Re: buffer assertion tripping under repeat pgbench load

On 24 December 2012 15:57, Greg Smith <greg(at)2ndquadrant(dot)com> wrote:

> 2012-12-24 06:08:46 EST [26015]: WARNING: refcount of base/16384/49169 is
> 1073741824 should be 0, globally: 0
>
> That is pgbench_accounts_pkey. 1073741824 =
> 0100 0000 0000 0000 0000 0000 0000 0000 = 2^30
>
> Pretty odd value to find in a PrivateRefCount. What makes me nervous about
> all of the PrivateRefCount coding is how it switches between references like
> PrivateRefCount[(bufnum) - 1] and PrivateRefCount[b]. Might this be an off
> by one error in one of those, where the wrong form was used?

PrivateRefCount is also one of the only large chunks of memory in a
running server where single bit errors aren't tolerated.

In shared buffers, most SLRUs or various other places, single bit
errors would often go unnoticed.

I wonder if you're having a hardware problem?,

If not, it might be worth looking to see if that block is the root,
maybe there's a problem in fastroot code, or something that only kicks
in with larger indexes.

--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services