On 05/08/10 21:28, Alvaro Herrera wrote:
> Excerpts from Tom Lane's message of jue ago 05 14:01:15 -0400 2010:
>> Maybe write-the-buffers-first is a sufficient longterm solution.
>
> Yeah, perhaps it is, though it's a pity that a single platform problem
> is going to slow down everyone else.
How about performing the truncate as usual, but if it fails, fill the
truncated portion with zeros instead? Zeroed pages should be handled
gracefully, and this would be a very non-invasive fix. Now if the
write() to zero the pages fails too, I guess we'll have to panic, but
that's not much different from flushing the dirty pages first.
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com