Re:Systemd, ten years later: a historical and technical retr...
by rezonant on Saturday May 16, 2020 @06:42PM
(#23207811)
> Keeping things in separate source databases with explicit interfaces is a well understood way of reducing the chances of ending up with a ball of mud.
I don't think this invalidates the thrust of your point, and I don't necessarily subscribe to the concept of monorepos, but Google is a monorepo, so senior engineers are definitely signing off on this sort of thing. Is it a ball of mud though? Actually does seem true more and more as time goes on.
Probably the biggest difference is that well-disciplined engineers utilizing monorepo do sustain those well-defined interfaces. The example of systemd-logind having no interface guarantees (cited in the linked article) is a sign of trouble.
Actually, it's possible for more than two things to be true at once. It could be:
1. It could be true sysv init sucked. (Actually, it wasn't that bad as a modular version of local.rc, but parallel and delayed startup was always just a kludge).
2. It could be true that systemd is a good init system. (For me it's a step change improvement sysv init for parallel and delayed start up. About my only complaint is you don't see failures on "systemctl start xxx"; you have to run journalctl.)
3. It could be that adding another 1M lines of code into the systemd source unrelated to starting stuff was an unsound engineering decision - almost inexplicable coming as it does from a senior engineer. Keeping things in separate source databases with explicit interfaces is a well understood way of reducing the chances of ending up with a ball of mud.
4. It could be that moving sending state changes via an RPC interface (dbus) so you have a million places you have to inspect to understand the state of the system and how it got there rather than writing the desired state of the system to disk with a journal where everyone can inspect it was a bad idea.