On Tue, 2018-04-17 at 15:09 -0400, Tom Lane wrote:
> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:
> > Andres was working on a radix tree structure to fix this problem, but
> > that seems to be abandoned now, and it seems a major undertaking. While
> > I agree that the proposed solution is a wart, it seems much better than
> > no solution at all. Can we consider Fujii's proposal as a temporary
> > measure until we fix shared buffers? I'm +1 on it myself.
>
> Once we've introduced a user-visible reloption it's going to be
> practically impossible to get rid of it, so I'm -1. I'd much rather
> see somebody put some effort into the radix-tree idea than introduce
> a kluge that we'll be stuck with, and that doesn't even provide a
> good user experience. Disabling vacuum truncation is *not* something
> that I think we should recommend.
This new option would not only mitigate the long shared_buffers scan,
it would also get rid of the replication conflict caused by the
AccessExclusiveLock taken during truncation, which is discussed in
https://www.postgresql.org/message-id/c9374921e50a5e8fb1ecf04eb8c6ebc3%40postgrespro.ru
and seems to be a more difficult problem than anticipated.
Could that tip the scales in favor of this stop-gap?
FWIW, I have always considered heap truncation on VACUUM to be something
strange anyway. VACUUM does not get rid of empty pages in the middle of
a relation, so why is it so important to do it at the end of the relation?
If the answer is "just because we can do it easily", then I think it would be
ok to disable the feature in cases where it causes problems.
Yours,
Laurenz Albe