When optimizing, a metric is like a lighthouse in heavy fog.
It's a proxy for the underlying reality.
When the lighthouse is in the distance, it's a good thing to sight off.
When the lighthouse is close, if you track it too closely you will crash in the shoals.
In metrics, as you get asymptotically closer to the goal, you get dangerously lost, chasing things that don't matter.
The metric is always a proxy.
When you're far away, the proxy is closer to the real endpoint than other things, helping point you in the right direction.
When you're close, the proxy is farther away from the real endpoint than other things, pulling you away from the right direction.
Metrics need to be in the middle-distance to be true guides.