django-cache-memoize 0.2.0

args_rewrite

Internally the decorator rewrites every argument and keyword argument to the function it wraps into a concatenated string. The first thing you might want to do is help the decorator rewrite the arguments to something more suitable as a cache key string. For example, suppose you have instances of a class whose __str__ method doesn’t return a unique value. For example:

class Record(models.Model):
    name = models.CharField(max_length=100)
    lastname = models.CharField(max_length=100)
    friends = models.ManyToManyField(SomeOtherModel)

    def __str__(self):
        return self.name

# Example use:
>>> record = Record.objects.create(name='Peter', lastname='Bengtsson')
>>> print(record)
Peter
>>> record2 = Record.objects.create(name='Peter', lastname='Different')
>>> print(record2)
Peter

This is a contrived example, but basically you know that the str() conversion of certain arguments isn’t safe. Then you can pass in a callable called args_rewrite. It gets the same positional and keyword arguments as the function you’re decorating. Here’s an example implementation:

from cache_memoize import cache_memoize

def count_friends_args_rewrite(record):
    # The 'id' is always unique. Use that instead of the default __str__
    return record.id

@cache_memoize(100, args_rewrite=count_friends_args_rewrite)
def count_friends(record):
    # Assume this is an expensive function that can be memoize cached.
    return record.friends.all().count()

prefix

By default the prefix becomes the name of the function. Consider:

from cache_memoize import cache_memoize

@cache_memoize(10, prefix='randomness')
def function1():
    return random.random()

@cache_memoize(10, prefix='randomness')
def function2():  # different name, same arguments, same functionality
    return random.random()

# Example use
>>> function1()
0.39403406043780986
>>> function1()
0.39403406043780986
>>> # ^ repeated of course
>>> function2()
0.39403406043780986
>>> # ^ because the prefix was forcibly the same, the cache key is the same

hit_callable

If set, a function that gets called with the original argument and keyword arguments if the cache was able to find and return a cache hit. For example, suppose you want to tell your statsd server every time there’s a cache hit.

from cache_memoize import cache_memoize

def _cache_hit(user, **kwargs):
    statsdthing.incr(f'cachehit:{user.id}', 1)

@cache_memoize(10, hit_callable=_cache_hit)
def calculate_tax(user, tax=0.1):
    return ...

miss_callable

Exact same functionality as hit_callable except the obvious difference that it gets called if it was not a cache hit.

store_result

This is useful if you have a function you want to make sure only gets called once per timeout expiration but you don’t actually care that much about what the function return value was. Perhaps because you know that the function returns something that would quickly fill up your memcached or perhaps you know it returns something that can’t be pickled. Then you can set store_result to False. This is equivalent to your function returning True.

from cache_memoize import cache_memoize

@cache_memoize(1000, store_result=False)
def send_tax_returns(user):
    # something something time consuming
    ...
    return some_none_pickleable_thing

def myview(request):
    # View this view as much as you like the 'send_tax_returns' function
    # won't be called more than once every 1000 seconds.
    send_tax_returns(request.user)

cache_exceptions

This is useful if you have a function that can raise an exception as valid result. If the cached function raises any of specified exceptions is the exception cached and raised as normal. Subsequent cached calls will immediately re-raise the exception and the function will not be executed. cache_exceptions accepts an Exception or a tuple of Exceptions.

This option allows you to cache said exceptions like any other result. Only exceptions raised from the list of classes provided as cache_exceptions are cached, all others are propagated immediately.

>>> from cache_memoize import cache_memoize

>>> class InvalidParameter(Exception):
...     pass

>>> @cache_memoize(1000, cache_exceptions=(InvalidParameter, ))
... def run_calculations(parameter):
...     # something something time consuming
...     raise InvalidParameter

>>> run_calculations(1)
Traceback (most recent call last):
...
InvalidParameter

# run_calculations will now raise InvalidParameter immediately
# without running the expensive calculation
>>> run_calculations(1)
Traceback (most recent call last):
...
InvalidParameter

cache_alias

The cache_alias argument allows you to use a cache other than the default.

