I think that was spurious. It looked good when we looked at it yesterday. The rest that failed seemed unrelated and was also taking on my SSL patch too.
I talked to Andres about the possibility of torn reads in the pg_class stats but those are all 4-byte columns so probably safe. And in any case that's a pre-existing possibility just more likely (if it's possible at all) by frequent truncates.
> So here I've done it that way. It is a bit of an unfortunate layering > since it means the heapam_handler is doing the catalog change but it > does seem inconvenient to pass relfrozenxid etc back up and have the > caller make the changes when there are no other changes to make.