Re: [CORE] [BUGS] BUG #13350: blindly fsyncing data dir considered harmful

* Magnus Hagander (magnus(at)hagander(dot)net) wrote:
> On May 25, 2015 5:12 PM, "Stephen Frost" <sfrost(at)snowman(dot)net> wrote:
> > This was back-patched all the way and released with the latest round of
> > minor releases, and given that it means crash recovery fails for a large
> > number of deployed systems, I think we need to fix (or revert) the
> > recursive fsync change (d8ac77ab178ddb2ae043b8c463cd30c031e793d0 and
> > related) and do new releases very shortly.
>
> Agreed, this is a pretty bad regression and we need to at least do
> something and out out a release asap - either revert or if we can find a
> better way (see the other thread about this issue for some other ideas).
>
> It happens to be the default shipment on Debian and Ubuntu but it's
> definitely not a platform specific problem I believe, so we should put out
> a "real" release and not expect packagers to carry a specific patch for it.

Agreed, there are certainly other reasons why a file might exist which
can't be written to by the postgres user, we really can't have crash
recovery fail because of it.

Further, I believe that a lot of the .deb-based distributions use
the same technique of using symlinks (Ubuntu included, but even those
who aren't downstream of Debian), and they would all have to be updated
with such a patch. Sadly, suggesting to stay on a prior version (eg:
9.4.1) really isn't acceptable either, given the corruption risks
which 9.4.2 addressed.

Thanks!

Stephen