Cacheops

A slick app that supports automatic or manual queryset caching and automatic granular event-driven invalidation.

It uses redis as backend for ORM cache and redis or filesystem for simple time-invalidated one.

And there is more to it:

decorators to cache any user function or view as a queryset or by time
extensions for django and jinja2 templates
transparent transaction support
dog-pile prevention mechanism
a couple of hacks to make django faster

Requirements

Python 3.5+, Django 2.1+ and Redis 4.0+.

Installation

Using pip:

$ pip install django-cacheops

# Or from github directly
$ pip install git+https://github.com/Suor/django-cacheops.git@master

Setup

Add cacheops to your INSTALLED_APPS.

Setup redis connection and enable caching for desired models:

CACHEOPS_REDIS = {
    'host': 'localhost', # redis-server is on same machine
    'port': 6379,        # default redis port
    'db': 1,             # SELECT non-default redis database
                         # using separate redis db or redis instance
                         # is highly recommended

    'socket_timeout': 3,   # connection timeout in seconds, optional
    'password': '...',     # optional
    'unix_socket_path': '' # replaces host and port
}

# Alternatively the redis connection can be defined using a URL:
CACHEOPS_REDIS = "redis://localhost:6379/1"
# or
CACHEOPS_REDIS = "unix://path/to/socket?db=1"
# or with password (note a colon)
CACHEOPS_REDIS = "redis://:password@localhost:6379/1"

# If you want to use sentinel, specify this variable
CACHEOPS_SENTINEL = {
    'locations': [('localhost', 26379)], # sentinel locations, required
    'service_name': 'mymaster',          # sentinel service name, required
    'socket_timeout': 0.1,               # connection timeout in seconds, optional
    'db': 0                              # redis database, default: 0
    ...                                  # everything else is passed to Sentinel()
}

# To use your own redis client class,
# should be compatible or subclass cacheops.redis.CacheopsRedis
CACHEOPS_CLIENT_CLASS = 'your.redis.ClientClass'

CACHEOPS = {
    # Automatically cache any User.objects.get() calls for 15 minutes
    # This also includes .first() and .last() calls,
    # as well as request.user or post.author access,
    # where Post.author is a foreign key to auth.User
    'auth.user': {'ops': 'get', 'timeout': 60*15},

    # Automatically cache all gets and queryset fetches
    # to other django.contrib.auth models for an hour
    'auth.*': {'ops': {'fetch', 'get'}, 'timeout': 60*60},

    # Cache all queries to Permission
    # 'all' is an alias for {'get', 'fetch', 'count', 'aggregate', 'exists'}
    'auth.permission': {'ops': 'all', 'timeout': 60*60},

    # Enable manual caching on all other models with default timeout of an hour
    # Use Post.objects.cache().get(...)
    #  or Tags.objects.filter(...).order_by(...).cache()
    # to cache particular ORM request.
    # Invalidation is still automatic
    '*.*': {'ops': (), 'timeout': 60*60},

    # And since ops is empty by default you can rewrite last line as:
    '*.*': {'timeout': 60*60},

    # NOTE: binding signals has its overhead, like preventing fast mass deletes,
    #       you might want to only register whatever you cache and dependencies.

    # Finally you can explicitely forbid even manual caching with:
    'some_app.*': None,
}

You can configure default profile setting with CACHEOPS_DEFAULTS. This way you can rewrite the config above:

CACHEOPS_DEFAULTS = {
    'timeout': 60*60
}
CACHEOPS = {
    'auth.user': {'ops': 'get', 'timeout': 60*15},
    'auth.*': {'ops': ('fetch', 'get')},
    'auth.permission': {'ops': 'all'},
    '*.*': {},
}

Using '*.*' with non-empty ops is not recommended since it will easily cache something you don't intent to or even know about like migrations tables. The better approach will be restricting by app with 'app_name.*'.

Besides ops and timeout options you can also use:

local_get: True: To cache simple gets for this model in process local memory. This is very fast, but is not invalidated in any way until process is restarted. Still could be useful for extremely rarely changed things.
cache_on_save=True | 'field_name': To write an instance to cache upon save. Cached instance will be retrieved on .get(field_name=...) request. Setting to True causes caching by primary key.

