Re: Corruption during WAL replay

At Mon, 17 Aug 2020 11:22:15 -0700, Andres Freund <andres(at)anarazel(dot)de> wrote in
> Hi,
>
> On 2020-08-17 14:05:37 +0300, Heikki Linnakangas wrote:
> > On 14/04/2020 22:04, Teja Mupparti wrote:
> > > Thanks Kyotaro and Masahiko for the feedback. I think there is a
> > > consensus on the critical-section around truncate,
> >
> > +1
>
> I'm inclined to think that we should do that independent of the far more
> complicated fix for other related issues.
...
> > Perhaps a better approach would be to prevent the checkpoint from
> > completing, until all in-progress truncations have completed. We have a
> > mechanism to wait out in-progress commits at the beginning of a checkpoint,
> > right after the redo point has been established. See comments around the
> > GetVirtualXIDsDelayingChkpt() function call in CreateCheckPoint(). We could
> > have a similar mechanism to wait out the truncations before *completing* a
> > checkpoint.
>
> What I outlined earlier *is* essentially a way to do so, by preventing
> checkpointing from finishing the buffer scan while a dangerous state
> exists.

Seems reasonable. The attached does that. It actually works for the
initial case.

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center