Presumably it works by mapping that file underneath a regular memory allocator, that means extremely small and random IO patterns. I think "must" is appropriate as a real backing file would likely mean a slowdown of 1-2 orders of magnitude! Compare to a typical DB system where the allocation structure always has some kind of large block IO locality by design
The alternative to 'must' is 'should', and people ignore that more easily, resulting in bug reports like 'memcached performance abysmal when running with disk', 'why does memcache cause 1MB of IO when I only read 256 4-byte keys', etc
Not every memory access to a memory mapped file is automatically paged out immediately, so depending on vm dirty bytes, etc, you may get away with it for a while.
But the worst case will be dismal and not unlikely.
The alternative to 'must' is 'should', and people ignore that more easily, resulting in bug reports like 'memcached performance abysmal when running with disk', 'why does memcache cause 1MB of IO when I only read 256 4-byte keys', etc