Impossible to remove all offload machines

Impossible to remove all offload machines

[Ian Eure]

Ran into this issue last week.  If you:

- Configure some offload build machines in your operating-system
  configuration.
- Reconfigure your system.
- Remove all offload build machines.
- Reconfigure your system again.

...then various guix operations will still try to connect to 
offload machines, even if you reboot the affected client.

This is caused by a bug in the `guix-activation' procedure:

   ;; ... and /etc/guix/machines.scm.
   #$(if (null? (guix-configuration-build-machines config))
         #~#f
         (guix-machines-files-installation
          #~(list #$@(guix-configuration-build-machines
           config))))

If there are no build machines defined in the configuration, no
operation is performed (#f is returned), which leaves the previous
generation’s /etc/guix/machines.scm in place.

The same issue appears to affect channels:

   ;; ... and /etc/guix/channels.scm...
   #$(and channels (install-channels-file channels))

I’d be happy to take a stab at fixing this, but I’m not certain 
what
direction to go, or how much to refactor to get there. Should the
channels/machines files be removed (ignoring errors if they don’t
exist)?  Should empty files be installed?  Should that happen 
inline
in `guix-activation', or in another procedure? Should the 
filenames be
extracted to %variables to avoid duplicating between the two 
places
they’ll be used?

If someone would like to provide answered, I would contribute a 
patch.

Thanks,

 — Ian

bug#72686: Impossible to remove all offload machines

[Ian Eure]

Tomas Volf <~@wolfsden.cz> writes:

> [[PGP Signed Part:Undecided]]
>
> Hello,
>
> Ian Eure <ian@retrospec.tv> writes:
>
>> Disregard this, I continued thinking after sending the email 
>> (as one does) and
>> realized that any managed file will be a link into the store -- 
>> so if the system
>> is reconfigured with no build-machines or channels *and* the 
>> corresponding file
>> is a store link, it should be removed; otherwise, it should 
>> remain untouched. I
>> can work with this.
>
> Will this correctly handle cases where user is managing the file 
> using
> for example extra-special-file?
>

No, it wouldn’t.


> I wonder whether fat-warning approach would not be better.
>

I think I agree.

  — Ian

bug#72686: Impossible to remove all offload machines

[Maxim Cournoyer]

Hi Ian,

Ian Eure <ian@retrospec.tv> writes:

[...]

> Apologies for the silence, life stuff has been eating most of my free
> time, but I have a bit of bandwidth to spend on this problem again.

As you can see, you are not alone :-).

> I took a swing at this, it wasn’t as difficult as I expected. While
> this approach gives a smooth upgrade path for those who’ve configured
> channels in a stateful way switching to declarative configuration,
> it’s possibly bumpy for those already using a declarative config.  If
> a machine with declarative channels is reconfigured, the channels will
> be duplicated from /etc/guix/channels.scm to
> /etc/guix/channels-declarative.scm. Using `delete-duplicates' on the
> merged channels should avoid major problems, but I think it still
> needs a loud entry in news and manual action (deleting
> /etc/guix/channels.scm) to upgrade. Given that both approaches will
> require manual action, I’m a bit inclined to go with the simpler, and
> take over the existing file. That said, I think the failure mode of
> the simpler approach (stomping on channels a user may have configured)
> is undeniably worse than potentially duplicating channels or
> continuing to pull in old ones unexpectedly.  Do either of you have a
> strong opinion or more information which would help guide this
> decision?

I still like the option to leave a handcrafted channels.scm alone; I
don't think this imperative variant will be obsolete in the future; it's
still useful to experiment quickly, avoiding costly reconfigure just to
change some bits about a new offload machine.

> The root issue at work behind all these problems is that activation
> code only sees the desired target config, rather than the current and
> target configs.  Comparing the current and target configs would allow
> the code to more precisely compute the needd change to move from one
> state to the next.  I think that could be a good change to make,
> though it’s obviously going to be much more involved, and IMO will
> require discussion outside the scope of this specific bug.

The activation is just a script, so if it's useful to check
the current value, it should be OK to do so.

> I have a draft patch series I hope to send up soon, but need to get
> Guix System up in a VM to test first.  It does separate declarative
> channels into their own config, but doesn’t do the same for build
> machines.  While I think many fewer users configure build machines
> than channels, it’s probably a good idea to use the same approach for
> both channels and machines.

Yes, it'd be nice to have a uniform approach.  The acl file could
benefit from the same treatment, I believe (that, and 'guix authorize'
should validate that we're not adding the same key multiple times --
warn something along "key XXXX is already authorized".

-- 
Thanks,
Maxim