Thread: Logs vanish after partial log destination change
Hi, I'm not entirely sure if it's a bug, or what, but it caught me by surprise. Tested it on Linux, on 12.6 and 15devel. I start with: log_destination = 'syslog' logging_collector = false Logs to to syslog, all is well. Then, I chance log_destination to 'csvlog' and logging_collector to 'on'. And then I reload config. I know I can chance log_destination, but I can't change logging_collector. I assumed it means that logging will go to where they did before. But - logs are gone. They are not in syslog. They are not in the new location. When I tried it on test system with 15devel, I noticed by chance that logs were redirected to terminal from which I issued `pg_ctl reload` from. Well, that's hardly useful. I'm not entirely sure what Pg should do in such case, perhaps just restrict changing any log_* if one of changed gucs require restart? But, it caught me by surprise, and figured that it might be problematic for others too. Best regards, depesz
hubert depesz lubaczewski <depesz@depesz.com> writes: > I start with: > log_destination = 'syslog' > logging_collector = false > Logs to to syslog, all is well. > Then, I chance log_destination to 'csvlog' and logging_collector to > 'on'. > And then I reload config. > I know I can chance log_destination, but I can't change > logging_collector. I assumed it means that logging will go to where they > did before. No. You effectively set log_destination to empty, since the csvlog item is ignored if the logging collector isn't active. That's not a case that we should disallow, IMO. I do not think there's anything that we can do here that wouldn't make other use-cases worse. regards, tom lane
On Sun, Sep 05, 2021 at 10:24:35AM -0400, Tom Lane wrote: > > I know I can chance log_destination, but I can't change > > logging_collector. I assumed it means that logging will go to where they > > did before. > No. You effectively set log_destination to empty, since the csvlog > item is ignored if the logging collector isn't active. That's not > a case that we should disallow, IMO. I do not think there's anything > that we can do here that wouldn't make other use-cases worse. Well, OK. Thanks, depesz