Redis can run anywhere from 50-200K IOPs, on a single core, dependent on CPU; you can find benchmarks at the following:

    * http://code.google.com/p/redis/wiki/Benchmarks
    * http://redis.io/topics/benchmarks

At the point where multi-core sharding/ring setups are involved, you begin to have NIC/network saturation. It's feasible to achieve 300K on a single box.

That's amazing, thank you. This would make me want to use redis for everything, if I weren't already using it for everything.

The OP actually addresses this. They're doing some form of clustering and had to add more nodes due to network saturation.

Just to clarify, they're using regular master-slave, as opposed to "clustering" which is an unreleased feature.

I love how they use sInterStore as a replacement for a fat SQL query.

It would be mostly reads so could be done with lots of slaves and sharding without redis cluster.

Very true, but now I wonder how many reads/sec each of those servers pushed.