bug#27155: [PATCH 0/2] Support service extensions on the "final" service values

[Alex Kost]

Ludovic Courtès (2017-06-03 23:21 +0200) wrote:

> Ludovic Courtès <address@hidden> skribis:
>
>> This patch adds support for service extensions that modify the
>> "final" values of a service.  This is meant to implement cross-cutting
>> concerns as well as system-wide customization as discussed with Alex
>> long ago:
>>
>>   https://lists.gnu.org/archive/html/guix-devel/2015-11/msg00623.html
>>   https://lists.gnu.org/archive/html/guix-devel/2016-09/msg01505.html
>>
>> To summarize, a "finalization extension" (for lack of a better name)
>> gets the final value of a service and returns a new value for that
>> service.
>
> I found a better name: “customizations”.

I kinda like "finalization" more :-)  But "customization" is fine with
me, not a big deal.

>> For example, for the /etc service, a "normal" extension can only add
>> entries for /etc.  A "finalization" extension can instead inspect and
>> change all the /etc entries.  IOW, it is a sort of a "sudo" for service
>> extensions; it's also quite inelegant compared to the "normal" extension
>> mechanism, but it's certainly useful.
>
> Not liking the “sudo” aspect of this patch, I thought it would be
> natural if service types could control how customizations apply.  That
> way, the PAM or /etc service could still guarantee, for instance, that
> customization does not add or remove entries, and so on.

Ouch, that's what I don't like.  I think a full control is better.
You'll never know what a user might want to do, and giving a user a full
freedom (even to break a system!) would be a great feature.  So I'm
against such guarantees that strict users in modifying their systems.

> In the end, this control by the service type makes it easier to reason
> about what extensions do, whereas the “sudo” style means that an
> extension can alter the service’s value in any possible way.

Right, "any possible way" is exactly what I want!

> So I started modifying this patch set to add a ‘customize’ field to
> <service-type>, next to ‘extend’.  For the PAM and /etc services,
> ‘customize’ would compose and apply procedures that modify an entry, for
> instance.
>
> Then I realized that the only difference between ‘customize’ and
> ‘extend’ would be the meaning attached to it.  IOW, both are some kind
> of an extension.
>
> So at this point, I started wondering whether we should just allow
> service types to declare several extension points.  So for PAM, we’d do:
>
> (define pam-service-addition
>   ;; The extension point to add PAM services.
>   (service-extension-point
>     (compose concatenate)
>     (extend append)))
>
> (define pam-service-cutomization
>   ;; The extension point to customize PAM services.
>   (service-extension-point
>     (compose compose)
>     (extend append)))
>
> (define pam-root-service-type
>   (service-type (name 'pam)
>                 (extensions (list (service-extension etc-service-type
>                                                      /etc-entry)))
>
>                 (extension-points (list pam-service-addtion
>                                         pam-service-customization))))
>
> But then ‘service-extension’ would need to specify not only the target
> service type but also the target extension point, which means more
> boilerplate, etc.

I don't have a deep understanding of services, but your suggestion seems
(to me) to have the following downsides:

- More additional work – to determine (and implement) what aspects of
  services should and what should not be modified by a user.

- Less freedom (comparing to your previous solution) for users in
  modifying services.

> So after so much thought and hacking, I feel like the ad hoc solution at
> <https://lists.gnu.org/archive/html/guix-devel/2016-09/msg01505.html>
> was not that bad after all.

He-he :-)

> Sorry to bother you with philosophical design questions when we already
> have two ways to solve the problem at hand, but I feel like there’s a
> pattern worth looking for!

No problem, looking for patterns is always an interesting occupation!
As for me, I agree with any solution that allows me to replace
"/etc/profile".  But in general, I vote for that solution that allows
users to customize as much things as possible.

-- 
Alex