Restarting services - minimize downtime

[danbst]

Issue description

During recent nixos-rebuild switch I got the following log

...
updating GRUB 2 menu...
stopping the following units: -.slice, alsa-store.service, audit.service, ddclient.service, kmod-static-nodes.service, network-addresses-eth0.service, network-link-eth0.service, network-local-commands.service, network-setup.service, nix-daemon.service, nix-daemon.socket, nscd.service, sshd-keygen.service, systemd-modules-load.service, systemd-sysctl.service, systemd-timesyncd.service, systemd-tmpfiles-clean.timer, systemd-tmpfiles-setup-dev.service, systemd-udev-trigger.service, systemd-udevd-control.socket, systemd-udevd-kernel.socket, systemd-udevd.service, systemd-vconsole-setup.service
Failed to stop -.slice: Operation refused, unit -.slice may be requested by dependency only.
See system logs and 'systemctl status -.slice' for details.
<... HERE A PAUSE FOR ONE MINUTE ...>
NOT restarting the following changed units: getty@tty1.service, systemd-journal-flush.service, systemd-random-seed.service, systemd-remount-fs.service, systemd-tmpfiles-setup.service, systemd-update-utmp.service, systemd-user-sessions.service, user@1001.service
activating the configuration...
setting up /etc...
restarting systemd...
reloading the following units: dbus.service, dev-hugepages.mount, dev-mqueue.mount, firewall.service, proc-sys-fs-binfmt_misc.mount, sys-kernel-debug.mount
restarting the following units: dhcpcd.service, sshd.service, systemd-journald.service, systemd-logind.service
starting the following units: alsa-store.service, audit.service, ddclient.service, kmod-static-nodes.service, network-addresses-eth0.service, network-link-eth0.service, network-local-commands.service, network-setup.service, nix-daemon.socket, nscd.service, sshd-keygen.service, systemd-modules-load.service, systemd-sysctl.service, systemd-timesyncd.service, systemd-tmpfiles-clean.timer, systemd-tmpfiles-setup-dev.service, systemd-udev-trigger.service, systemd-udevd-control.socket, systemd-udevd-kernel.socket, systemd-vconsole-setup.service
the following new units were started: systemd-journal-catalog-update.service

Looks like Nixos restarting logic has two separate algorithms:

1. stop in parallel, then start in parallel
2. restart in paralell

Why does the first algorithm even exist? In one of my deployments it resulted in nginx downtime for several minutes. (Speaking of nginx, there is #1988)

How can I change the default behavior of a service to second algorithm?

Technical details

• System: 17.03pre101636.183eeb3

[fpletz]

You can set stopIfChanged to false in the systemd service config to get the second behavior.

[danbst]

thanks