This makes a cross context call be treated as one level when calculating the probability. this makes previous contexes not be completely invalidated when a cross context call is sent.
While working on the Alarm skill I discovered several issues with the
event scheduler. This PR cleans up those findings and resolves several
other potential issues:
1) To avoid thread synchronization issues, the EventScheduler had several
queues which independently held objects to be added/deleted/updated. However, the order of the events was undefined and got mixed since they were all batched together. So, for instance, if skill code performed:
self.add_event("foo", self.handle_foo)
if SomeReason:
self.cancel_event("foo")
The actual order of queue handling would perform Remove first, then Add which resulted in "foo" not being found for delete, but then added and left as an active event.
Now the EventScheduler protects the list using a Lock and the queues have been removed. Modifications to the list happen immediately after obtaining the lock and are not batched up.
2) One-time events were triggered while the event was still in the EventScheduler list. Now the entry is removed before invoking the handler.
3) Within the MycroftSkill.add_event(name, handler) is a local 'wrapper' method that actually makes the callback. The MycroftSkill.remove_event(name) method attempted to find entries in the events list and the associated handler entries in the self.emitter to remove. However, the emitter actually held the wrapper(handler), not the handler itself. So the emitter handlers were left behind.
This was a quiet bug until the next time you scheduled an event of the same name. When that second event finally triggered, it would fire off both the new and the old handler -- which snowballed in the 'skill-alarm:Beep' case, doubling and redoubling with every beep.
Now this cancels all the emitter listeners by name. There is a very slim chance that someone has registered a listener with the same name, but since it is namespaced to "skill-name:Event" I think this is pretty safe.
Not technically related, but a failure that has been lurking for
some time and is a French unit test that doesn't work depending
on the time of day when the test is run.
Åke
jarbasal
Steve Penrod
Matthew D. Scholefield