2015-06-03 14:47 GMT+02:00 Thierry Goubier <thierry.goubier@gmail.com>:

2015-06-03 14:19 GMT+02:00 Nicolai Hess <nicolaihess@web.de>:

2015-06-03 13:22 GMT+02:00 Thierry Goubier <thierry.goubier@gmail.com>:

2015-06-03 11:41 GMT+02:00 Nicolai Hess <nicolaihess@web.de>:

2015-06-03 8:22 GMT+02:00 Marcus Denker <marcus.denker@inria.fr>:
Indeed�� this has to do something with caching: when looking a .mcz the system just takes the one it finds, which might be
one with the same name but different.

here is a minimal change to the way it uses the cache, or omitting the cache after the correct version
wasn't find in the cache.

What do you think, may this work?

I'm not sure the problem is there. Often, in configurations, you don't have the full version info, just the mcz name :( And I would expect that a given repo has a single mcz with a given name (i.e. the repo cache should be working properly)...

Nice find that each file based repository has a cache in addition to the main package cache....

The problem is that the package "loading" again searches the cache, without the version id:

search for package with name+id
-> search cache for package with name+id -> "not found
-> search other repos
�� -> other repo:
�� -> search again the cache for package with name+id -> "not found"
�� -> load version with name
�� -> search cache for package with *name only* -> "Found"
�� "other" repo thinks it has loaded its package with that name
�� -> compare name+id -> "not found"
<- nothing found, even if it is (online) in other repo.

Are you sure? Because the code you describe does not correct/affect that.

'self versionFromFileNamed: fileName'

looks only into that special repository cache, not the global one.

self versionFromFileNamed: fileName

is called after it isn't found in the MCCacheRepository

and if it is not found in its own special repository cache , it is loaded with
self loadVersionFromFileNamed:

and this again looks up in the MCCacheRepostory:

loadVersionFromFileNamed: aString
(MCCacheRepository uniqueInstance includesFileNamed: aString)
�� ifTrue: [ ^ MCCacheRepository uniqueInstance loadVersionFromFileNamed: aString].
��
�� ^ self versionReaderForFileNamed: aString do: [:r | r version]

but this time it searches the package in the MCCacheRepitory by its name only and load

the version info from that file.

��

We need to build test repositories to have correct coverage of those issues, because at the moment I'm not sure I understand the case you are describing :(

Try to load
SLICE-Issue-15646-Cleaning-method-category-api-should-be-protocol-part-1-StephaneDucasse.1
from pharo 5.0 inbox repository.

This slice has dependencies to packages
DebuggerActions-StephaneDucasse.75, Kernel-StephaneDucasse.2042

which are both in pharo 5.0 main and inbox repository but with different uuids

If pharo 5.0 main repository is searched first, it will only find packages with the wrong version info,
even if the packages in the 5.0 inbox have the correct version info.

��

Thierry
��

��

Package loads with version info happens in two cases:
- dependency loading (i.e. slices)
- history loading (i.e. browsing history and merges).

Anything else is loading by name, since this is what is recorded in Configurations and Gofer scripts.

Thierry

nicolai

��

Maybe we could for real put (part of) the UID in the filename?

In a disconnected system the case that the filename is the same for different files will always happen if the name is contructed
as it is now.

�� Marcus

> On 02 Jun 2015, at 22:37, stepharo <stepharo@free.fr> wrote:
>
> I hate so much this bug....
>
>
> Le 1/6/15 15:04, Ben Coman a ��crit :
>> Stef, I can now see all the dependent packages for the new slice, but
>> I still have a strange error.�� However I'm not sure if its a bug or
>> something unique at my end.
>>
>> Can someone try merging
>> SLICE-Issue-15646-Cleaning-method-category-api-should-be-protocol-part-1-StephaneDucasse.1
>>
>> What I see at top of stack is two calls to MCVersionMerger>>addVersion:
>>
>>�� MCVersionMerger>>addVersion: aVersion
>>�� records add: (MCMergeRecord version: aVersion).
>>�� aVersion dependencies
>>�� do: [:ea | | dep satisfied |
>>�� dep := ea resolve.
>>�� satisfied := (records anySatisfy: [:r | r version = dep]).
>>�� satisfied ifFalse: [self addVersion: dep]]�� "<<< race? "
>>�� displayingProgress: [ :ea| 'Searching dependency: ', ea
>> package name]
>>�� "15646Note: variable /satisfied/ added for reporting/debugging"
>>
>> One level down from where the error occurs the debugger shows...
>>
>>�� /aVersion/ --> a
>> MCVersion(SLICE-Issue-15646-Cleaning-method-category-api-should-be-protocol-part-1-StephaneDucasse.1)
>>
>>�� /ea/ --> a MCVersionInfo(DebuggerActions-StephaneDucasse.75)
>>
>>�� /dep/ --> nil
>>
>>�� /satisfied/�� --> false
>>
>> and the following which contradicts the value in /satisfied/
>>
>>�� (records anySatisfy: [:r | r version = dep]) --> true.
>>
>> so there seems to be a race such that the ifFalse block is improperly
>> executed, such that the recursive call on top of stack has...
>>
>>�� /aVersion/-->nil
>>
>> hence MNU receiver of "dependencies" is nil.
>>
>> cheers -ben
>>
>>
>> On Mon, Jun 1, 2015 at 10:36 AM, Ben Coman <btc@openinworld.com> wrote:
>>> I tried, but it seems some packages are missing from the inbox.
>>> cheers -ben
>>>
>>> On Sun, May 31, 2015 at 2:19 PM, stepharo <stepharo@free.fr> wrote:
>>>> Hi
>>>>
>>>> I continued to clean that classes have categories and method protocols
>>>> because it was not finished.
>>>> This entry is just adding protocol in the classes that were missing it,
>>>> adding comments, and fixing some local senders
>>>> It does not remove the category API but puts it in a accessing-backward
>>>> protocol and in a second step I will fix all the senders I can (ie not
>>>> Metacello for example).
>>>> Category is really overloaded and we get lost when trying to understand
>>>> code.
>>>> I want to rename RBRule 'category' into 'kind' for this reason.
>>>>
>>>> Stef
>>>>
>>
>
>