Additionally, you can tell cacheops to degrade gracefully on redis fail with:

CACHEOPS_DEGRADE_ON_FAILURE = True

There is also a possibility to make all cacheops methods and decorators no-op, e.g. for testing:

from django.test import override_settings

@override_settings(CACHEOPS_ENABLED=False)
def test_something():
    # ...
    assert cond

Usage

Automatic caching

It's automatic you just need to set it up.

Manual caching

You can force any queryset to use cache by calling its .cache() method:

Article.objects.filter(tag=2).cache()

Here you can specify which ops should be cached for the queryset, for example, this code:

qs = Article.objects.filter(tag=2).cache(ops=['count'])
paginator = Paginator(objects, ipp)
articles = list(pager.page(page_num)) # hits database

will cache count call in Paginator but not later articles fetch. There are five possible actions - get, fetch, count, aggregate and exists. You can pass any subset of this ops to .cache() method even empty - to turn off caching. There is, however, a shortcut for the latter:

qs = Article.objects.filter(visible=True).nocache()
qs1 = qs.filter(tag=2)       # hits database
qs2 = qs.filter(category=3)  # hits it once more

It is useful when you want to disable automatic caching on particular queryset.

You can also override default timeout for particular queryset with .cache(timeout=...).

Function caching

You can cache and invalidate result of a function the same way as a queryset. Cached results of the next function will be invalidated on any Article change, addition or deletion:

from cacheops import cached_as

@cached_as(Article, timeout=120)
def article_stats():
    return {
        'tags': list(Article.objects.values('tag').annotate(Count('id')))
        'categories': list(Article.objects.values('category').annotate(Count('id')))
    }

Note that we are using list on both querysets here, it's because we don't want to cache queryset objects but their results.

Also note that if you want to filter queryset based on arguments, e.g. to make invalidation more granular, you can use a local function:

def articles_block(category, count=5):
    qs = Article.objects.filter(category=category)

    @cached_as(qs, extra=count)
    def _articles_block():
        articles = list(qs.filter(photo=True)[:count])
        if len(articles) < count:
            articles += list(qs.filter(photo=False)[:count-len(articles)])
        return articles

    return _articles_block()

We added extra here to make different keys for calls with same category but different count. Cache key will also depend on function arguments, so we could just pass count as an argument to inner function. We also omitted timeout here, so a default for the model will be used.

Another possibility is to make function cache invalidate on changes to any one of several models:

@cached_as(Article.objects.filter(public=True), Tag)
def article_stats():
    return {...}

As you can see, we can mix querysets and models here.

View caching

You can also cache and invalidate a view as a queryset. This works mostly the same way as function caching, but only path of the request parameter is used to construct cache key:

from cacheops import cached_view_as

@cached_view_as(News)
def news_index(request):
    # ...
    return render(...)

You can pass timeout, extra and several samples the same way as to @cached_as(). Note that you can pass a function as extra:

@cached_view_as(News, extra=lambda req: req.user.is_staff)
def news_index(request):
    # ... add extra things for staff
    return render(...)

A function passed as extra receives the same arguments as the cached function.

Class based views can also be cached:

class NewsIndex(ListView):
    model = News

news_index = cached_view_as(News, ...)(NewsIndex.as_view())

Invalidation

Cacheops uses both time and event-driven invalidation. The event-driven one listens on model signals and invalidates appropriate caches on Model.save(), .delete() and m2m changes.

Invalidation tries to be granular which means it won't invalidate a queryset that cannot be influenced by added/updated/deleted object judging by query conditions. Most of the time this will do what you want, if it won't you can use one of the following:

from cacheops import invalidate_obj, invalidate_model, invalidate_all

invalidate_obj(some_article)  # invalidates queries affected by some_article
invalidate_model(Article)     # invalidates all queries for model
invalidate_all()              # flush redis cache database

And last there is invalidate command:

./manage.py invalidate articles.Article.34  # same as invalidate_obj
./manage.py invalidate articles.Article     # same as invalidate_model
./manage.py invalidate articles   # invalidate all models in articles

And the one that FLUSHES cacheops redis database:

./manage.py invalidate all

Don't use that if you share redis database for both cache and something else.

