On 01/31/2018 09:51 PM, Jan Wieck wrote:
>
>
> On Wed, Jan 31, 2018 at 12:32 PM, Stefan Blanke
> <stefan.blanke@framestore.com <mailto:stefan.blanke@framestore.com>> wrote:
>
> >
> > I'll bet you it's not that. It's quite unlikely that would fail with
> > exactly 1GB request size. It seems much more like a buffer that we keep
> > to be power of 2. The question is which one.
>
> I had dismissed corruption before writing in. It's exactly 1GB every
> time this has happened - and we can dump the full dataset
> periodically without issue.
>
> >> I have my money on a corrupted TOAST entry. Is this happening on
> >> trustworthy hardware or beige box with no ECC or RAID?
>
> It's good quality commercial hardware in our colo - no exactly sure
> what.
>
>
> If it is a sporadic issue and you can dump the full dataset, then I just
> lost my money (Tomas, you coming to PGConf in Jersey City?).
>
Unfortunately no, but I'm sure there will be other opportunities to buy
me a beer ;-) Like pgcon, for example.
>
> But then, if this is a plain COPY to stdout ... I am wondering what
> could possibly be in that path that wants to allocate 1GB. Or is this
> not so plain but rather a COPY ... SELECT ...?
>
That's what I've been guessing, and why I was asking for a query plan.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services