Make package RECOMMENDS a stronger dependency about libsolv

[mlschroe]

But isn't that already what libsolv does? It tries to install all recommended packages, unless it's not possible. And there's already a flag to turn this off (SOLVER_FLAG_IGNORE_RECOMMENDED).

[kirgene]

Actually it doesn't.
The following test fails:

repo system 0 testtags <inline>
#>=Pkg: A 1 0 all
#>=Pkg: C 1 0 all
#>=Pkg: Z 1 0 all
#>=Rec: A = 1-0
#>=Rec: C = 1-0
repo test 0 testtags <inline>
#>=Pkg: A 2 0 all
#>=Pkg: B 1 1 all
#>=Pkg: C 2 0 all
#>=Pkg: Z 2 0 all
#>=Rec: A = 2-0
#>=Rec: B = 1-1
#>=Rec: C = 2-0
system i686 deb system

job install name Z [orupdate]
result transaction,problems <inline>
#>install B-1-1.all@test
#>upgrade A-1-0.all@system A-2-0.all@test
#>upgrade C-1-0.all@system C-2-0.all@test
#>upgrade Z-1-0.all@system Z-2-0.all@test

with the result:

Results differ:
-upgrade A-1-0.all@system A-2-0.all@test
-upgrade C-1-0.all@system C-2-0.all@test

[mlschroe]

Yes, that's true. It solves your example that way because it currently has a focus on doing minimal changes to your system, and thus it prefers not updating A and C and breaking the recommends dependency over the update. Is this some real issue that needs fixing?

[mlschroe]

I can add a solver flag that makes it add weak rules for the recommends, this will make the solver try hard to fulfill all recommends.

[kirgene]

Good idea.

[adelcast]

The libsolv backend of opkg has the same problem, it's a regression from when the internal solver is used. A flag that will allow opkg to keep the same behavior across backends would be fantastic!

https://bugzilla.yoctoproject.org/show_bug.cgi?id=9939

[mlschroe]

Now done: the SOLVER_FLAG_STRONG_RECOMMENDS makes the solver create rules for recommends dependencies. Pretty much untested, though...

[adelcast]

While testing SOLVER_FLAG_STRONG_RECOMMENDS, I found a potential issue. In a repo that has the following (pseudo code form opkg ATS):

o.add(Package="x")
o.add(Package="a", Recommends="b")
o.add(Package="b", Recommends="c")
o.add(Package="c", Conflicts="x")

And package 'x' is installed, when an 'install a' operation is run, only 'a' is installed. Seems to me like a better solution would be to install 'a' AND 'b', which is what you get without SOLVER_FLAG_STRONG_RECOMMENDS. Would it make sense to have SOLVER_FLAG_STRONG_RECOMMENDS follow that behavior? I can open an issue, if you think this suggestion makes sense.

[adelcast]

@mlschroe