VCL reload is currently broken on cp3032:
Nov 2 09:55:25 cp3032 puppet-agent[39299]: (/Stage[main]/Cacheproxy::Instance_pair/Varnish::Instance[text-backend]/Exec[retry-load-new-vcl-file]/returns) Starting VCC-compiler: fork() failed: Resource temporarily unavailableVCL compilation failed
This is due to the unit reaching TaskMax:
10:00:21 ema@cp3032.esams.wmnet:~ $ sudo systemctl status varnish | grep Tasks Tasks: 7372 (limit: 7372)
Interestingly:
- all other varnish backends have a number of tasks significantly lower than cp3032's, and in any case below 1000. For example, varnish.service on cp3033 currently has 719 task
- on cp[1075-1090].eqiad.wmnet TaskMax is 41779 instead of 7372, the limit on all other hosts. This is due to DefaultTaskMask being set to 15% of kernel.pid_max, which is 278528 on cp[1075-1090] and 49152 on the others
- DefaultTaskMask was originally set to 512 by systemd upstream, then changed to 15% of kernel.pid_max: https://github.com/systemd/systemd/issues/3211
- systemd 232 in stretch ships debian/patches/debian/Revert-core-enable-TasksMax-for-all-services-by-default-a.patch which claims to be reverting https://github.com/systemd/systemd/commit/9ded9cd14 but is in reality just changing systemd-system.conf(5) and a commented value in system.conf (DefaultTasksMax)