Hierarchical EDS weighted load balancing

[htuch]

EDS now expresses weights at both the endpoint and locality level, these need to be combined to produce an effective LB weighting. The issue tracks the implementation work on this.

[rshriram]

Define locality: data center or az ?
Or weighted clusters? If latter, the labels are already chosen in route block.

[alyssawilk]

This includes proper support for all of the load balancers: fixing the 'stickiness' for LeastRequestLoadBalancer, proper weighted round robin for, ring hash, etc.

Once this is done we can remove the suggested load balancing weight limits in the eds config.

[alyssawilk]

Given #1929 took longer than expected I'm probably not going to be able to pick this one in up in the near future.

[mpuncel]

is this a duplicate of #2725 ?

[alyssawilk]

Not quite? For example I don't think we want to remove suggested weight limits for eds until/unless we refactor/replace LeastRequestLoadBalancer

[htuch]

I don't think we're going to be removing the weight limits, given how the hierarchical Maglev balancer works today. I think we should close this one out, as the work is complete, feel free to open additional tracking issues for point features.

[ZackButcher]

Should we update the comment in the API about this issue given it's been closed?

envoy/api/envoy/api/v2/endpoint/endpoint.proto

Lines 109 to 114 in 6a8b843

	// .. attention::
	//
	// The limit of 128 is somewhat arbitrary, but is applied due to performance
	// concerns with the current implementation and can be removed when
	// `this issue <https://github.com/envoyproxy/envoy/issues/1285>`_ is fixed.
	google.protobuf.UInt32Value load_balancing_weight = 3

[htuch]

Agree, PRs welcome :)

[ZackButcher]

The resolution is not clear to me from this issue given it was closed with no changes. Based on current LB impl, the limit of 128 sticks around due to perf concerns and the reference to this issue should be removed, or have the underlying perf concerns been addressed?

[mattklein123]

There aren't any perf concerns anymore w/ weighting for both RR and LR. Both have been fixed. I don't recall exactly how Maglev works. TBH I would probably just update the comment to remove the attention link until if/when someone complains that the weight max needs to be larger than 128.