Clarification on scaling limitations and performance considerations about clusternet

[dixudx]

@clarkmcc Recently I did a perf test on clusternet-hub, as below.

In the perf test, clusternet-hub and kube-apiserver are running with single instance. When managing 2k clusters, clusternet-hub takes about 250 MB memory, while kube-apiserver takes about 1GB, the watch latency on sub-cluster resources reaches about 600 ms.

Both clusternet-hub and kube-apiserver are stateless. With more replicas the latency could be reduced as well.

Also I've run perf tests with a larger scale, managing 1M clusters (one cluster = one clusternet-agent goroutine), single clusternet-hub (I run 30 replicas of this) agent can manage ~40k clusters.

[dixudx]

The design goal of Clusternet is to support as many Kubernetes clusters as practically required. This in fact is a major differentiator of Clusternet, compared to many other multi-cluster Kubernetes solutions.

Currently we're still doing performance optimizations and enhancements. Related benchmarks will be given in v0.5.0.

Yes, you are very right that 'hundreds of millions' is an exaggerated expression for the current state of the project. While we do have a dream to manage 'billions of things' (as in IoT), for now we want to be more accurate. We will update the README accordingly. Thank you so much for pointing out. We are eager to know more about your use case with thousands (hundreds of thousands) of agents.