Re: Re: AW: Re: OID wraparound: summary and proposal

Hannu Krosing <hannu(at)tm(dot)ee> writes:
> Will it be easier to make Relation shared and persistent or creating
> a new shared structure that has just a counter+lock for each
> relation oid ?

The latter. Relation (by which I mean a whole relcache entry with all
its subsidiary structure, not only struct RelationData) is too large,
complex and heavyweight a structure to be a good candidate for moving
into shared memory. It also contains a lot of backend-local status
data in its current incarnation.

Some kind of shared cache for sequence generators (essentially,
generalizing the existing shared OID counter into N counters) is
probably the answer. But it would have to be a cache, not the whole
truth, so there'd need to be an underlying table that holds counters not
currently swapped into cache. That part we don't have a good model for
in the existing OID-generator code, nor in the existing sequence code.

regards, tom lane