RFE: flag to keep multilib package versions in sync

[ignatenkobrain]

repo system 0 empty
repo available 0 testtags <inline>
#>=Pkg: foo 2 1 x86_64
#>=Pkg: foo 1 1 i686
system x86_64 rpm system
poolflags implicitobsoleteusescolors
solverflags allowvendorchange keepexplicitobsoletes bestobeypolicy keeporphans yumobsoletes
job install pkg foo-2-1.x86_64@available
job install pkg foo-1-1.i686@available
result transaction,problems

results into:

Transaction summary:

2 installed packages:
  - foo-1-1.i686
  - foo-2-1.x86_64

From https://bugzilla.redhat.com/show_bug.cgi?id=1326157

[mlschroe]

Yes. Do you really want to enforce it for explicit user requests as well?

[ignatenkobrain]

@jsilhan, what do you think?

[mlschroe]

Btw, you can "fix" that in hawkey/libdnf with SOLVER_NOAUTOSET, so that the solver doesn't assume the arch is intentional for that job.

If SOLVER_NOAUTOSET is used, libsolv does not attempt to guess the intention from the job. I.e. it does not assume SOLVER_SETNAME for SOLVER_SOLVABLE_NAME installs, SOLVER_SETNAME | SOLVER_SETARCH | SOLVER_SETVENDOR | SOLVER_SETREPO | SOLVER_SETEVR for SOLVER_SOLVABLE installs, and so on. See jobtodisablelist() in solver.c

[ignatenkobrain]

that seems like workaround... Would be nice to get flag which would completely enforce to be all multilib arches in sync

[mlschroe]

Well, I won't add it if it will not be used, and your question above is not answered yet...

[jsilhan]

Currently DNF do queries on packages, find package ids based on filters and then put it into selector to resolve one of given packages. If the flag would handle multilib_policy=best/all by itself that would be an improvement. I would especially invite this flag during VERIFY job so it would report these multilib violation when the new flag will be set (if it;s not libsolv doing already).

[dmach]

We want the packages to be at the same version.
The system is better auditable and users know exactly what's installed.
Having a libsolv feature (a flag?) for that would be much nicer than hacking this in libdnf/dnf directly.

[ignatenkobrain]

@mlschroe friendly ping, we want to lock-in 32bit and 64bit versions of same package to same evr in Fedora.

[ignatenkobrain]

(by having solverflag or poolflag)

[ignatenkobrain]

@mlschroe we would really like to have this flag for Fedora.

[ignatenkobrain]

Yes. Do you really want to enforce it for explicit user requests as well?

I think this is still not answered… Yes, this should make sure that multilib packages are kept in sync all the time.

[mlschroe]

I can easily add a flag that makes the solver not disable "infarch" rules, but that means that there will be no way to install that foo-1-1.i686 package with dnf (as there is no corresponding foo-1.1.x86_64 package).

[ignatenkobrain]

@mlschroe I'll check the YUM behavior and write it back here.

[dmnks]

I promised to chime in with some information on how yum handles this.

In yum, we have the following config option (see yum.conf(5)):

protected_multilib

Either `1' or `0'.  This tells yum whether or not it should perform a check to
make sure that multilib packages are the same version.  For example, if this
option  is  off  (rpm  behavior)  then  in  some  cases it might be possible
for pkgA-1.x86_64 and pkgA-2.i386 to be installed at the same time.  However
this is very rarely desired.  Install only packages, like the kernel, are
exempt from this check.  The default is `1'.

If the option is enabled, yum will verify in the transaction building phase in YumBase.buildTransaction() that any duplicate packages (as in, having the same name) in the just resolved transaction set are of the same version. In case they're not, a verbose message is printed.