On 28/07/16 17:52, Jason Dusek wrote: > With regards to write amplification, it makes me think about about > OIDs. Used to be, every row had an OID and that OID persisted across > row versions. > > https://www.postgresql.org/docs/9.5/static/runtime-config-compatible.html#GUC-DEFAULT-WITH-OIDS > > Would reintroducing such a feature address some of Uber's concerns > about multiple indexes? It could, and would do so without the implicit > requirement of a foreign key; but it would also require a fast OID to > CTID mapping.
Would it be best to increase OIDs to 64 bits?
Possibly a choice of 32/64 to be decided when the DB is created???
Moving to 64 bit OIDs would make the OID->CTID mapping take more memory and could, consequently, make it slower; but OIDs would seem to be required to support temporal tables so maybe they should make a comeback?