188 karma · joined April 30, 2021
It's not that bad to use a fork during 2-3 release cycles give that: 1) It's not highly divergent fork (the patchset is small). 2) They patchset will be smoothly decrease over that period of time. Also, even 2-3 release cycles isn't that bad in comparison with decades while we have small progress in wicked problems.
Also, I have to mention that zheap development is inactive for now. And its undo implementation was never complete. Zheap undo works only if you don't update any indexed column. Otherwise, it works as plain heap with write-amplification and bloat.
I would say "Snapshot too old" isn't essential drawback of undo log. DBMS can keep the the undo log records until all snapshots, which needs them, are released. And this is how OrioleDB behaves. This is also kind of "bloat", but cleanup is cheaper. You just have to cut unclaimed undo log instead of expensive vacuum.
The design document will be published later.
1. The first goal is to become a pure extension. That should be done in 2-3 PostgreSQL release cycles. 2. The long term goal is to become a part of PostgreSQL.