On Fri, Oct 21, 2016 at 1:37 PM, Merlin Moncure <mmonc...@gmail.com> wrote: > On Fri, Oct 21, 2016 at 8:03 AM, Kevin Grittner <kgri...@gmail.com> wrote: >> On Tue, Oct 18, 2016 at 8:45 AM, Merlin Moncure <mmonc...@gmail.com> wrote: >> >>> Most or all the damage seemed to be to the system catalogs with >>> at least two critical tables dropped or inaccessible in some >>> fashion. A lot of the OIDs seemed to be pointing at the wrong >>> thing. >> >> While the oid in pg_class often matches the filename, that is not >> true after some operations (like CLUSTER or VACUUM FULL). It is >> the relfilenode column that is the definitive link to the file. > > no such operations happened. In the first instance at least one table > dropped from the system catalogs. I have a hunch that the heap is > fine (supported by the size of the database on disk). At this > precise moment I'm restoring the database to another fileserver in > order to do some forensic analysis, also in the hopes of getting the > second database online in order to expedite recovery. > > ah -- done. :-) deleting the init file didn't help, but starting up > single user allowed the start up to gracefully fail with a FATAL cache > lookup.
OK, I have some good (very- in the specific case of yours truly) news to report. Doing a filesystem level copy to a test server I was able to relfilenode swap one of the critical tables over the place of the refilenode of the stored backup. Not being able know the file to copy from, I figured out the source node by judging the size and using 'strings' utility. Data recovery for that table at least appears to be 100%. For those following along, this simple process is only likely to work easily if the table contains only system types; no user types, enums, composites, etc, since those have a unique ID for each data restore. merlin -- Sent via pgsql-hackers mailing list (email@example.com) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers