Proposal: Latency-aware WRR for k8gb (external metrics → dynamic DNS weights)

[mariks1]

Proposal: Latency-aware WRR for k8gb (external metrics → dynamic DNS weights)

Note: This is part of a university student project (ITMO University) in the course “Cloud and Fog Computing”. Selected topic:
“Limitations in Open-Source Tools for Real-time Network Latency Optimization Across Geographically Dispersed Cloud Regions”.
The goal for the subject is to make a pull request to an open source project that closes a practical aspect of the topic.

Proposal

1. Documentation + running example (no changes to the main code):
   A short example showing how to get latency metrics, calculate weights, and update k8gb WRR weights.

2. Optional controller logic:
   Behind the feature flag, operator can read p50/p95 delays for each region and automatically update WRR weights. Disabled by default.

Is this feature useful for k8gb?
If yes, can I implement it?

[kuritka]

Thanks a lot for the proposal — before diving in, please make sure to review our Weighted Round Robin caveats. It describes the key limitation that applies directly to your idea.

In short: k8gb correctly publishes weighted records via CoreDNS, but most production setups use a parent DNS (e.g. BIND, Infoblox, Route53). Those layers usually randomize, cache, or reorder answers, so the final client-side split rarely matches the configured weights. Because of that, a latency-aware WRR can work in special cases (when client app query CoreDNS directly), but not as a predictable production feature.

That said, this area would actually be very valuable to investigate further. If as part of your demo you could explore which DNS servers or configurations preserve WRR ordering or support deterministic weight policies, that would be extremely helpful for everyone using k8gb in enterprise setups. There’s not much public data on this, and your findings could feed directly into our documentation and best practices.

No need to touch CoreDNS or k8gb logic — a documentation-oriented analysis or demo setup would be perfect with high impact.

Looking forward to seeing what you discover!