Like all my notes in this series, I’ve only included portions I’ve found interesting.
Why not datastore x?
I was thinking this question when I first heard of this paper.
The keys and values can be made up of multiple columns
(Date, Publisher, Country) -> (ClickCount, Cost)
x < y)
The delta compaction policy determines how many deltas and individual versions to keep around and when to create new deltas to subsume smaller deltas
Specifically: * What deltas must be created before a version can be queried * For performance, I assume * When deltas should be created outside of the update path * When deltas can be deleted