Summary
A fleet of 2,500 applications was moving to Kubernetes, and the first question was how many clusters to run: one for the whole fleet, one per department (three), or one per business domain (six). The article compares a single cluster with several on the same workload. One cluster is cheaper, simpler to administer and uses its nodes better; but every upgrade reaches every application, isolation rests only on namespaces and policies, and a flaw spreads to everything. Beyond a proof of concept, running production on a single cluster is strongly discouraged.
Two questions weighed more than capacity: node upgrades, which follow three Kubernetes minor versions a year and can break what already runs, and performance campaigns, run by a dedicated team in each department. The recommendation took the isolated business domain as the unit. Two pairs of domains were merged, giving four groups, and each group got two permanent clusters, one for non-production work and one for pre-production and production: eight permanent clusters, from 7 to 50 worker nodes off production and from 60 to 230 in production. Outside the eight, a sandbox cluster of one master and one node is created on demand for a limited time, to test tools, qualify upgrades, train teams and run performance campaigns.
The leadership point is that cluster topology mirrors the organisation. Each of the three splits, by environment, by product or by team, assumes a different owner. The real question behind “how many clusters” is who can decide alone on an upgrade, a policy or a maintenance window: a cluster shared by several owners turns every change into a negotiation, a cluster per owner turns it into a decision.
Key ideas
- Beyond a proof of concept, a single production cluster trades low cost and efficient use of nodes for weak isolation, low resilience and flaws that spread to the whole fleet.
- Capacity limits rarely decide the split. Upgrades, with three Kubernetes minor versions a year, and performance campaigns do.
- The chosen topology took the isolated business domain as its unit: four domain groups of two permanent clusters, non-production and production, eight in all.
- A sandbox cluster of one master and one node, created on demand for a limited time, takes in tool trials, upgrade qualification, training and performance campaigns.
- Cluster topology mirrors ownership: draw the cluster boundary where one owner can decide an upgrade alone.