<TL;DR>
I am tuning an NFSv4-only file server based on the current Debian 10 Buster and the kernel ntfsd; systemd v241. The nfs-kernel-server package's systemd scripting from the distribution strikes me as a little bit odd. A few service definition files, including the nfs-server.service itself, come with the setting DefaultDependencies=no, such that the unit does not automatically get the Conflicts=shutdown.target dependency, according to systemd.service(5):
[With
DefaultDependencies=yes] [s]ervice units will have dependencies [...] of typeConflicts=andBefore=onshutdown.target. These ensure that normal service units [...] are terminated cleanly prior to system shutdown.
Neither of these is provided explicitly, unlike what I saw in other, systemd's own packages. The command
systemctl show nfs-server.service | egrep '^(Want|Requ|Bind|Bound|Before|After|Confl)'
confirms that this is in fact true: there are no such dependencies. The manual continues,
Only services involved with early boot or late system shutdown should disable this option.
which the NFS server exactly isn't, as it cannot begin serving until the network is fully up, and should probably stop accepting new requests and lameduck from under load as soon as the system shutdown begins.
This is not a single service in the package that has a similar setup, but this one is worrying me the most. I'm rolling out single-purpose VMs in a cloud setup, and a file server may have an ungodly amount of RAM (64-128G), all stuffed to the neck with the filesystem cache, as htop(1) shows. And since this is a file storage machine, I am at a loss for words to express how very much I want the server to, quoth the manual, “terminate cleanly prior to system shutdown,” especially given that I trade a bit of reliability for performance with the exported filesystem's ext4 mount options data=writeback and nobarrier¹.
</TL;DR>
So my question distilled down to one sentence is:
What happens to a systemd service that does not have the Conflicts= and Before= dependencies on the shutdown.target when the system in fact shuts down?
¹ Это хорошо продуманный инженерный компромисс, оцененный с учетом соглашения об уровне обслуживания (SLA) поставщика облачных услуг и результатов серии тестов производительности, и он совершенно не имеет отношения к сути вопроса.