Turning off and postponing invalidation

There is also a way to turn off invalidation for a while:

from cacheops import no_invalidation

with no_invalidation:
    # ... do some changes
    obj.save()

Also works as decorator:

@no_invalidation
def some_work(...):
    # ... do some changes
    obj.save()

Combined with try ... finally it could be used to postpone invalidation:

try:
    with no_invalidation:
        # ...
finally:
    invalidate_obj(...)
    # ... or
    invalidate_model(...)

Postponing invalidation can speed up batch jobs.

Mass updates

Normally qs.update(...) doesn't emit any events and thus doesn't trigger invalidation. And there is no transparent and efficient way to do that: trying to act on conditions will invalidate too much if update conditions are orthogonal to many queries conditions, and to act on specific objects we will need to fetch all of them, which QuerySet.update() users generally try to avoid.

In the case you actually want to perform the latter cacheops provides a shortcut:

qs.invalidated_update(...)

Note that all the updated objects are fetched twice, prior and post the update.

Simple time-invalidated cache

To cache result of a function call or a view for some time use:

from cacheops import cached, cached_view

@cached(timeout=number_of_seconds)
def top_articles(category):
    return ... # Some costly queries

@cached_view(timeout=number_of_seconds)
def top_articles(request, category=None):
    # Some costly queries
    return HttpResponse(...)

@cached() will generate separate entry for each combination of decorated function and its arguments. Also you can use extra same way as in @cached_as(), most useful for nested functions:

@property
def articles_json(self):
    @cached(timeout=10*60, extra=self.category_id)
    def _articles_json():
        ...
        return json.dumps(...)

    return _articles_json()

You can manually invalidate or update a result of a cached function:

top_articles.invalidate(some_category)
top_articles.key(some_category).set(new_value)

To invalidate cached view you can pass absolute uri instead of request:

top_articles.invalidate('http://example.com/page', some_category)

Cacheops also provides get/set primitives for simple cache:

from cacheops import cache

cache.set(cache_key, data, timeout=None)
cache.get(cache_key)
cache.delete(cache_key)

cache.get will raise CacheMiss if nothing is stored for given key:

from cacheops import cache, CacheMiss

try:
    result = cache.get(key)
except CacheMiss:
    ... # deal with it

File Cache

File based cache can be used the same way as simple time-invalidated one:

from cacheops import file_cache

@file_cache.cached(timeout=number_of_seconds)
def top_articles(category):
    return ... # Some costly queries

@file_cache.cached_view(timeout=number_of_seconds)
def top_articles(request, category):
    # Some costly queries
    return HttpResponse(...)

# later, on appropriate event
top_articles.invalidate(some_category)
# or
top_articles.key(some_category).set(some_value)

# primitives
file_cache.set(cache_key, data, timeout=None)
file_cache.get(cache_key)
file_cache.delete(cache_key)

It has several improvements upon django built-in file cache, both about high load. First, it's safe against concurrent writes. Second, it's invalidation is done as separate task, you'll need to call this from crontab for that to work:

/path/manage.py cleanfilecache
/path/manage.py cleanfilecache /path/to/non-default/cache/dir

Django templates integration

Cacheops provides tags to cache template fragments. They mimic @cached_as and @cached decorators, however, they require explicit naming of each fragment:

{% load cacheops %}

{% cached_as <queryset> <timeout> <fragment_name> [<extra1> <extra2> ...] %}
    ... some template code ...
{% endcached_as %}

{% cached <timeout> <fragment_name> [<extra1> <extra2> ...] %}
    ... some template code ...
{% endcached %}

You can use None for timeout in @cached_as to use it's default value for model.

To invalidate cached fragment use:

from cacheops import invalidate_fragment

invalidate_fragment(fragment_name, extra1, ...)

If you have more complex fragment caching needs, cacheops provides a helper to make your own template tags which decorate a template fragment in a way analogous to decorating a function with @cached or @cached_as. This is experimental feature for now.

