My impression is that databases have known of the shortcomings of mmap since, like, the 90s. A critical flaw is the lack of error handling -- in addition to the unpredictable performance characteristics and naive caching. I'm looking forward to reading the paper.
