The taint is where this trade actually lands. The hub serves one tenant, itself, so it gets a single worker node, and the masters absorb whatever that worker cannot hold once the NoSchedule taint comes off [9]. The hub's tenant list is Prometheus, Grafana, Loki, ArgoCD, Harbor and Vault [1]. If the overflow is the log store and the registry, the three nodes running etcd are also serving image pushes and chunk writes. There is no resource figure to check that against: the only nodes sized anywhere in the write-up are the HAProxy pair at 1C/2G/30G each [7].
Two prep steps sharpen it. Swap is off across masters and workers, and the stated reason is that kubelet will not start with it enabled [19]. On an untainted hub master, that means a memory spike gets resolved by the kernel killing something on a box that also holds an etcd member. Then ufw is stopped and disabled [20], on every master and worker but not on the haproxy nodes, which never run the prep [18]. Consolidation puts the platform's secrets and images on those same nodes [4][5], with no host packet filter left in front of them. Disabling the host firewall is the cheapest line in any runbook and the one most likely to come back at review.
The roughly four-second VIP move is a claim about this author's Keepalived configuration, not a property of the pattern [13]. For the number to transfer, your VRRP advert interval and your haproxy health check have to match theirs, and it only ever covers the load balancer tier of one cluster. Every failure mode documented here sits inside a single cluster [28]. What a hub outage does to the spokes is not written down, and the dependencies are explicit: spoke image pulls go to hub Harbor [4], spoke secrets arrive from hub Vault through agents or sidecar injectors [5], and spoke deploys come from hub ArgoCD [6].
Count the machines. Hub: two HAProxy, three masters, one worker. Each spoke: two HAProxy, three masters, and eight workers, since the spoke worker range runs 01 to 08 at .80 through .87 [15][24]. Three byte-for-byte identical spokes [11] plus the hub is 45 nodes [25]. The claimed saving is 4x infrastructure, 4x maintenance and 4x dashboards [2]. Across four clusters, running the stack once removes three of the four copies, or 75 percent [27]. The copy that survives is the one carrying four clusters' metrics, logs and images, and it is the copy with one worker.
What is explicit here is genuinely explicit. Identical spoke topology means one template rather than three near-duplicates [11]. Each cluster gets its own join token, with reuse across clusters ruled out [17], so the one thing deliberately not shared is the credential that would make consolidation a blast radius. The network prep explains each knob instead of listing it, down to pod traffic stopping at the node boundary without ip_forward [22]. The consolidation trade is the part still asserted rather than sized. Two numbers would settle it: CPU and memory for the untainted hub masters, and a written answer for what a spoke can still do with the hub gone.
