adsj@novozymes.com (Adam Sjøgren) writes: >> [... still waiting for the result, I will return with what it said >> when the server does ...]
> It did eventually finish, with the same result:
Huh. So what we have here, apparently, is that regular MVCC snapshots think there is exactly one copy of the 1698936148/0 row, but TOAST fetches think there is more than one. This is darn odd, not least because we never do UPDATEs in toast tables, only inserts and deletes, so there certainly shouldn't be update chains there.
It seems like you've got some corner case wherein SnapshotToast sees a row that isn't visible according to MVCC --- probably a row left over from some previous cycle of life. That is, I'm imagining the OID counter wrapped around and we've reused a toast OID, but for some reason there's still a row in the table with that OID. I'm not sure offhand how we could get into such a state. Alvaro, does this ring any bells (remembering that this is 9.3)?
FWIW one of our support customers reported a very similar TOAST table corruption issue last week which nearly caused an outage. After a lot of analysis, I think I've now fully understood the reasons behind the corruption, the underlying bug(s) and possible remedy. I am currently working on writing a reproducible test case to demonstrate the problem and writing the fix. More details on that soon.
I've posted a reproducer and a proposed fix to -hackers [1]
In the particular case that I investigated, a database crash/recovery was involved. But I think we should be able to create a scenario where OID wrap-around or a standby promotion triggers the problem. I don't know if any of that was involved in the cases reported on this thread, but I've a strong suspicion that the underlying bug is probably the same.