To use it create myapp/templatetags/mycachetags.py and add something like this there:

from cacheops import cached_as, CacheopsLibrary

register = CacheopsLibrary()

@register.decorator_tag(takes_context=True)
def cache_menu(context, menu_name):
    from django.utils import translation
    from myapp.models import Flag, MenuItem

    request = context.get('request')
    if request and request.user.is_staff():
        # Use noop decorator to bypass caching for staff
        return lambda func: func

    return cached_as(
        # Invalidate cache if any menu item or a flag for menu changes
        MenuItem,
        Flag.objects.filter(name='menu'),
        # Vary for menu name and language, also stamp it as "menu" to be safe
        extra=("menu", menu_name, translation.get_language()),
        timeout=24 * 60 * 60
    )

@decorator_tag here creates a template tag behaving the same as returned decorator upon wrapped template fragment. Resulting template tag could be used as follows:

{% load mycachetags %}

{% cache_menu "top" %}
    ... the top menu template code ...
{% endcache_menu %}

... some template code ..

{% cache_menu "bottom" %}
    ... the bottom menu template code ...
{% endcache_menu %}

Jinja2 extension

Add cacheops.jinja2.cache to your extensions and use:

{% cached_as <queryset> [, timeout=<timeout>] [, extra=<key addition>] %}
    ... some template code ...
{% endcached_as %}

{% cached [timeout=<timeout>] [, extra=<key addition>] %}
    ...
{% endcached %}

Tags work the same way as corresponding decorators.

Transactions

Cacheops transparently supports transactions. This is implemented by following simple rules:

Once transaction is dirty (has changes) caching turns off. The reason is that the state of database at this point is only visible to current transaction and should not affect other users and vice versa.
Any invalidating calls are scheduled to run on the outer commit of transaction.
Savepoints and rollbacks are also handled appropriately.

Mind that simple and file cache don't turn itself off in transactions but work as usual.

Dog-pile effect prevention

There is optional locking mechanism to prevent several threads or processes simultaneously performing same heavy task. It works with @cached_as() and querysets:

@cached_as(qs, lock=True)
def heavy_func(...):
    # ...

for item in qs.cache(lock=True):
    # ...

It is also possible to specify lock: True in CACHEOPS setting but that would probably be a waste. Locking has no overhead on cache hit though.

Multiple database support

By default cacheops considers query result is same for same query, not depending on database queried. That could be changed with db_agnostic cache profile option:

CACHEOPS = {
    'some.model': {'ops': 'get', 'db_agnostic': False, 'timeout': ...}
}

Sharing redis instance

Cacheops provides a way to share a redis instance by adding prefix to cache keys:

CACHEOPS_PREFIX = lambda query: ...
# or
CACHEOPS_PREFIX = 'some.module.cacheops_prefix'

A most common usage would probably be a prefix by host name:

# get_request() returns current request saved to threadlocal by some middleware
cacheops_prefix = lambda _: get_request().get_host()

A query object passed to callback also enables reflection on used databases and tables:

def cacheops_prefix(query):
    query.dbs    # A list of databases queried
    query.tables # A list of tables query is invalidated on

    if set(query.tables) <= HELPER_TABLES:
        return 'helper:'
    if query.tables == ['blog_post']:
        return 'blog:'

NOTE: prefix is not used in simple and file cache. This might change in future cacheops.

Custom serialization

Cacheops uses pickle by default, employing it's default protocol. But you can specify your own it might be any module or a class having .dumps() and .loads() functions. For example you can use dill instead, which can serialize more things like anonymous functions:

CACHEOPS_SERIALIZER = 'dill'

One less obvious use is to fix pickle protocol, to use cacheops cache across python versions:

import pickle

class CACHEOPS_SERIALIZER:
    dumps = lambda data: pickle.dumps(data, 3)
    loads = pickle.loads

Using memory limit

If your cache never grows too large you may not bother. But if you do you have some options. Cacheops stores cached data along with invalidation data, so you can't just set maxmemory and let redis evict at its will. For now cacheops offers 2 imperfect strategies, which are considered experimental. So be careful and consider leaving feedback.

First strategy is configuring maxmemory-policy volatile-ttl. Invalidation data is guaranteed to have higher TTL than referenced keys. Redis however doesn't guarantee perfect TTL eviction order, it selects several keys and removes one with the least TTL, thus invalidator could be evicted before cache key it refers leaving it orphan and causing it survive next invalidation. You can reduce this chance by increasing maxmemory-samples redis config option and by reducing cache timeout.

Second strategy, probably more efficient one is adding CACHEOPS_LRU = True to your settings and then using maxmemory-policy volatile-lru. However, this makes invalidation structures persistent, they are still removed on associated events, but in absence of them can clutter redis database.