# Given settings like:
# CACHES = {
#     'default': {...},
#     'other': {...},
# }

@cache_memoize(1000, cache_alias='other')
def myfunc(start, end):
    return random.random()

Cache invalidation

When you want to “undo” some caching done, you simply call the function again with the same arguments except you add .invalidate to the function.

from cache_memoize import cache_memoize

@cache_memoize(10)
def expensive_function(start, end):
    return random.randint(start, end)

>>> expensive_function(1, 100)
65
>>> expensive_function(1, 100)
65
>>> expensive_function(100, 200)
121
>>> exensive_function.invalidate(1, 100)
>>> expensive_function(1, 100)
89
>>> expensive_function(100, 200)
121

An “alias” of doing the same thing is to pass a keyword argument called _refresh=True. Like this:

# Continuing from the code block above
>>> expensive_function(100, 200)
121
>>> expensive_function(100, 200, _refresh=True)
177
>>> expensive_function(100, 200)
177

There is no way to clear more than one cache key. In the above example, you had to know the “original arguments” when you wanted to invalidate the cache. There is no method “search” for all cache keys that match a certain pattern.

History

Mozilla Symbol Server is written in Django. It’s a web service that sits between C++ debuggers and AWS S3. It shuffles symbol files in and out of AWS S3. Symbol files are for C++ (and other compiled languages) what sourcemaps are for JavaScript.

This service gets a LOT of traffic. The download traffic (proxying requests for symbols in S3) gets about ~40 requests per second. Due to the nature of the application most of these GETs result in a 404 Not Found but instead of asking AWS S3 for every single file, these lookups are cached in a highly configured Redis configuration. This Redis cache is also connected to the part of the code that uploads new files.

New uploads are arriving as zip file bundles of files, from Mozilla’s build systems, at a rate of about 600MB every minute, each containing on average about 100 files each. When a new upload comes in we need to quickly be able find out if it exists in S3 and this gets cached since often the same files are repeated in different uploads. But when a file does get uploaded into S3 we need to quickly and confidently invalidate any local caches. That way you get to keep a really aggressive cache without any stale periods.

This is the use case django-cache-memoize was built for and tested in. It was originally written for Python 3.6 in Django 1.11 but when extracted, made compatible with Python 2.7 and as far back as Django 1.8.

django-cache-memoize is also used in SongSear.ch to cache short queries in the autocomplete search input. All autocomplete is done by Elasticsearch, which is amazingly fast, but not as fast as memcached.

“Competition”

There is already django-memoize by Thomas Vavrys. It too is available as a memoization decorator you use in Django. And it uses the default cache framework as a storage. It used inspect on the decorated function to build a cache key.

In benchmarks running both django-memoize and django-cache-memoize I found django-cache-memoize to be ~4 times faster on average.

Another key difference is that django-cache-memoize uses str() and django-memoize uses repr() which in certain cases of mutable objects (e.g. class instances) as arguments the caching will not work. For example, this does not work in django-memoize:

from memoize import memoize

@memoize(60)
def count_user_groups(user):
    return user.groups.all().count()

def myview(request):
    # this will never be memoized
    print(count_user_groups(request.user))

However, this works…

from cache_memoize import cache_memoize

@cache_memoize(60)
def count_user_groups(user):
    return user.groups.all().count()

def myview(request):
    # this *will* work as expected
    print(count_user_groups(request.user))

Code style is all black

All code has to be formatted with Black and the best tool for checking this is therapist since it can help you run all, help you fix things, and help you make sure linting is passing before you git commit. This project also uses flake8 to check other things Black can’t check.

To check linting with tox use:

tox -e lint-py36

To install the therapist pre-commit hook simply run:

therapist install

When you run therapist run it will only check the files you’ve touched. To run it for all files use:

therapist run --use-tracked-files

And to fix all/any issues run:

therapist run --use-tracked-files --fix