Re: [PATCH 2/2] centralize multistrap configuration generation

[Jan Kiszka <jan.kiszka@siemens.com>]

On 2018-02-15 12:27, Baurzhan Ismagulov wrote:
> On Thu, Feb 15, 2018 at 11:15:02AM +0100, Jan Kiszka wrote:
>>> That shows to me that upstream might be interested in some of these
>>> improvements we would need to implement.
>>
>> Exactly. We should really use the community-proven, continuously tested
>> official approach of Debian, in that in a way that it foresees for it.
>>
>> Let's bury multistrap. Its a dead horse we rode to long now.
> 
> First off, thanks for cutting the many lines of code not being referred to :) .
> 
> Regarding multistrap, both approaches have advantages and disadvantages.
> Moreover, there are other implementations, such as cdebootstrap.
> 
> 
> Multistrap is currently useful in the following ways:
> 
> 1. Generate the list of packages to install using several apt repos.
> 
> 2. Provide the package paths to download from apt repos.
> 
> 3. Generate the base system from several apt repos.
> 
> Unfortunately, all three seem to be unique among the three candidates.
> 
> 
> (1) and (2) are necessary for reproducibility. My preferred solution would be
> to use an apt library that would do that for us (e.g., extract it from apt,
> extend python-debian, or look whether aptly provides that). Moving to
> *debootstrap would require solving those in a community-unmaintained way for
> some time.
> 
> (3) is useful for people with customizations cutting off significant dependency
> chains from Debian (like SLIND did with Perl dependencies). Moving to
> *debootstrap would mean first installing those, then apt-getting our layers
> above the base system, then removing the packages that are no more required.
> 
> This was the main reason for Neil Williams to create multistrap. After trying
> to wrap debootstrap for some time, he ended up with problems not solvable in
> that way (not sure which, maybe some kinds of circular dependencies or
> dependencies on exact package versions). So, it would be wise to consider his
> experience here.
> 
> That said, on DebConf17, he told us that that use case is not interesting for
> him any more, advised to use debootstrap, and said that he doesn't support
> multistrap any more.
> 
> Regarding maintaining multistrap for some time: I personally don't like Perl.
> That said, Alex looked at that recently; it's some reasonable wrapping around
> apt-get. If the problem in the end is solved with apt-get, we should possibly
> evaluate that, too.
> 
> 
> Regarding debootstrap, the problem is that it is implemented in shell, which is
> horrible for any serious improvements. Thus the question about cdebootstrap.
> 
> 
> All in all, I'd like to see a solution that would offer serious benefits while
> not introducing new problems. Currently, multistrap is an internal issue not
> directly affecting user-visible features. There are many features we can do for
> far better Debian support. Changing one bit with another, not better one
> wouldn't change the big picture. So, I'd rather use this opportunity to start a
> bigger effort of defining the requirements, choosing the tools to merge with,
> provide a working PoC, and start working with Debian on reusability of those
> tools in Isar. IMHO, *debootstrap alone wouldn't bring major architectural
> improvements.

Let's wait for a first prototype. Maybe it will become clearer that we
actually need, if debootstrap can do it, or if we are better of writing
the core and best all the bootstarps in Isar recipes.

Fact is (as I have been told): debootstarp is Debian standard, all
others are secondary or even historic variants. So let's go standard
first, and then see if we need more again.

Jan

-- 
Siemens AG, Corporate Technology, CT RDA IOT SES-DE
Corporate Competence Center Embedded Linux