I've reached a point where my default design is relational database augmented with memcache/redis. The relational DB is the authoritative source of truth, redis absorbs all (most) reads and memcache is, well a distributed cache. Writes go to both redis and relational. Redis data can always be rebuilt, and is regularly rebuilt, from relational data.

At project onset, usually only redis or memcache are in use.

It's a model that has served me well, apart from bugs that cause divergence between redis and relational. These are usually the class of bugs that would corrupt data were I relying on redis alone, so I just think of them as a lesser evil.

There's an obvious bottleneck on writes to the relational DB but I've so far managed to keep from hitting that ceiling.