block cache sketch

[kylebarron]

• We should have separate public block caching APIs for sync and async backends. This way a backend doesn't have to know how to become sync or async itself.
• The idea is that the underlying cache can be independent of whether the fetching is sync or async. (This assumes that there's no-overhead synchronous access to the cache. If we were to add file-based caching in the future, then maybe we'd want to have an async cache as well).
  • In the future we could even make the idea of a cache generic. Where the cache can declare specific methods like ranges_for_request, get, set. And then a generic SyncCacheStore and AsyncCacheStore could use a generic cache, while not tying the cache implementation to the name of the wrapper. (I.e. we currently have a BlockCache class, where the cache implementation is tied to the specific wrapper)
• BlockStoreReader shouldn't implement a generic read method. Instead the block cache should just implement and expose publicly the same get_range & get_ranges or get_range_async & get_ranges_async as the underlying backend. Then we can have a generic Read wrapper that takes an arbitrary backend that supports get_range and get_ranges and implements file methods on them.