Redis uses a memory layout that, unless the instance is completely idle, will make it very easy to touch every page even if there are just a few queries per second.
So fare in two years we never saw an instance where there was a latency problem because part of the dataset was swapped on disk by the OS. This is why we currently don't use mlockall().
How can they guarantee that, when RAM is little more than a disk cache on modern operating systems? Is there a syscall that prevents specified pages of virtual memory from being paged out to disk? Or is Redis implemented as a kernel module?