2015-06-03 15:09 GMT+02:00 Nicolai Hess <nicolaihess@web.de>:
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.
Thanks for finding that. Interesting, that semantic missmatch ... which amounts to loosing the version info in places during the loading process. Notes for documentation : - FileTree changes versionFromFileNamed: to disable the internal repository cache (but cache is still an instance variable because of inheritance, and it goes through PackageCache). - GitFileTree avoids the PackageCache altogether and is the only type of repo not to have that bug in Pharo4 (Yes! But I had no idea why :)).
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
Yuck :( So, if I understood well, what is happening is: - Reading from Pharo5Main, DebuggerActions-StephaneDucasse.75 is loaded in PackageCache (and in the cache of Pharo5Main). - Comparing UUIDs say that this is not the right one. - Reading from Pharo5Inbox, DebuggerActions-StephaneDucasse.75 is loaded in the cache of Pharo5Inbox - First read inside package cache say that the file exist but has the wrong UUID - Second read has matching file name inside Pharo5Inbox - So it loads the version from Pharo5Inbox - which test PackageCache for DebuggerActions-StephaneDucasse.75 ... which is true - But uuids don't match - so it fails.... I hate package-cache.
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.
You're right. So your fix should work. I would not hesitate however: - to move the version reading line inside versionWithInfo:ifAbsent: - to factor the cache update into an #updateCacheWith: v (so that it can also be called from versionFromFileNamed:) called from versionWithInfo:ifAbsent: Thierry
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 écrit :
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