Cluster Cost Management

A Kubernetes bill is mostly nodes you are paying for whether or not your Pods use them. Cost management is the practice of keeping spend proportional to value — right-sizing requests, packing workloads densely, using cheaper capacity for the right work, and attributing cost so waste is visible.

The levers are mostly things covered elsewhere, applied with a cost lens. The recurring theme: idle, over-requested, and unattributable capacity is where the money leaks.

Where the Money Goes

The dominant cost is compute — the nodes — and the biggest waste is the gap between what Pods request (and thus reserve) and what they actually use. Because the scheduler reserves based on requests (Topic 25), an over-requested cluster runs half-empty nodes you fully pay for. Other real costs hide in storage (provisioned but unused volumes), and especially network egress and cross-zone traffic, which is easy to ignore until the bill arrives.

Right-Sizing and Bin-Packing

The first lever is right-sizing: set requests close to measured real usage so reservations reflect reality, recovering the idle headroom. The VPA's recommendations (Topic 28) help find the numbers. The second is bin-packing — letting the scheduler and Cluster Autoscaler consolidate Pods onto fewer, fuller nodes and remove the empties. Workloads must be movable (PDBs, no node-local storage) for consolidation to actually reclaim nodes.

Cheaper Capacity

Not all work needs on-demand, full-price nodes. Spot/preemptible nodes cost a fraction but can be reclaimed with little notice — ideal for fault-tolerant, restartable workloads (batch, stateless web behind enough replicas), wrong for stateful singletons. Reserved/committed capacity discounts the steady baseline you will run regardless. The pattern is a committed baseline plus on-demand for variability plus spot for the interruptible portion.

Attribution and Autoscaling

You cannot manage what you cannot see. Cost attribution — labeling workloads by team/product and using tools like OpenCost/Kubecost to allocate node cost back to namespaces and labels — turns an opaque bill into accountable line items, which is what actually drives teams to clean up waste (showback/chargeback). Combined with autoscaling that scales to demand (and to zero where possible) and the levers above, cost tracks value. The anti-goal worth naming: never cut reliability — replicas, PDBs, headroom — to shave cost, because an outage costs more than the nodes.