Update:

< * Consider compressing indexes by storing key prefix values shared by
> * Consider compressing indexes by storing key values duplicated in
735a736,737
>
>   This is difficult because it requires datatype-specific knowledge.