Django's cache layer sucks. Simply stated and simply true. Any time I decide I can cache something, I should ask myself if I could have built it before I even had the request in the first place. Doing that with the template caches simply isn't possible. It should be possible and it should be the first path you take, instead of forcing us to go out of our way to do the better thing. Anything I might want to cache, I also might want to be sure I'm not doing in more place than once, and forcing them inline in my templates does not help this. The template caches imply a copy-and-paste method of reuse when a cached portion is used in more place than one. When I define a cache block, I name it and I specify a set of keys. This is exactly the information, that when changed, I should just generate that block as a static snippet to be inserted. If it weren't for the lacking in reuse mechanics, I would advocate parsing all your templates for cache blocks and pre-generating them. Instead, we need to pull the cached contents out of the normal templates and use the existing names and keys to find the generated snippets.
On the more basic level, there are some abstractions that need to be injected into Django-proper to really be useful, by means of what they would standardize. We have no current means of standardizing our cache keys in a way that different applications can cooperate about what data is where and how to get it. Even the types that are taken for granted in Django have no useful standards. If they did, I would be able to drop a QuerySet object into the cache in a way that another query can find to reuse. And, when memcached is by far the most likely cache backend to be used, we would be providing a mechanism that abstracted away its limitations in entry size, allowing us to trust dropping our QuerySet in safely.
Denormalization should be normal. I have revision tracking in a document system, and from a normalization perspective it makes sense that each version hold a foreign key to either its previous or next version, but not both. From a practicality perspective, if I have one version I want to know the previous and next versions without doing a new query. Our Resources might offer a solution, by giving us some place outside of our model to allow denormalized data. I could generate a record of my documents with all the revision information queried and built and stored in one flat record, while keeping my base model clean.
Queuing work should be as accessible as doing work. There is little or nothing inhibiting a developing from dropping one little query or action into an existing operation. I've recently built a weighted sort to replace our basic date and time based order for posts. This means generating scores for all the posts and updating those when posts or votes change. Now, whenever we calculate scores we account for the age of all votes and the relative scores and age of all posts and votes together. In other words, this is something I'd prefer not to add to the cost of a user actually posting content or voting on something. It would have been extremely easy for me to call one generate_scores() function, but it takes thought, planning, and infrastructure to have this done after the request is handled.
Borrowing from existing Python canon makes sense, so I think multiprocessing is a candidate for use here, in one form or another. multiprocessing.Pool.apply_async() without a result returned fits the bill for an interface to call some function at another time, possibly in another process. Any function that works when passed through multiprocessing into another process should also work when queued up for execution at some later time, so borrowing here reusing existing semantics developers should be familiar with.