Michael Guerin <guerin(at)rentec(dot)com> writes:
>> Hmm, that makes it sound like a plain old data-corruption problem, ie,
>> trashed xmin or xmax in some tuple header. Can you do a "select
>> count(*)" from this table without getting the error?
>>
> no, select count(*) fails around 25 millions rows.
OK, so you should be able to narrow down the corrupted row(s) and zero
them out, which'll at least let you get back the rest of the table.
See past archives for the standard divide-and-conquer approach to this.
regards, tom lane