Using @functools.lru_cache with dictionary arguments

user_760767 0
Viewed 18943

I have a method that takes (among others) a dictionary as an argument. The method is parsing strings and the dictionary provides replacements for some substrings, so it doesn't have to be mutable.

This function is called quite often, and on redundant elements so I figured that caching it would improve its efficiency.

But, as you may have guessed, since dict is mutable and thus not hashable, @functools.lru_cache can't decorate my function. So how can I overcome this?

Bonus point if it needs only standard library classes and methods. Ideally if it exists some kind of frozendict in standard library that I haven't seen it would make my day.

PS: namedtuple only in last resort, since it would need a big syntax shift.

hashable python-3.x dictionary python
8 Answers

Instead of using a custom hashable dictionary, use this and avoid reinventing the wheel! It's a frozen dictionary that's all hashable.

https://pypi.org/project/frozendict/

Code:

from frozendict import frozendict

def freezeargs(func):
    """Transform mutable dictionnary
    Into immutable
    Useful to be compatible with cache
    """

    @functools.wraps(func)
    def wrapped(*args, **kwargs):
        args = tuple([frozendict(arg) if isinstance(arg, dict) else arg for arg in args])
        kwargs = {k: frozendict(v) if isinstance(v, dict) else v for k, v in kwargs.items()}
        return func(*args, **kwargs)
    return wrapped

and then

@freezeargs
@lru_cache
def func(...):
    pass

Code taken from @fast_cen 's answer

Note: this does not work on recursive datastructures; for example, you might have an argument that's a list, which is unhashable. You are invited to make the wrapping recursive, such that it goes deep into the data structure and makes every dict frozen and every list tuple.

(I know that OP nolonger wants a solution, but I came here looking for the same solution, so leaving this for future generations)

user_58107470

Based on @Cedar answer, adding recursion for deep freeze as suggested:

def deep_freeze(thing):
    from collections.abc import Collection, Mapping, Hashable
    from frozendict import frozendict
    if thing is None or isinstance(thing, str):
        return thing
    elif isinstance(thing, Mapping):
        return frozendict({k: deep_freeze(v) for k, v in thing.items()})
    elif isinstance(thing, Collection):
        return tuple(deep_freeze(i) for i in thing)
    elif not isinstance(thing, Hashable):
        raise TypeError(f"unfreezable type: '{type(thing)}'")
    else:
        return thing


def deep_freeze_args(func):
    import functools

    @functools.wraps(func)
    def wrapped(*args, **kwargs):
        return func(*deep_freeze(args), **deep_freeze(kwargs))
    return wrapped
user_10963700

After deciding to drop lru cache for our use case for now, we still came up with a solution. This decorator uses json to serialise and deserialise the args/kwargs sent to the cache. Works with any number of args. Use it as a decorator on a function instead of @lru_cache. max size is set to 1024.

def hashable_lru(func):
    cache = lru_cache(maxsize=1024)

    def deserialise(value):
        try:
            return json.loads(value)
        except Exception:
            return value

    def func_with_serialized_params(*args, **kwargs):
        _args = tuple([deserialise(arg) for arg in args])
        _kwargs = {k: deserialise(v) for k, v in kwargs.items()}
        return func(*_args, **_kwargs)

    cached_function = cache(func_with_serialized_params)

    @wraps(func)
    def lru_decorator(*args, **kwargs):
        _args = tuple([json.dumps(arg, sort_keys=True) if type(arg) in (list, dict) else arg for arg in args])
        _kwargs = {k: json.dumps(v, sort_keys=True) if type(v) in (list, dict) else v for k, v in kwargs.items()}
        return cached_function(*_args, **_kwargs)
    lru_decorator.cache_info = cached_function.cache_info
    lru_decorator.cache_clear = cached_function.cache_clear
    return lru_decorator
user_1592410

The solution might be much simpler. The lru_cache uses parameters as the identifier for caching, so in the case of the dictionary, lru_cache doesn't know how to interpret it. You can serialize dictionary parameter to string and unserialize in the function to the dictionary back. Works like a charm.

the function:

    @lru_cache(1024)
    def data_check(serialized_dictionary):
        my_dictionary = json.loads(serialized_dictionary)
        print(my_dictionary)

the call:

    data_check(json.dumps(initial_dictionary))
user_36733280
Related