On Thu, Aug 5, 2010 at 7:28 PM, Alvaro Herrera
<alvherre@commandprompt.com> wrote:
>
> The scope is further reduced by the fact that this only seems to happen
> on Windows, and then only when the antivirus is messing around with the
> files.
So I suspect this could be triggered lots of ways. Imagine a ZFS
volume that runs out of space temporarily. Even truncate would need
additional blocks to replace the meta information. Or a network
filesystem like AFS where your kerberos tickets have expired and you
get a permission failure until they've been renewed. Or, in the case
of a very large table being truncated I suspect there's a
CHECK_FOR_INTERRUPTS lying around that can cancel the backend at the
wrong time.
It would be nice to have a regression test harness that caused system
calls to fail randomly -- the difficult part would be testing the
results.
--
greg
