smemo 0.3.2

Explicit session memoization

smemo: Explicit session memoization

Memoization of function return values is a very common optimization technique. Even the standard Python 3 library provides multiple facilities for it, e.g., functools.cached_property and functools.lru_cache. Needless to say, many user-contributed facilities are also available.

Most of them focuses on ensuring that the cache is invisible to the user. When you call a function decorated by functools.lru_cache, you don't really need to know that the function value is cached. That is usually a big merit, but at times this is a big drawback.

The problem is on invalidation. Most libraries spent a lot of efforts to keep track of them magically so that we don't need to worry about them. But at times, all these work are not necessary or outright counterproductive. The run of your program might be divided into parts, or sessions, where each session is benefited by having return values of functions, usually many of them, saved rather than recomputed. But after a session, the saved values of the function is of little value or even may be invalid, so there is not much to be gained by keeping them in memory.

In such cases, most other caching tools either do not help at all, or requires you to carefully find the functions being used by the session and invalidate their caches one by one.

This module uses a more explicit approach. A "session" object is used, providing the cache required by the memoization process. The availability of this object makes a few operations, tricky in other libraries, to become trivial. They include:

• Invalidation of all caches of a session.
• One single function uses different caches for different calls.
• Fine-grain control of whether values of a function should skip caching.
• Injection of values to a particular function for a session.

In fact, the implementation of the module is so simple that the core of it can be done in just a couple of hours.

Usage

Functions requiring caching should be written like this:

@smemo.cached
def efib(session: smemo.BaseSession, n: int,
         a0: float, a1: float) -> float:
    "Return an extended Fibonacci number where efib[0] = a0, efib[1] = a1"
    if n == 0:
        return a0
    if n == 1:
        return a1
    return efib(session, n - 1, a0, a1) + efib(session, n - 2, a0, a1)

Note that it explicitly includes the session object as the first argument. All functions decorated by @smemo.cached does the same.

It can then be called like this:

session = smemo.Session()
print(efib(session, 5, 1, 3))

After the call, the session contains entries like {((2, 1, 3), ()): (4, None)}, the key contains a hashable representation of the positional and keyword arguments, while the values are the return value or exception raised. The value can be observed without triggering a call to the inner function, by:

print(session.get_cache(efib, 5, 1, 3))

The returned value of the function decorated by @smemo.cached goes through copy.deepcopy, so that the caller can freely manipulate the returned object without affecting the cache. If that is undesirable (usually because it slows things down or you really want the actual object returned), you would replace @smemo.cached by @smemo.rcached.

For many purposes these are good enough. But at times some of the values (e.g., 3 and 1 above) are really constants for a session, and you can save a lot of space by not including them in the keys of the cache. In such cases you can rewrite your function as follows:

@smemo.cached
def efib(session: smemo.BaseSession, n: int) -> float:
    "Return an extended Fibonacci number where efib[0] = a0, efib[1] = a1"
    if n == 0:
        return session.getval('a0')  # Same as smemo.getter(session, 'a0')
    if n == 1:
        return session.getval('a1')
    return efib(session, n - 1) + efib(session, n - 2)

You call your function like this instead:

session = smemo.Session()
session.putval(1, 'a0')
session.putval(3, 'a1')
print(efib(session, 5))

This is a trivial function to implement. It just composes other session facilities with a "getter" function, the latter simply raises an exception. On the other hand, this is a very useful mechanism, because it enables you to use memoization where previously you might not even think you want to. For example, you can read a whole config file and put it to your session. Now your functions can freely use the configuration values, and you don't need to worry about having cache which has huge entries, and at the same time you don't need to worry about cross-talk when your function is called concurrently with different sessions.

Nested session

At times you may have some of your data to live for a shorter lifetime than the rest. In such case, you might find it beneficial to have a hierarchy of sessions. The Session class __init__ takes a "parent" BaseSession argument. When cache is fetched and no entry is found, this parent will be consulted. Whenever cache entry is set it is local. This way you can populate some entry in a global session, and create a "sub-session" which will hold any additional cache entries.

