Re: [gentoo-portage-dev] [PATCH] ArchChecks: don't mix arches between ebuilds

[Zac Medico <zmedico@gentoo.org>]

On 04/24/2016 02:45 PM, Zac Medico wrote:
> On 04/24/2016 01:00 AM, Brian Dolbec wrote:
>> On Sun, 24 Apr 2016 08:15:56 +0200
>> Michał Górny <mgorny@gentoo.org> wrote:
>>
>>> On Sat, 23 Apr 2016 16:57:21 -0700
>>> Zac Medico <zmedico@gentoo.org> wrote:
>>>
>>>> Fix ArchChecks to not mix arches of ebuilds together, so that
>>>> errors are only reported for those arches that the ebuild has
>>>> keywords for.
>>>> ---
>>>> Applies to the *repoman* branch.
>>>>
>>>>  pym/repoman/modules/scan/arches/arches.py | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/pym/repoman/modules/scan/arches/arches.py
>>>> b/pym/repoman/modules/scan/arches/arches.py index 4df25a8..6e1c17d
>>>> 100644 --- a/pym/repoman/modules/scan/arches/arches.py
>>>> +++ b/pym/repoman/modules/scan/arches/arches.py
>>>> @@ -69,7 +69,7 @@ class ArchChecks(ScanBase):
>>>>  				arches.add(('**', '**', ('**',)))
>>>>  		# update the dynamic data
>>>>  		dyn_arches = kwargs.get('arches')
>>>> -		#dyn_arches.clear()
>>>> +		dyn_arches.clear()
>>>>  		dyn_arches.update(arches)  
>>>
>>> Is that some crazy way of replacing contents of an existing dict? May
>>> be worth a comment since the code looks suspicious, at least.
>>>
>>
>> It was an interim step so that modules didn't pass back
>> arbitrary data to be re-used by other modules after it.  This way
>> maintains the data pointer to the set (not dict) object in the calling
>> function. Just assigning it a new set doesn't set the 'arches' pointer
>> in the calling function so that it can pass that data along to other
>> modules.
> 
> I feel like the plugin structure has lead to decoupling of things that
> really should be coupled. For example, the code in the ArchChecks could
> be a function that's called by the ProfileDependsChecks plugin, given
> that ProfileDependsChecks seems to be the only consumer of this data.

We'll have to go through all of the plugins and look for other things
that have too aggressively decoupled. The fact Scanner._scan_ebuilds
calls plugins in a hardcoded order is troubling. It means that there may
be coupling between plugins that makes the order matter, yet this
coupling is hidden and non-trivial to discover. If we couple more things
together where appropriate (like ArchChecks and ProfileDependsChecks),
then it will make the system easier for a person to reason about.
-- 
Thanks,
Zac