I wrote:
> What I'm thinking of doing is inventing a "soft_stop" variant of
> Cluster::stop that won't complain if pg_ctl stop fails, and then
> having 017_shm's poll_start() call that before retrying the start
> call.
I did that, modeling the details on Cluster::start's fail_ok option.
> It seems like it'd also be a good idea if the END block used
> soft_stop instead of regular stop.
I still like this idea, but I left it alone for the moment,
figuring I've used up my quota of buildfarm destabilization
for now.
regards, tom lane