[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Debian Buster release to partially drop non-systemd support

Philipp Kern <pkern@debian.org> writes:
> On 17.10.2018 06:52, Russ Allbery wrote:

>> I think a package of a daemon that does not inherently require any
>> systemd-specific features and would work straightforwardly with
>> sysvinit, but has only a systemd unit file and no init script, is not
>> only buggy but RC-buggy.  That's what Policy currently says.

> And it would not be buggy if it does not ship any init integration or
> relies on a non-init service supervision system like runit. (The crux
> being that systemd is not modular to be run that way and not portable.)

No, I think that's also buggy.  If it's a daemon that should be started at
boot and it doesn't include init integration, I think that's an obvious
bug.  It's very hard to write a Policy rule that says this, since it's
more of a common sense sort of bug.

> You say "more than adequate". I don't particularly see it as providing a
> solid system as you don't get restart on failure. Now I can see how
> people say that this is not a problem as daemons should not crash in the
> first place. Maybe I just value reliability of service more highly than
> being woken up at night being told that my service is unreliable.

My point is that sysvinit is a non-default configuration and it's
reasonable to expect that it may be missing some features or robustness.
If you have the time and resources to put into equaling the robustness
that you would get under systemd, that's great, but sysvinit is much more
of a fire-and-forget system and is known to in general not have that
robustness property.  sysvinit users who care will run something like
monit that watches health externally and takes appropriate action.

I think every packager owes it to the social fabric of the project to make
the effort to provide a basic init script, in the same way that we do
basic porting to other architectures and investigate build failures.
There is some point, which is subjective, beyond which I think it's
reasonable to ask someone who cares about sysvinit to do more complicated
work, but I think a basic init script tested under systemd's init support
is a reasonable request.

My personal concern is more about the social community of the project than
about the technical details.  I consider providing an init script even if
I don't personally use sysvinit to be extending the hand of community and
solidarity to other Debian community members who use it.  To say to them
that their concerns have been heard and supported, even if I don't agree
with their concerns.  Personally, I find that extremely important, a
principle that's as important as the technical quality bar we try to reach
in our packages.

> Is a possible answer here to ship an init script and then provide
> additional supervision options using override files - to enhance
> whatever is provided by the sysv generator?

I personally would tend to maintain separate init and systemd unit files
and only lightly test the init script unless I knew there was something
tricky going on, but this answer varies based on the complexity of the

> This statement also does not address a daemon that only runs through
> socket activation - passing an fd to the daemon. But I don't have any
> example handy and this might be a strawman argument. Except that I could
> see that for simplicity newer daemons might be written in this way as it
> does cut a significant amount of socket handling code.

Yes, at the point where we have an upstream daemon that was written
explicitly for use with systemd in that fashion, I think that turns into a
different sort of question.  This is now more akin to porting, and that's
a different bar.  I don't think we can place hard requirements on what a
maintainer chooses to do here, other than asking for them to take patches
from porters if those are offered.

> To some degree I regret that we cannot provide a fully integrated
> distribution by mandating that the core layers (be it kernels or init
> systems) can be switched out. systemd still supports init scripts but on
> the other side there's pretty much complete stagnation with the onus on
> the systemd maintainers to keep things working. There could as well have
> been an effort to support a subset of the unit language for sysvinit.

I do completely agree with the general thrust of this thread: sysvinit
needs an active maintenance team, and without that work, it will
eventually die.  There's a limit to how much people who don't use it
should feel obligated to keep it working.  Hopefully those who do use it
will be able to find the resources to support and improve it (and indeed
being able to support the unit *syntax* would be a huge step towards
making sysvinit support in Debian more robust).

> I suppose one answer would be a cron generator. Given that policy
> specifies naming schemes for /etc/cron.{hourly,daily,weekly,monthly,d},
> there could probably be a mapping from filename to timer. But the cron
> language itself does not contain identifiers, so there's still the
> question what to do if you encounter multiple lines and how you'd map
> them. Init scripts with their 1:1 mapping and metadata headers were much
> easier to handle. Plus we'd need a way to tell cron not to execute those
> cronjobs in case systemd is running, which I guess means adding systemd
> guards to /etc/crontab (a conffile). And you'd of course still have cron
> wake up to execute the shell statement.

> But it's not like timer units are forbidden. Just like my
> introductionary statement of "but if you use a different system not
> considered an init system, you are fine", there's nothing in policy
> mandating periodic jobs to work in a particular way. It just talks about
> what to do if you do ship a cronjob.

Yes, Policy is behind the curve here.

Timer units are also a more complicated problem since they're not a
superset of cron behavior.  They do some things better than cron jobs;
they do other things much *worse* than cron jobs.  I have cron jobs that I
wanted to convert to timer units and discovered I couldn't because timers
simply wouldn't work for what I was trying to do.

But it does feel like there's a common subset of functionality that we
could embrace, and unblock people using timer units for the things they're
good at without degrading functionality for people not using systemd.

Russ Allbery (rra@debian.org)               <http://www.eyrie.org/~eagle/>

Reply to: