Polling has several benefits over pushing: a misbehaving endpoint can't knock out the whole system by having too much data, but at worse will just not get polled. Similarly load shedding is pretty simple.

In my experience the discovery and configuration is needed anyway to do something with the metrics.