aasmith / docker-cpu-agent Goto Github PK

When a node is overloaded, the node will report a very low idle number. The consumer, such a proxy will back off connections significantly to help reduce workload on the node. Once this work is complete, the node will likely be very idle, as the upstream proxy will not be sending in many connections. The node will then report 99 idle or something very high. If the node is somewhat degraded, it may not be able to handle peak load, that it will now certainly get from the proxy. It will report near zero and perform below average, continuing this whipsaw effect.

The proposed solution is to gradually report an increase in idle over a given number of samples, therefore allowing the proxy to ramp traffic gradually and discover the right load for the node.

Currently zero idle is a hidden condition. This is desirable for my use case, but may not be for others. It should be a runtime option.

Small detail, but the internal port should be something sensible, not 12345.