On Wed, Dec 1, 2021 at 5:15 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Justin Pryzby <pryzby@telsasoft.com> writes:
> > +1 to document it, but it seems like the worse problem is allowing the admin to
> > write a configuration which causes the server to fail to start, without having
> > issued a warning.
>
> > I think you could fix that with a GUC check hook to emit a warning.
> > ...
>
> Considering the vanishingly small number of actual complaints we've
> seen about this, that sounds ridiculously over-engineered.
> A documentation example should be sufficient.
I don't know if this will tip the scales, but I'd like to lodge a
belated complaint. I've gotten myself in this server-fails-to-start
situation several times (in development, for what it's worth). The
syntax (as Bharath pointed out in the original message) is pretty
picky, there are no guard rails, and if you got there through ALTER
SYSTEM, you can't fix it with ALTER SYSTEM (because the server isn't
up). If you go to fix it manually, you get a scary "Do not edit this
file manually!" warning that you have to know to ignore in this case
(that's if you find the file after you realize what the fairly generic
"FATAL: ... No such file or directory" error in the log is telling
you). Plus you have to get the (different!) quoting syntax right or
cut your losses and delete the change.
I'm over-dramatizing this a bit, but I do think there are a lot of
opportunities to make mistakes here, and this behavior could be more
user-friendly beyond just documentation changes. If a config change is
bogus most likely due to a quoting mistake or a typo, a warning would
be fantastic (i.e., the stat() check Justin suggested). Or maybe the
FATAL log message on a failed startup could include the source of the
problem?
Thanks,
Maciek