Re: Re: In-core regression tests for replication, cascading, archiving, PITR, etc.

On Mon, Dec 07, 2015 at 02:34:39PM +0900, Michael Paquier wrote:
> On Mon, Dec 7, 2015 at 12:06 PM, Noah Misch <noah@leadboat.com> wrote:
> > On Wed, Dec 02, 2015 at 06:59:09PM -0300, Alvaro Herrera wrote:
> >> +sub DESTROY
> >> +{
> >> +     my $self = shift;
> >> +     return if not defined $self->{_pid};
> >> +     print "# signalling QUIT to $self->{_pid}\n";
> >> +     kill 'QUIT', $self->{_pid};
> >
> > On Windows, that kill() is the wrong thing.  I suspect "pg_ctl kill" will be
> > the right thing, but that warrants specific testing.
> 
> I don't directly see any limitation with the use of kill on Windows..
> http://perldoc.perl.org/functions/kill.html
> But indeed using directly pg_ctl kill seems like a better fit for the
> PG infrastructure.

From http://perldoc.perl.org/perlwin32.html, "Using signals under this port
should currently be considered unsupported."  Windows applications cannot
handle SIGQUIT: https://msdn.microsoft.com/en-us/library/xdkz3x12.aspx.  The
PostgreSQL backend does not generate or expect Windows signals; see its
signal.c emulation facility.

> > Postmaster log file names became less informative.  Before the commit:
> > Should nodes get a name, so we instead see master_57834.log?
> 
> I am not sure that this is necessary.

In general, you'd need to cross-reference the main log file to determine which
postmaster log corresponds to which action within the test.  I did plenty of
"grep $PATTERN src/bin/pg_rewind/tmp_check/log/master.log" while debugging
that test.  I'd like to be able to use /*master*.log, not rely on timestamps
or on scraping regress_log_002_databases to determine which logs are master
logs.  Feel free to skip this point if I'm the only one minding, though.

> There is definitely a limitation
> regarding the fact that log files of nodes get overwritten after each
> test, hence I would tend with just appending the test name in front of
> the node_* files similarly to the other files.

They got appended, not overwritten.  I like how you changed that to not
happen, but it doesn't address what I was reporting.

nm