Redis for us almost operates as a set of application servers. We push a great deal of logic into Lua that we compile so that tons of back and forth operations aren't necessary. This also limits the data sent back to the API servers since only the structures needed are sent back.
We're rather coupled with Redis as a result with much of the application logic in Lua and our data structure designed around Redis structure types and commands.
we used Tokumx for http://blog.maxcdn.com/learned-stop-worrying-love-logs/ and the write performance vs mongo was incredible. I think the cluster is doing ~250,000 writes per second across six boxes with SSD space being the bottleneck. I'm really curious to spin up AerospikeDB. There's no Tokumx comparison I could find and they talk about very high reads. Curious to see write heeavy applications.
I'm testing it out and it seems pretty legit if you need a big, big key value store and don't want to wrangle redis or memcached shards.
Non-Java drivers are a bit behind, but nothing seems to have the same 95%tile latency, fast failover, SSD performance for actual read loads, and relatively easy scale out.
We're rather coupled with Redis as a result with much of the application logic in Lua and our data structure designed around Redis structure types and commands.