> So ... not one of the buildfarm members that are running TAP tests > likes this. ... > I think probably what's happening is that pg_dump is still trying to dump > directly from the catalog, when what it needs to do now is dump from the > view, in case it's not running as superuser.
I experimented with extracting the required data from the view, and there are at least two show-stopper problems:
* The view doesn't expose pg_statistic_ext.oid, which pg_dump has to have for dependency tracking purposes. I think we could just add it though. Now that OIDs are ordinary columns it won't even look very odd.
* Rather than just not exposing the critical data for stats you don't have privilege to read, the view doesn't expose anything at all. I do not think that's acceptable; it creates a significant hazard of data loss during pg_dump, for no very good reason. What we should be doing, IMO, is still showing all the rows but filling the data-value columns with nulls in rows where the caller can't access the underlying data.
Hang on. Isn't the real problem that we should be revoking public access from pg_statistic_ext_data, not pg_statistic_ext? Normal users should still be able to see the stats definitions in the catalog table, but not the stats data, so I think pg_dump wouldn't need changing.