Re: Postgres restart in the middle of exclusive backup and the presence of backup_label file

On 03/01/22 12:32, Nathan Bossart wrote:
> On Tue, Mar 01, 2022 at 11:09:13AM -0500, Chapman Flack wrote:
>> That way, at least, there would be a period of time where procedures
>> that currently work (by passing exclusive => false) would continue to work,
>> and could be adapted as time permits by removing that argument, with no
>> behavioral change.
>
> I'm not sure if there's any advantage to kicking the can down the road. At
> some point, we'll need to break existing backup scripts. Will we be more
> prepared to do that in v17 than we are now?

Yes, if we have provided a transition period the way we did in 7117685.
The way the on-ramp to that transition worked:

- Initially, your procedures used exclusive mode.
- If you changed nothing, they continued to work, with no behavior change.
- You then had ample time to adapt them to non-exclusive mode so now
they work that way.
- You knew a day would come (here it comes) where, if you've never
gotten around to doing that, your unchanged exclusive-mode procedures
are going to break.

So here we are, arrived at that day. If you're still using exclusive mode,
your stuff's going to break; serves you right. So now limit the cases we
care about to people who made use of the time they were given to change
their procedures to use exclusive => false. So the off-ramp can look like:

- Initially, your procedures pass exclusive => false.
- If you change nothing, they should continue to work, with no
behavior change.
- You should then have ample time to change to a new spelling without
exclusive => false, and have them work that way.
- You know some later day is coming where, if you've never
gotten around to doing that, they're going to break.

Then, when that day comes, if you're still passing exclusive at all,
your stuff's going to break; serves you right. If you have made use of
the time you were given for the changes, you'll be fine. So yes, at that
point, I think we can do it with clear conscience. We'll have made the
off-ramp as smooth and navigable as the on-ramp was.

> We could maintain two sets of
> functions for a few releases and make it really clear in the documentation
> that pg_start/stop_backup() are going to be removed soon (and always emit a
> WARNING when they are used). Would that address your concerns?

That would. I had been thinking of not changing the names, and just making
the parameter go away. But I wasn't thinking of your concern here:

> What we need to do is make sure that an older installation won't silently
> work in a broken way, i.e. if we remove the exclusive flag somebody
> expecting the pre-9.6 behavior might not receive an error and think
> everything is OK. That would not be good.

So I'm ok with changing the names. Then step 3 of the off-ramp
would just be to call the functions by the new names, as well as to drop
the exclusive => false.

The thing I'd want to avoid is just, after the trouble that was taken
to make the on-ramp navigable, making the off-ramp be a cliff.

Regards,
-Chap