexed by opcode dis.opnames. These could be combined into a single bidict.

  • Python 3’s html.entities module maintains separate html.entities.name2codepoint and html.entities.codepoint2name dicts. These could be combined into a single bidict.


Non-Atomic Mutation

As with built-in dicts, mutating operations on a bidict are not atomic. If you need to mutate the same bidict from different threads, use a synchronization primitive to coordinate access. 1


See also: [2], [3]

Equivalent but distinct Hashables

Consider the following:

>>> d = {1: int, 1.0: float}

How many items do you expect d to contain? The actual result might surprise you:

>>> len(d)

And similarly,

>>> dict([(1, int), (1.0, float), (1+0j, complex), (True, bool)])
{1: <... 'bool'>}
>>> 1.0 in {True}

(Note that 1 == 1.0 == 1+0j == True.)

This illustrates that a mapping cannot contain two items with equivalent but distinct keys (and likewise a set cannot contain two equivalent but distinct elements). If an object that is being looked up in a set or mapping is equal to a contained object, the contained object will be found, even if it is distinct.

With a bidict, since values function as keys in the inverse mapping, this behavior occurs in the inverse direction too, and means that a bidict can end up with a different but equivalent key from the corresponding value in its own inverse:

>>> b = bidict({'false': 0})
>>> b.forceput('FALSE', False)
>>> b
bidict({'FALSE': False})
>>> b.inverse
bidict({0: 'FALSE'})

nan as a Key

In CPython, nan is especially tricky when used as a dictionary key:

>>> d = {float('nan'): 'nan'}
>>> d
{nan: 'nan'}
>>> d[float('nan')]  
Traceback (most recent call last):
KeyError: nan
>>> d[float('nan')] = 'not overwritten'
>>> d  
{nan: 'nan', nan: 'not overwritten'}

In other Python implementations such as PyPy, nan behaves just like any other dictionary key. But in CPython, beware of this unexpected behavior, which applies to bidicts too. bidict contains no special-case logic for dealing with nan as a key, so the behavior will match dict’s wherever bidict is running.

See e.g. these docs for more info (search the page for “nan”).

For more in this vein, check out Learning from bidict.