Will a Kata Pool Fit Through Maintenance?
Use Kubernetes RuntimeClass overhead and measured node headroom to size the Kata worker pool, then price the node capacity needed during a planned drain.
Calculate capacity for a rolling drain
Take a snapshot of capacity available to the pool after accounting for system workloads and any other reserved requests. Sum the worker's container requests and add its admitted RuntimeClass overhead. Calculate the number of whole workers each node can fit in each resource, then use the tighter bound. Now remove the nodes you expect to drain simultaneously and repeat the fit check. On an illustrative node with 4,000 mCPU and 8,192 MiB available, workers requesting 750 mCPU and 1,024 MiB including overhead fit five per node. Two nodes can hold ten workers normally. Draining either leaves capacity for five. An eight-worker minimum would therefore require a third equivalent node during that maintenance window. Repeat the calculation for each node shape if your pool is heterogeneous.
Pod Overhead | KubernetesProve replacement works on the destination node
Kata's Kubernetes isolation boundary is the pod VM. Treat replacement on another node as creation of a fresh sandbox and test the handoff explicitly. In a staging pool, start a harmless fixture job that checkpoints after processing its first input. Cordon a test node through your normal maintenance procedure and let the platform replace the worker. Record where the replacement lands, when it becomes ready, which checkpoint it reads, and how many final records are emitted. A resumed job should produce one final record per input even if the first input is replayed. Compare deployment with extra node headroom against deployment with fewer concurrent drains using the same fixture. If replacement stays pending, inspect scheduling. If it becomes ready but produces duplicate records, inspect the application handoff.
kata-containers/docs/design/virtualization.md at main · kata-containers/kata-containers · GitHubChoose what a warm worker contains
Kata puts the VM boundary around the Kubernetes pod. Put an agent runner and any helper that belongs to the same trust domain in one pod, and assign a separate pod to each tenant that needs its own VM boundary. Define a warm worker as a ready pod waiting for a job. This gives you a consistent unit for measuring both startup latency and reserved capacity.
To decide how many to keep ready, replay a representative arrival burst in a test cluster. Record when each job enters the queue and when its first command starts, first with pods created on demand and then with a small warm pool. A pool passes when the share of jobs that start within your target meets the service objective. If both runs pass, start on demand.
kata-containers/docs/design/virtualization.md at main · kata-containers/kata-containers · GitHubTurn pod overhead into node cost
Read the overhead from the RuntimeClass you actually deploy. Kubernetes adds it to container requests when scheduling, so include it whether an admitted warm pod is busy or idle. Divide the node CPU and memory budgets available to the pool by the corresponding per-pod totals, round each result down, and use the smaller count as the planning limit.
Illustrative inputs only. Suppose a node has 4 vCPU and 8 GiB available to the pool, and each worker requests 0.5 vCPU and 1 GiB plus configured overhead of 0.25 vCPU and 128 MiB. CPU fits five workers and memory fits seven, so six warm workers require two nodes under these assumptions. At a hypothetical $0.20 per node-hour, two dedicated nodes held for ten hours cost $4 in compute. Replace the rate with your provider's billed rate for the exact region, node type, and purchase model.
Pod Overhead | KubernetesFind where the pilot stalls
For a Kubernetes pilot, choose a bare-metal test worker or confirm that your VM worker exposes nested virtualization. Install the kata-deploy Helm chart at the release you intend to evaluate. Check that the staged installation job succeeds on the intended worker and that the chart creates a Kata RuntimeClass. Set your test pod's runtimeClassName to that resource, then run one representative task that reads a test input and writes a known result to the workspace.
If the pod stays Pending without a node assignment, inspect scheduling events and compare the pod's full request with remaining node capacity. If it has a node assignment but never starts and events report sandbox creation failures, inspect runtime logs and verify virtualization on that worker. Count the pilot as successful when the selected runtime starts the pod and the task produces the expected result.
kata-containers/docs/installation.md at main · kata-containers/kata-containers · GitHub