Nathan Bossart <nathandbossart@gmail.com> writes:
> I think the main difference between your patch and mine is that I've
> exposed vac_update_datfrozenxid() via a function instead of a VACUUM
> option. IMHO that feels a little more natural, but I can't say I feel too
> strongly about it.
I thought about that but it seems fairly unsafe, because that means
that vac_update_datfrozenxid is executing inside a user-controlled
transaction. I don't think it will hurt us if the user does a
ROLLBACK afterward --- but if he sits on the open transaction,
that would be bad, if only because we're still holding the
LockDatabaseFrozenIds lock which will block other VACUUMs.
There might be more hazards besides that; certainly no one has ever
tried to run vac_update_datfrozenxid that way before.
regards, tom lane
