Re: Timing-sensitive case in src/test/recovery TAP tests

Michael Paquier <michael(dot)paquier(at)gmail(dot)com> writes:
> On Tue, Jun 27, 2017 at 3:44 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> Looks good as far as it goes, but I wonder whether any of the other
>> get_slot_xmins calls need polling too. Don't feel a need to add such
>> calls until someone exhibits a failure there, but I won't be very
>> surprised if someone does.

And behold, we have here
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=nightjar&dt=2017-08-08%2020%3A54%3A09

# Failed test 'xmin of cascaded slot non-null with hs feedback'
# at t/001_stream_rep.pl line 224.
# got: ''
# expected: anything else

That's one of only four calls lacking a preceding wait_slot_xmins call.
The ones at lines 173 and 177 seem safe because nothing has happened yet.
I think the one at line 300 should be safe because the standby_2 server is
shut down at that point, but is there any way that the standby_1's view
hasn't updated by the time that happens?

> I got the same thought, wondering as well if get_slot_xmins should be
> renamed check_slot_xmins with the is() tests moved inside it as well.
> Not sure if that's worth the API ugliness though.

Mmm, doesn't seem like that's worth doing, but I'm half tempted to merge
wait_slot_xmins into get_slot_xmins so you can't skip it ...

regards, tom lane