Re: ALTER TABLE .. SET WITH / WITHOUT OIDS

Rod Taylor <rbt(at)rbt(dot)ca> writes:
> Works somewhat like adding / dropping a column with the default of 0.

After looking at this patch I'm inclined to reject it. The doc patch
claims

Adds or removes the <literal>OID</literal> column from a table.
When enabling (setting with) OIDs in a table, the currently
existing tuples are not immediately numbered.

but the actual behavior is that the existing tuples are not numbered and
*never can become numbered* --- only freshly-inserted rows will acquire
OIDs. That puts a severe, probably fatal, limit on the usefulness of
SET WITH OIDS.

I can see some value in the SET WITHOUT OIDS half of the patch, as this
gives a way of reclaiming space in a table that was unthinkingly created
with OIDs. But I'm not sure we should expend the code space to provide
just a one-way transformation.

Or we could modify heap_update to assign an OID if it reads a zero OID
from the old tuple in a relation that is marked as having OIDs. Not
sure whether I like that behavior either, but at least it would provide
a mechanism for causing the old rows to acquire OIDs.

Comments?

regards, tom lane