Keeping stats

Cacheops provides cache_read and cache_invalidated signals for you to keep track.

Cache read signal is emitted immediately after each cache lookup. Passed arguments are: sender - model class if queryset cache is fetched, func - decorated function and hit - fetch success as boolean value.

Here is a simple stats implementation:

from cacheops.signals import cache_read
from statsd.defaults.django import statsd

def stats_collector(sender, func, hit, **kwargs):
    event = 'hit' if hit else 'miss'
    statsd.incr('cacheops.%s' % event)

cache_read.connect(stats_collector)

Cache invalidation signal is emitted after object, model or global invalidation passing sender and obj_dict args. Note that during normal operation cacheops only uses object invalidation, calling it once for each model create/delete and twice for update: passing old and new object dictionary.

CAVEATS

Conditions other than __exact, __in and __isnull=True don't make invalidation more granular.
Conditions on TextFields, FileFields and BinaryFields don't make it either. One should not test on their equality anyway. See CACHEOPS_SKIP_FIELDS though.
Update of "selected_related" object does not invalidate cache for queryset. Use .prefetch_related() instead.
Mass updates don't trigger invalidation by default. But see .invalidated_update().
Sliced queries are invalidated as non-sliced ones.
Doesn't work with .raw() and other sql queries.
Conditions on subqueries don't affect invalidation.
Doesn't work right with multi-table inheritance.

Here 1, 2, 3, 5 are part of the design compromise, trying to solve them will make things complicated and slow. 7 can be implemented if needed, but it's probably counter-productive since one can just break queries into simpler ones, which cache better. 4 is a deliberate choice, making it "right" will flush cache too much when update conditions are orthogonal to most queries conditions, see, however, .invalidated_update(). 8 is postponed until it will gain more interest or a champion willing to implement it emerges.

All unsupported things could still be used easily enough with the help of @cached_as().

Performance tips

Here come some performance tips to make cacheops and Django ORM faster.

When you use cache you pickle and unpickle lots of django model instances, which could be slow. You can optimize django models serialization with django-pickling.
Constructing querysets is rather slow in django, mainly because most of QuerySet methods clone self, then change it and return the clone. Original queryset is usually thrown away. Cacheops adds .inplace() method, which makes queryset mutating, preventing useless cloning:
```
items = Item.objects.inplace().filter(category=12).order_by('-date')[:20]
```
You can revert queryset to cloning state using .cloning() call.

Note that this is a micro-optimization technique. Using it is only desirable in the hottest places, not everywhere.
Use template fragment caching when possible, it's way more fast because you don't need to generate anything. Also pickling/unpickling a string is much faster than a list of model instances.
Run separate redis instance for cache with disabled persistence. You can manually call SAVE or BGSAVE to stay hot upon server restart.
If you filter queryset on many different or complex conditions cache could degrade performance (comparing to uncached db calls) in consequence of frequent cache misses. Disable cache in such cases entirely or on some heuristics which detect if this request would be probably hit. E.g. enable cache if only some primary fields are used in filter.

Caching querysets with large amount of filters also slows down all subsequent invalidation on that model. You can disable caching if more than some amount of fields is used in filter simultaneously.

Writing a test

Writing a test for an issue you are experiencing can speed up its resolution a lot. Here is how you do that. I suppose you have some application code causing it.

Make a fork.
Install all from requirements-test.txt.
Ensure you can run tests with ./run_tests.py.
Copy relevant models code to tests/models.py.
Go to tests/tests.py and paste code causing exception to IssueTests.test_{issue_number}.
Execute ./run_tests.py {issue_number} and see it failing.
Cut down model and test code until error disappears and make a step back.
Commit changes and make a pull request.

TODO

faster .get() handling for simple cases such as get by pk/id, with simple key calculation
integrate previous one with prefetch_related()
shard cache between multiple redises
respect subqueries?
respect headers in @cached_view*?
group invalidate_obj() calls?
a postpone invalidation context manager/decorator?
fast mode: store cache in local memory, but check in with redis if it's valid
an interface for complex fields to extract exact on parts or transforms: ArrayField.len => field__len=?, ArrayField[0] => field__0=?, JSONField['some_key'] => field__some_key=?
custom cache eviction strategy in lua
cache a string directly (no pickle) for direct serving (custom key function?)

ducva/django-cacheops