Improved memcached cache backend for Django
Newcache is an improved memcached cache backend for Django. It provides four major advantages over Django’s built-in cache backend:
- It supports pylibmc.
- It allows for a function to be run on each key before it’s sent to memcached.
- It supports setting cache keys with infinite timeouts.
- It mitigates the thundering herd problem.
It also has some pretty nice defaults. By default, the function that’s run on each key is one that hashes, versions, and flavors the key. More on that later.
The simplest way is to just set it as your cache backend in your settings.py, like so:
CACHE_BACKEND = 'newcache://127.0.0.1:11211/?binary=true'
Note that we’ve passed an additional argument, binary, to the backend. This is because pylibmc supports using binary mode to talk to memcached. This is a completely optional parameter, and can be omitted safely to use the old text mode. It is ignored when using python-memcached.
Earlier we said that by default it hashes, versions, and flavors each key. What does this mean? Let’s go through each item in detail.
Keys in memcached come with many restrictions, both on their length and on their contents. Practically speaking, this means that you can’t put spaces in your keys, and they can’t be very long. One simple solution to this is to create an sha1 hash of whatever key you want, and use the hash as your key instead. That is what we do in newcache. It not only allows for long keys, but it also lets us put spaces or other characters in our key as well.
Sometimes it’s necessary to clear the entire cache. We can do this using memcached’s flushing mechanisms, but sometimes a cache is shared by many things instead of just one web app. It’s a shame to have everything lose its fresh cache just because one web app needed to clear its cache. For this, we introduce a simple technique called versioning. A version number is added to each cache key, and when this version is incremented, all the old cache keys will become invalid because they have an incorrect version.
This is exposed as a new setting, CACHE_VERSION, and it defaults to 1.
Finally, we found that as we split our site out into development, staging, and production, we didn’t want them to share the same cache. But we also didn’t want to spin up a new memcached instance for each one. So we came up with the idea of flavoring the cache. The concept is simple–add a FLAVOR setting and make it something like ‘dev’, ‘prod’, or ‘test’. With newcache, this flavor string will be added to each key, ensuring that there are no collisions.
Concretely, this is what happens:
# CACHE_VERSION = 2 # FLAVOR = 'staging' cache.get('games') # ... would actually call ... cache.get('staging-2-9cfa7aefcc61936b70aaec6729329eda')
All of the above is simply the default, you may provide your own callable function to be run on each key, by supplying the CACHE_KEY_MODULE setting. It must provide a get_key function which takes any instance of basestring and output a str.
The thundering herd problem manifests itself when a cache key expires, and many things rush to get or generate the data stored for that key all at once. This is doing a lot of unnecessary work and can cause service outages if the database cannot handle the load. To solve this problem, we really only want one thread or process to fetch this data.
Our method of solving this problem is to shove the old (expired) value back into the cache for a short time while the first process/thread goes and updates the key. This is done in a completely transparent way–no changes should need to be made in the application code.
With this cache backend, we have provided an extra ‘herd’ keyword argument to the set, add, and set_many methods–which is set to True by default. What this does is transform your cache value into a tuple before saving it to the cache. Each value is structured like this:
(A herd marker, your original value, the expiration timestamp)
Then when it actually sets the cache, it sets the real timeout to a little bit longer than the expiration timestamp. Actually, this “little bit” is configurable using the CACHE_HERD_TIMEOUT setting, but it defaults to 60 seconds.
Now every time we read a value from the cache, we automatically unpack it and check whether it’s expired. If it has expired, we put it back in the cache for CACHE_HERD_TIMEOUT seconds, but (and this is the key) we act as if it were a cache miss (so we return None, or whatever the default was for the call.)