On Tue, May 22, 2018 at 09:17:15AM -0400, Robert Haas wrote:
> Another interesting question is how the system got into this state in
> the first place, of course. Was it a bug, or did you do something
> that corrupted the database?
Perhaps a side effect which has been fixed in d2599ecf for 9.6.9?
The thing is that pruning chains broken by bugs like freeze-the-dead
could remain undetected for some time. On top of that autovacuum does
not check after duplicated OIDs registered when building the list, and
executes each table in its own transaction, so an autoanalyze would not
have seen the issue, still it was doing twice the work. One thing could
be to run diagnostic checks on all system catalogs (pg_catcheck can
help here).
--
Michael
