The destruction order of objects during shutdown is totally unclear to me for now. Since people could potentially hold a smart pointer to a service object everywhere, it is always possible that this object is never released. This is of course a weakness in the design and should be corrected.

When I write preferences (outside of the preferences dialog), I always automatically call flush() to make sure my changes get written to disk (you never know if a bundle will crash the application in the next seconds...). I would suggest to communicate this as being the "proper policy" when handling preferences.

I will also add a call to flush in the workbenchs shutdown code.

Michi fixed the preferences problem, see T1850 comment #30. Removing 3M keyword.