On Wed, Mar 4, 2020 at 6:15 AM Michael Paquier <michael@paquier.xyz> wrote:
>
> On Tue, Mar 03, 2020 at 06:25:51PM +0900, Michael Paquier wrote:
> > Or actually, a more simple solution is to abuse of the two existing
> > routines so as the dependency switch is done the other way around,
> > from the new index to the old one. That would visibly work because
> > there is no CCI between each scan, and that's faster because the scan
> > of pg_depend is done only on the entries in need of an update. I'll
> > look at that again tomorrow, it is late here and I may be missing
> > something obvious.
>
> It was a good inspiration. I have been torturing this patch today and
> played with it by injecting elog(ERROR) calls in the middle of reindex
> concurrently for all the phases, and checked manually the handling of
> entries in pg_depend for the new and old indexes, and these correctly
> map. So this is taking care of your problem. Attached is an updated
> patch with an updated comment about the dependency of this code with
> CCIs. I'd like to go fix this issue first.
Thanks for the patch! I started to look at it during the weekend, but
I got interrupted and unfortunately didn't had time to look at it
since.
The fix looks good to me. I also tried multiple failure scenario and
it's unsurprisingly working just fine. Should we add some regression
tests for that? I guess most of it could be borrowed from the patch
to fix the toast index issue I sent last week.