The in-memory size of ICU is important. Both that and performance are closely related to how we cache. We definitely have the opportunity to improve both size and performance by doing better caching. Right now we don't cache things we should cache, and we don't dump things out of the cache when they are sufficiently old or unused. Because we have separate caches for different objects, we also can't have a good overall strategy for removal when no longer needed.
Here are some thoughts on how to do a better job. At least part of this involves making the job of caching easier, so that it isn't as onerous to "do the right thing" for programmers.
At a low level, we have a global cache. To get something out of the cache, you supply a service key, a specific key, and a functor. The functor is specific to the service, and knows how to create the requested object if not available in the cache. The cache also supplies iteration over service keys, and within that over specific keys. Public methods can also flush the cache, either partially (to meet some maximum size objective or other criterion) or completely. Of course, the cache is thread-safe. We also provide low-level hooks, so that people can replace the caching mechanism with their own.
At a higher level, we use the low level cache to create classes (by subclassing) that make this syntactically easier for the service. Each such class knows about a type of object, and how to create one, so it will supply the functor needed for the global cache. These classes may provide some semantic sugar to cover common cases:
- immutable objects
- mutable objects internal to the service
- mutable objects that the service will clone and allow access to (eg through getInstance())
Some of this overlaps with the current registration mechanisms, so we should look at those at the same time; we may be able to do some sharing of code, although the registration has somewhat different goals.