John Lister <john.lister-ps@kickstone.com> writes:
> GMT LOG: relation "pg_class" TID 15538/4: dead HOT-updated tuple ---
> cannot shrink relation
Hmm. The comments in vacuum.c about this case suppose that it could
only be a transient condition, ie the tuple became dead while we were
looking at the page. Evidently it's persistent for you, which means
that for some reason heap_page_prune() is failing to remove an
already-dead tuple. I suspect this implies a corrupt HOT chain, but
maybe the data is okay and what you've got is really a bug in
heap_page_prune.
Could you send a dump of page 15538 of pg_class, using pg_filedump?
The source code for it is here:
http://sources.redhat.com/rhdb/utilities.html
Best would be -i -f format, iepg_filedump -i -f -R 15538 $PGDATA/base/something/1259
regards, tom lane