Re: BufFileRead() error signalling

On Thu, May 28, 2020 at 4:16 AM Alvaro Herrera <alvherre(at)2ndquadrant(dot)com> wrote:
> On 2020-Jan-27, Robert Haas wrote:
> > OK, now that I've waxed eloquent on that topic, let me have a go at
> > your actual questions. Regarding back-patching, I don't mind
> > back-patching error handling patches like this, but I don't feel it's
> > necessary if we have no evidence that data is actually getting
> > corrupted as a result of the problem and the chances of it actually
> > happening seems remote.
>
> I do have evidence of postgres crashes because of a problem that could
> be explained by this bug, so I +1 backpatching this to all supported
> branches.
>
> (The problem I saw is a hash-join spilling data to temp tablespace,
> which fills up but somehow goes undetected, then when reading the data
> back it causes heap_fill_tuple to crash.)

Ooh.

> Thomas, if you're no longer interested in seeing this done, please let
> me know and I can see to it.

My indecision on the back-patching question has been resolved by your
crash report and a search on codesearch.debian.org. BufFileRead() and
BufFileWrite() aren't referenced by any of the extensions they
package, so by that standard it's OK to change this stuff in back
branches. I'll post a rebased a patch with Robert and Ibrar's changes
for last reviews later today.