The obvious volatile memory choices were either Memcached or Redis. We eventually sided with Redis as it natively handled replication, and this replication fell under the PCI responsibility scope of AWS. This added duribility meant that if a redis node were to die, our downtime would be in the seconds, rather than 15 minutes which we would incur using Memcached

The requirement was the classic "cache the results of a SQL query for a period of time."

While the Internet is full of "Redis is fuller featured" posts, the key issue for us was the actual performance. We discovered, in various stress scenario testing, that Memcached outperformed Redis for simple key-value retrieval dramatically (over twice as fast.) That's not to say that Redis is bad - we use that in other places where the requirements are more sophisticated than simple key/value retrieval.

 This was roughly 4 years ago at this point. We had been using an old iteration of memcache on Windows as the data cache per server for while and had for whatever reason opted to store our session data (to which the application is heavily dependent) in App Fabric. App Fabric had come to EOL and we needed to move away from it. As a quick search showed throughput capacity to be higher and overall features of Redis were better we initially implemented it. The number of changes required were minimal and we were able to migrate away to a more resilient system pretty quickly.
 We hit a snag in that the implementation of the Redis session handler at that point only took a single IP so we had to do use keepalived and HAProxy to display the application consistently between master and slave failovers. This caused issues on occasion of dropped connections to the backend service. We upgraded our client and could put both members into our config files and it stopped timing out. We were all happy in this and it was (for it's own part) a significant upgrade. Generally performance was better for all pages. 
 We found however, that our application was serializing all requests and locking on the thread via session lock for a great many things and this caused us and our users considerable headaches on occasion. We found that the implementation of the couchbase session handler gave us the option of ignoring session write locks and not be required to rewrite dozens of pages to handle the no session requirement as the application was working fine without the need to be threadsafe. Though the maximum throughput was not as good compared to Redis the application performance was considerably better as a result of the change. The multi-master write was a big benefit and the cross data center replication was a nice thing to have as that would allow our users to remain logged in even across DataCenter fail-over events (praying that never happened). Overall we used that for session handling and chose to use couchbase (in a 2nd cluster) to handle memcache requests as that gave us greater capacity to handle larger objects more efficiently. 
 We are, all these years later, looking to move into the newer features of couchbase to give us more and better use of this product that really has been the answer to a bunch of the growing pains we experienced. Since the decision performance has not been on wild rides and stability has never been better. So I sing the praises of couchbase to anyone that will listen.