Switch Balancer group balancing from Random to EDF #2965
Conversation
EDF is strictly better than random
| // RandomWRR constructor, to be modified in tests. | ||
| var newRandomWRR = wrr.NewRandom | ||
| // WRR constructor, to be modified in tests. | ||
| var newWRR = wrr.NewEDF |
There was a problem hiding this comment.
RandomWRR is the behavior agreed on by all implementations of different gRPC languages. We cannot change the default.
We can make it configurable so you can override it in your client.
Changing the algorithm will also be easier in the future as there are also plans to support different locality picking algorithms.
There was a problem hiding this comment.
Was this discussion public? It looks a bit strange that gRPC uses different algorithm compared to Envoy (which uses EDF for this exact scenario)
There was a problem hiding this comment.
The xds gRFC is not public yet.
Is there any specific reason you care so much about this picking algorithm?
There was a problem hiding this comment.
If we decide to switch to EDF to be consistent with envoy, we should do that in all languages. I will bring this up with them.
There was a problem hiding this comment.
We care for endpoint lb (as edf provides fair distribution of the load across endpoints as in opposite to ≈random one).
For cross-zone lb we don't care that much (as number of zones is not that big) and the main/only motivator is consistency with other data plane (envoy).
|
Closing this PR for now, as it's not decided whether we will want it yet. |
EDF is strictly better than random.
This is continuation of #2957.