BTW: The system is tailored to SSDs (small random reads and large
sequential writes), because all revisions are written sequentially and
reconstructing nodes (or in the future any other stored data) requires
reading scattered data (even more, for instance during reconstruction of
incremental and differential page deltas).
