[bentorfs]

> The pseudo-LRU redis used to have is appropriate for very high-speed caches (it's essentially a 5-way associative cache), but is so very not appropriate when going out to disk, where the access latency is horrible.

Redis is an in-memory database, so what's the problem?

Also, keep in mind that it has to fit in 24 bits, so not all solutions are possible, even if they are theoretically better

[onetwotree]

It's possible that he refers to a cache miss "going out to disk" by requiring a read from a different data store (a SQL database, for example).

But this is really kind of silly - the entire idea with caching is that misses are much slower than hits. The statement is something like "you need a stupid good cache because misses hurt". Software engineering is always about tradeoffs, and antirez does a fantastic job of describing them in this post.