Currently, we keep all usable SDCH dictionaries in RAM.  This is a nice, simple way to guarantee that we can successfully decompress files compressed using SDCH dictionaries that we have in memory, but it has some problems.

It limits the number of SDCH dictionaries we can use and advertise (Currently capped at 1MB), so if two sites try and use SDCH, they can evict each other's dictionaries long before we run into cache contention.

It also takes up a fair bit of memory.  This is particularly an issue on mobile, where memory is more limited, and paging isn't supported.  It can lead to more frequent process eviction.