dynamic up/down weighting

[willscott]

we should look at the 'average error rate' that we're seeing across the pool of nodes.

If all nodes are failing 1% of the time, the weight-ranking should start to consider that 'normal' and not penalize nodes too much for it. We may want to track a moving window of success/failures per member, and do periodic re-weightings based on the states of those, rather than directly in response to failed requests.