Description Cedric Brun 2015-10-02 08:51:57 EDT

I found this issue while investigating tests failures of org.eclipse.sirius.tests.unit.table.unit.export.ExportToCsvTest after migrating it to AQL.
Issue : some metamodels are "unknown" by a specific interpreter whereas it is explicitely referenced in the .odesign in such specific conditions.
Symptoms : failures to parse or evaluate

Analysis : 

in DViewOperations.updateSelectedViewpointsData(IProgressMonitor) the metamodels are gathered and then passed to the model accessor 
and to the interpreters for initialization : 

            Collection<MetamodelDescriptor> metamodels = MetamodelDescriptorManager.INSTANCE.provides(getSelectedViewpoints(false));
            if (accessor != null) {
                accessor.activateMetamodels(metamodels);
            }
            session.getInterpreter().activateMetamodels(metamodels);

Accessor first and then the interpreter. Initializing the ModelAccessor ends up calling 
EcoreIntrinsicExtender.updateMetamodels(Collection<? extends MetamodelDescriptor>) which will diff the given list of metamodel descriptors with the one which has been previously registered
BUT the diff is done by modifying the list passed as a parameters. Because of this some descriptors might have been removed when the interpreter.activiateMetamodels() method is called.

You might wonder : how come we have not seen this problem before ? First this might explain some "hard to reproduce" problems with the MTL interpreter not knowing some metamodels. These 
are typically hard to reproduce as the bug will only manifest itself depending on the previously selected viewpoints.
Second : up until the fix for Bug 474807 AQL would not drop the AST's when the metamodel would change, hence would have been very likely to keep a valid one even if the next "activateMetamdels" call would not 
contain the needed EPackage instance.

To validate this ticket : just make sure ExportToCsvTest, migrated to AQL, is green.