[RFI] How does this differentiate between a dependent service getting slower and standard too much concurrency impacting latency?

[huggsboson]

I can absolutely see as this thing is tuning connection limits up and down watching average latency being a great signal. What happens when you're at steady state and one of your dependencies (say a database) becomes unresponsive and your latencies spike? It's hard to prescribe the perfect one-size meets-all response to this situation but in some scenarios I'd want it to add more concurrency to counteract the blocked threads waiting on a DB connection timeout.

Ideally it would incorporate a cpu or blocked threads metric into the equation to help determine if it's a dependency or the service itself that is impacting latency.