Queue length metrics vs Queue delay metrics is confusing

[xemul]

A rough experiment -- disk capacity is 1000 requests/second. Tick is 1ms, which means scheduler can dispatch 1 request/tick.
Next, 100 requests get queued instantly. Scheduler will dispatch them every 1ms. The total queue delay is going to be 1 + 2 + ... + 100 = 5050 ms.

Average delay is 50ms
Disk consumption is 10%
Queue length is likely 0, because it's scraped every few minutes

A load-average-like queue lengths might be helpful, it should make those "spikes" more evident