Re: pg_class.relpartbound definition overly brittle

On Wed, May 31, 2017 at 3:40 PM, Mark Dilger <hornschnorter(at)gmail(dot)com> wrote:
> recent changes have introduced the :location field to the partboundspec
> in pg_catalog.pg_class. This means that if two identical tables with
> identical partitioning scheme are created, but one is done before a change
> to gram.y, and the other after a change to gram.y, the relpartbound fields
> for those two tables could show up as different.
>
> Can we perhaps remove the :location field here? If not, could somebody
> please defend why this belongs in the catalog entries for the table? Sorry
> if I am missing something obvious...

Yeah, that probably wasn't a good decision, although calling a
decision might be giving it too much credit. I think the easiest
thing to do here would be to change transformPartitionBound() to set
result_spec->location to -1, although maybe a better idea would be to
have two different structures -- one that represents the partition
bound specification *before* parse analysis and another that
represents it *after* parse analysis, rather than reusing the same
structure for both. Then again, maybe making two different node types
for this is overkill. Not sure. Opinions?

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company