Re: Cleaning up historical portability baggage

Hi,

On Tue, Jun 10, 2025 at 12:59:16PM +0200, Michael Banck wrote:
> On Tue, Jun 10, 2025 at 09:05:03AM +1200, Thomas Munro wrote:
> > https://savannah.gnu.org/task/?7050
>
> 3. make check nowadays succeeds the subscription test that it failed in
> the log mentioned in the above task, but it still fails the stats test:

I take that back, the subscription test is flakey, but less so than the
stats test as it only has one of those stats_reset comparisons, compared
to several for stats. So I am seeing failures there only every few test
runs.

The reason those tests fail is that the Debian 12 era [1] GNU Mach I am
currently using has 10ms timer resolution, so some of those stats_reset
calls in quick succession do not lead to different timestamps and the
checks fail.

There are several "SELECT pg_sleep(0.1);" calls in stats.sql already, so
I wonder whether a few more would be acceptable, or at least using
pg_sleep(0.01)? Using the attached, I no longer see regression failures,
even after running it for 10+ times.

The good news is maybe two-fold:

1. GNU Mach recently got better timer resolution and the version in
Debian unstable got that backported [2], so going forward this should
not be a problem anymore. I'll try upgrading my VM and see whether that
helps without the patch.

2. The fact that nobody else complained about those new(er) timestamp-
comparison additions appears to imply that there are no 100ms resolution
machines we support anymore. So did we consider switching those
pg_sleep(0.1) calls in stats.sql to pg_sleep(0.01) to save a bit of
time?

Michael

[1] As hurd-i386 is not a release architecture, there is no Debian 12
release for it. But the Debian GNU/Hurd porters made their own release
based on a snapshot of the hurd-i386 archive at time of the bookworm
release.
[2] https://salsa.debian.org/hurd-team/gnumach/-/blob/master/debian/patches/git-hpet?ref_type=heads