On Tue, Aug 6, 2019 at 11:11 AM Larry Rosenman <ler@lerctr.org> wrote:
> As a followup, btcheck found another index that had issues, and a toast
> table was missing a chunk.
>
> I have ALL the data I used to create this table still around so I just
> dropped it and am reloading the data.
It sounds like there is a generic storage issue at play here. Often
TOAST data is the apparent first thing that gets corrupted, because
that's only because the inconsistencies are relatively obvious.
I suggest that you rerun amcheck using the same query, though this
time specify "heapallindexed=true" to bt_check_index(). Increase
maintenance_work_mem if it's set to a low value first (ideally you can
crank it up to 600MB). This type of verification will take a lot
longer, but will find more subtle inconsistencies that could easily be
missed.
Please let us know how this goes. I am always keen to hear about how
much the tooling helps in the real world.
--
Peter Geoghegan