Tomas Vondra <tv(at)fuzzy(dot)cz> writes:
> So maybe we shouldn't cling to the WAL-logging approach too much. Maybe
> Heikki's idea from to abandon the full checkpoint and instead assume
> that once the transaction commits, all the files were fsynced OK. Of
> couse, this will do nothing about the replay hazards.
Well, I'm not insisting on any particular method of getting there, but
if we're going to touch this area at all then I think "fix the replay
hazards" should be a non-negotiable requirement. We'd never have accepted
such hazards if CREATE DATABASE were being introduced for the first time;
it's only like this because nobody felt like rewriting a Berkeley-era
kluge.
regards, tom lane