I ditched memcached for redis a while ago, just because I was already using redis for my queueing, and it made sense to eliminate redundant dependencies.
I'm now using Redis clusters to distribute cache reads so that each webserver has its own local copy of the cache which it reads from, but there's a single master that accepts writes. This does occassionally mean a double-write, but since we're talking about cache data and redis is atomic, this isn't a problem, and it splits the load of my reads, so that I don't end up bottlenecking if the master is being slow.
I'm now using Redis clusters to distribute cache reads so that each webserver has its own local copy of the cache which it reads from, but there's a single master that accepts writes. This does occassionally mean a double-write, but since we're talking about cache data and redis is atomic, this isn't a problem, and it splits the load of my reads, so that I don't end up bottlenecking if the master is being slow.