On Mon, Jan 18, 2016 at 7:08 PM, Andres Freund <
andres@anarazel.de> wrote:
>
> On 2015-12-21 16:26:25 +0900, Michael Paquier wrote:
> > On Sun, Dec 20, 2015 at 10:14 PM, Michael Paquier
> > <
michael.paquier@gmail.com> wrote:
> > > Speaking of which, this patch was registered in this CF, I am moving
> > > it to the next as a bug fix.
> >
> > I found a stupid bug in my previous patch: when issuing XLOG_SWITCH it
> > is possible that the return LSN pointer is on the new segment that has
> > been forcibly archived if RequestXLogSwitch is called multiple times
> > and that subsequent calls are not necessary. Patch updated.
>
> I find this patch rather unsatisfactory. Yes, it kinda solves the
> problem of archive timeout, but it leaves the bigger and longer standing
> problems of unneccessary checkpoints with wal_level=hs in place. It's
> also somewhat fragile in my opinion.
>
> I think we rather want a per backend (or perhaps per wal insertion lock)
> flag that says 'last relevant record inserted at', and a way to not set
> that during insertion. Then during a checkpoint or the relevant bgwriter
> code, we look wether anything relevant happened in any backend since the
> last time we performed a checkpoint/logged a running xacts snapshot.
>
Sounds to be a more robust way of dealing with this problem. Michael,
if you don't disagree with above proposal, then we can mark this patch
as Waiting on Author?