Cache control

The big differentiation between smemo and other solutions is cache control. We do not do implicit cached control like LRU at all, all cached values are kept there until explicitly invalidated. In fact, we don't even store those information (e.g., when is a value last needed), so the overheads in our module is quite small.

Having a session object, explicit cache control can be done trivially. For example, you can clear the cache either for one call, for one function or for all functions, by the following respectively:

session.invalidate(efib, 7)
session.invalidate_all(efib)
session.invalidate_all()

For the first call, the call needs to match the calling arguments, but mypy will not be able to catch it if there is a type error. An alternative method can be used instead, which does trigger mypy to catch a type error:

efib(session.inv, 7)

Because it is just another call to efib, it has to match the type correctly to pass mypy.

For the last session.invalidate call, it actually doesn't invalidate all function values cached. If a function is marked as "persistent", the values cached for it can only be invalidated more explicitly. This is done by using @smemo.gcached(persistent=True) instead of @smemo.cached, or (more commonly) using @smemo.gcached(ref=True, persistent=True) instead of @smemo.rcached. Under the hood, getval and putval is just cache of a function, where the "persistent" mechanism is used to avoid getting values removed.

You can force a value into the cache, like one of these:

session.cache(efib, 5.0, 7)
session.cache_exc(efib, RuntimeError('My error'), 7)

In this case, it causes efib(7, session) to return 5.0 and raises "My error" respectively. This is very useful during unit tests, where at times you don't want your functions to actually be called. Again, we have alternative methods which allows mypy to catch type errors:

efib(session.setcache(5.0), 7)
efib(session.setcache(exception=RuntimeError('My error')), 7)

So, if you want to avoid getval and setval (which mypy cannot check for type errors), you can create functions returning the correct type and mark them as persistent instead:

@smemo.gcached(ref=True, persistent=True)
def mydict(session: smemo.BaseSession, a: str) -> typing.Dict[str, float]:
    raise RuntimeError(f'No value set for {a}')

Then you can put and get values as follows:

mydict(session.setcache({'pi': 3.1415936, 'e': 2.7182818}), 'const')
print(mydict(session, 'const')['pi'])

Note that this mechanism, when combined with nested sessions, can be used as a poor man dependency injection mechanism. All you need to do is to pass a session when you invoke any top-level function, and obtain all dependencies from the session. The main program would populate the main session and perhaps create sub-sessions as needed, before calling the top-level functions. Your unit tests can use mock objects instead of actual objects as desired.

Finally, you can call a function without caching it:

efib(session.callonly, 7)

The skipping of caching does not extend to the calls made by efib. If you want that, you can do the following:

with session.nocache:
    print(efib(session, 7))

Caching would be disabled for all the duration of the above efib call.

BaseSession interface

All the functionality of the memoization system is provided by the BaseSession interface, which the Session class is written against. If you want your own facilities, you might want to provide your own implementation. It is a rather simple interface:

• get_cache(func, *args, **kwargs)

  • When the user calls the defined function, the function object and the positional/keyword arguments are passed to this method to find a cache entry, which should be in the form (ret_val, exception), where the ret_val is used if exception is None.

• do_call(func, actual, args, kwargs)

  • When cache entry is not found or is skipped, this function will be called to obtain a value to return to the function caller.

• cache_exc(func, _exc, *args, **kwargs)

  • When the do_call method generates an exception, this function is called to ask the session object to cache the resulting exception. Note that we usually define the method to have the exception argument named like _exc, because the name could clash with names in kwargs.

• cache_val(func, _val, *args, **kwargs)

  • When the do_call method returns a value, this function is called to ask the session object to cache the result. Note that we usually define the method to have the exception argument named like _val, because the name could clash with names in kwargs.

Cache control is actually done by creative use of the interface. For example, session.inv is a BaseSession object which returns None upon get_cache or do_call, but invalidate cache entry upon cache_val. And session.setcache returns an object which is just the same as session.inv but instead of invalidating the cache, it forces a return value or exception to the cache.