>> But can't we read it only when needed ?
>> Just stat'ing it should be much cheaper than reading it each time.
We'd already created a precedent for read-on-HUP for postgresql.conf,
and no one seemed to be complaining about that. So although this change
will doubtless annoy some existing users, I don't see a big problem
with it. I'm happy to avoid the stat() call --- every kernel call we
can remove from the startup sequence is another small win.
regards, tom lane