Robert Haas <robertmhaas@gmail.com> writes:
> The problem is that if a relfilenode ever gets assigned by
> GetNewRelFileNode() during a binary-upgrade restore, that OID may turn
> out to be used by some other object later in the dump. And then
> you're dead, because the dump restore will fail later on complaining
> about, well, I forget the error message wording exactly, but,
> basically, an OID collision.
Right.
> So if we change the rules in such a way
> that objects which currently lack TOAST tables acquire them, we need
> to first restore all of the objects *without* adding any new TOAST
> tables, and then at the end create any new TOAST tables once we have a
> complete list of the relfilenodes that are actually used.
toasting.c explains why this currently doesn't seem necessary:
/* * In binary-upgrade mode, create a TOAST table if and only if * pg_upgrade told us to (ie, a
TOASTtable OID has been provided). * * This indicates that the old cluster had a TOAST table for the
* current table. We must create a TOAST table to receive the old * TOAST file, even if the table seems not to
needone. * * Contrariwise, if the old cluster did not have a TOAST table, we * should be able to
getalong without one even if the new version's * needs_toast_table rules suggest we should have one. There is a
lot * of daylight between where we will create a TOAST table and where * one is really necessary to avoid
failures,so small cross-version * differences in the when-to-create heuristic shouldn't be a problem. *
Ifwe tried to create a TOAST table anyway, we would have the * problem that it might take up an OID that will
conflictwith some * old-cluster table we haven't seen yet. */
I don't really see any near-term reason why we would need to change this.
In the long run, it would certainly be cleaner if pg_upgrade dropped
the force-the-relfilenode-assignment approach and instead remapped
relfilenodes from old cluster to new. But I think it's just for
cleanliness rather to fix any live or foreseeable bug.
regards, tom lane