Currently FastAPI calls inspect.signature, inspect.isgeneratorfunction, inspect.isasyncgenfunction, inspect.iscoroutinefunction, etc for each callable found in dependency tree, and this is done for each received request.

This can be avoided by introducing a cache for methods result. As callable objects are usually created once and never change after that, this cache can be bound to endpoint or even whole application. Using WeakKeyDictionary helps to avoid memory leaks then callables are created in runtime (e.g. by using functools.partial).

This cache is different from dependency cache - dependency cache stores values of resolved dependencies (non reusable between requests), then callable info cache stores signatures and other information about code (reusable between requests).

Before - app (fastapi/routing.py) takes 68.27% of all method calls:

After - app (fastapi/routing.py) takes 54.52% of all method calls:

The code is ugly as walking on dependency tree requires passing cache deep into dependency resolvers. Also lots of intermediate functions have default argument values, which are not used in most cases, but I have to follow this pattern.