Sorry to necro, but I think this is important. At least one active instance incurs costs (got caught by this recently), so it would be nice if we could scale up from zero without an AWS lambda instance or something.
As for the QueueDepth metric, what I noticed is that there is two versions of it: one for the overall queue and one for the fleet. The rule-based metric goes off the fleet one, which for some reason remains at zero. This is what I mean: (dark blue is Queue depth and cyan is fleet Queue depth)
This is the configuration that I would like:
Ideally the process would be client calls
StartGameSessionPlacement -> policy
scale goes into effect -> add 1 new instance
Then, when active sessions drop to zero, we scale down until eventually zero instances (and no costs )