> As a whole, this has to be backpatched, because we are just ignoring
> recovery if the REDO record is simply missing while the checkpoint
> record is found. For the back branches, the PANIC is actually what
> I'm planning to go with, to match with what is happening when the
> checkpoint record is missing. On HEAD, let's use a softer FATAL to
> give a way to test this driver error moving forward. There is a
> secondary argument for softening the PANIC when the checkpoint record
> is missing to a FATAL, but let's discuss that separately. Hence that
> would make two patches:
> - Something simpler than the attached, without the test with a PANIC
> for the redo record missing, for all the branches.
> - A second patch lowering this PANIC to a FATAL, with the test
> included, only for HEAD.
>
> And done all of that, with the test added to HEAD and a backpatch down
> to v14 for the main fix.
Thanks for fixing the test code. The plan for what goes into the back
branches and what goes into HEAD makes sense. Thanks for committing
it.
> There is a
> secondary argument for softening the PANIC when the checkpoint record
> is missing to a FATAL, but let's discuss that separately.
I was planning to start a separate thread for this point, but since it
was a small change I had included it here earlier. I understand the
considerations involved even for these minor adjustments. I will start
a separate thread for this.
Best Regards,
Nitin Jadhav
Azure Database for PostgreSQL
Microsoft
On Tue, Dec 16, 2025 at 11:35 AM Michael Paquier <michael@paquier.xyz> wrote:
>
> On Mon, Dec 15, 2025 at 05:48:29PM +0900, Michael Paquier wrote:
> > I have done a couple of tests in the CI and locally, and that was
> > looking stable. Attached is the result of what would happen on HEAD,
> > where the change in xlogrecovery.c would include the back-branch
> > versions.
> >
> > Thoughts or comments are welcome.
>
> And done all of that, with the test added to HEAD and a backpatch down
> to v14 for the main fix.
> --
> Michael
