Re: Flaky 003_start_stop.pl test

Hi,

On 2025-12-17 16:46:25 +0200, Heikki Linnakangas wrote:
> > The pattern is similar in other failed tests. There has been a spate of errors
> > on postgres/postgres' CI, which made me look at this again.
>
> I see what's going on here. Before the loop that opens all the connections
> with the SSLRequests, the test does this:
>
> > if (!$node->raw_connect_works())
> > {
> > plan skip_all => "this test requires working raw_connect()";
> > }
>
> That's the connection you see in the log without the SSLRequest.
>
> When the test fails, what happens is that the backend handling that
> connection is slow. It lingers around, even though the perl client has
> already closed the socket. The backend finally exits _just_ when the test
> has opened all the "dead end" connections, and checks that the final
> connection fails with the "FATAL: sorry, too many clients already" error.
> Because the backend exits at just the right moment, it frees up a connection
> slot which is then used by that final connection. The test fails, because it
> expects all the connection slots to be in use and for that final connection
> to launch a dead-end backend.
>
> The 002_connection_limits.pl test has potential for a similar issue. I added
> a "sleep(1)" to proc_exit to simulate slow backend exit, and the test
> started failing, because the "safe_psql" calls used in the initialization
> was still consuming a connection slot.
>
> I've pushed a fix to both tests, by restarting the server after the
> initialization steps. That's a little heavy-weight, but it's a simple way to
> ensure that all the backends are gone.

Thanks!

Let's hope that this makes cfbot run a bit more quietly aga<oh no>. The test
run for this commit failed with an independent flakiness... [4]

Greetings,

Andres Freund

[4] https://cirrus-ci.com/task/6281872222715904