Thanks [LWN.net]

Thanks

Posted Nov 21, 2014 18:48 UTC (Fri) by tjc (guest, #137)
In reply to: Thanks by anselm
Parent article: Today's Debian technical committee resignation: Ian Jackson

> The problem with Upstart is that it has severe shortcomings that – at least according to its original lead developer – would require a major redesign to overcome.

What are the shortcomings? I've been using Upstart for several years without incident, but I don't require much of an init system (other than to "init").

to post comments

Thanks

Posted Nov 21, 2014 19:31 UTC (Fri) by mjg59 (subscriber, #23239) [Link] (6 responses)

https://bugs.launchpad.net/ubuntu/+source/mountall/+bug/4... is an example.

Thanks

Posted Nov 22, 2014 16:10 UTC (Sat) by smurf (subscriber, #17840) [Link] (5 responses)

To summarize: While an event-based init system seemed like a good idea at the time, upstart quickly ran into a whole lot of corner cases which are basically unsolvable without adding job state as separate concepts. Upstart has tried working around this by adding even more events (like "starting to start" and "finished starting") but they don't help in general.

Given this fact, a dependency-based system which uses events just to change the units' state simplifies the whole system and affords easier debugging, as it's statically introspectable – meaning you don't need an event history to figure out what's wrong, just the current job states and the dependency graph.

Thanks

Posted Nov 22, 2014 21:14 UTC (Sat) by mgb (guest, #3226) [Link] (4 responses)

Forgy invented http://en.wikipedia.org/wiki/Rete_algorithm in the 70's and it was widely adopted in the early 80's.

Sad to see the wheel painfully being reinvented again.

Thanks

Posted Nov 22, 2014 21:46 UTC (Sat) by smurf (subscriber, #17840) [Link] (1 responses)

Dependency resolution in an init system doesn't need anything remotely this elaborate. We don't even need alternates (what for?).

Thanks

Posted Nov 22, 2014 22:08 UTC (Sat) by mgb (guest, #3226) [Link]

> Dependency resolution in an init system doesn't need anything remotely this elaborate.

Rete is not elaborate. It is a clean elegant modular algorithm.

> We don't even need alternates (what for?).

Alternate networks. Sometimes alternate data stores. Alternate entropy sources - you want sshd ASAP but maybe not until you have some entropy. "If I can't get any entropy bring up sshd only on the private LAN and if that doesn't work fall back to telnetd on the private LAN."

As the "new" init systems try to invade serious servers they will keep on running into problems and adding more keywords and reliving the 70's until they eventually relearn the lessons of the early 80's.

Thanks

Posted Nov 22, 2014 23:33 UTC (Sat) by mjg59 (subscriber, #23239) [Link] (1 responses)

How would you apply this to an event-based init system?

Thanks

Posted Nov 23, 2014 0:15 UTC (Sun) by mgb (guest, #3226) [Link]

> How would you apply this to an event-based init system?

Your own http://lwn.net/Articles/622742/ answers that.

If you want functionality beyond inittab, an event-based init system is even worse than an ad-hoc dependency-based init system that has to keep adding keywords like RequiresMountsFor.

Thanks

Posted Nov 21, 2014 20:14 UTC (Fri) by Felix (subscriber, #36445) [Link] (2 responses)

jspaleta also compiled an impressive list https://lwn.net/Articles/582585/

Thanks

Posted Nov 21, 2014 21:02 UTC (Fri) by jspaleta (subscriber, #50639) [Link] (1 responses)

One request...

If someone does fork upstart and fix all the deep deep design problems. Please rename it as a new project. What you end up building will be a very different animal.

I have a couple of humble name suggestions:
upyours, upset, upchuck

-jef

Thanks

Posted Nov 22, 2014 15:23 UTC (Sat) by zuki (subscriber, #41808) [Link]

Funny... but not quite funny enough. Please give it a rest.