> You risk being without a caching layer for a large period of time if you need to restart Redis, where Memcache, due to the lack of persistance is ready right away. Honestly I'm not sure it that still an issue, but I believe it is.
The bigger issue here is starting an empty (cold) caching instance and overloading backends because all the requests are MISSes and go to the backends. Persistence allows you to avoid this problem. Even if you run caching servers in some geo distributed setup, persistence will still help by reducing unnecessary geo bandwidth.
It's pretty hard to find even a single use case where in-memory caching server without persistence is better than the one with.
I agree with you 100%. I've personally suffered from caching stampedes (https://en.wikipedia.org/wiki/Cache_stampede) and having, at least, a partial cache is a phenomenal feature.
(Needless to say, I have learned the hard way not to let evict all cache items at the same time but to have a TTL which is a random number between min and max.)
Regardless of this, I still feel bad every time a TTL based cache is used since I strongly favour event based evictions, but this is a totally different (and very long) topic.
I think it really depends if and how you're going to use the Redis features. If you plan to make Redis a big part of your overall design, then maybe separating caching out into Memcache servers still makes sense.
One issue I've run into is shutdown and start up time for Redis, when you're storing large amounts of data. You risk being without a caching layer for a large period of time if you need to restart Redis, where Memcache, due to the lack of persistance is ready right away. Honestly I'm not sure it that still an issue, but I believe it is. Of cause this is only an issue if you're using Redis to store other data than just your cache, but the temptation to do so is pretty high once Redis is available to you.
You could of cause have two instances of Redis, but if you're not using the features in one of the Redis installation, then maybe you could just as easily use Memcache.
As with so much else, it's a question of your use case, how much data you have, what type of data and how you expect to be able to access it.