Decompose the container boundary
A Linux container combines mechanisms rather than introducing a separate kernel by definition. A process can appear as PID one in its own namespace while having another process identifier on the host. Mount and network namespaces change other resource views; cgroups address resource accounting and control. For a design review, list which host resources are mounted, which privileges are granted, and which kernel is shared. Separate images are useful packaging boundaries, but an image name by itself does not describe the security boundary, the available authority or protection against a compromised host.
Work a CPU-control example
Assume a cgroup CPU quota of twenty-five thousand microseconds in a period of one hundred thousand microseconds. This describes an aggregate CPU-time limit over the period, not a reservation guaranteeing service or a pinning rule selecting one core. The example is a configuration interpretation, not an executed performance result. A memory limit likewise needs an operational account of what happens near capacity. Placement planning must combine the demands of simultaneously running instances and leave room for supporting processes; setting individual limits alone does not establish that a whole node can sustain the intended workload.
Reason about reconciliation
Suppose the desired state is three replicas. If one disappears, a controller can create a replacement to reduce the difference between desired and observed state. Creation is not equivalent to readiness: the replacement may still be initializing or waiting for dependencies. Define what startup, readiness and liveness observations mean for the application. Reconciliation also does not establish business-level exactly-once behavior. A worker that retries an external side effect after replacement still needs an idempotency or transaction design. The controller maintains an execution arrangement, while application correctness remains a separate responsibility in this example.
Place durable state intentionally
A replaceable worker cannot rely on process memory surviving an update. Distinguish in-memory state, local writable storage, a persistent volume and an external database by their intended lifetimes and failure boundaries. For a queue processor, ask where unfinished work is recorded and how ownership is recovered after the worker disappears. A rollout test should include draining old work and reading data written by the previous version, rather than stopping after a new process starts. Kubernetes workload documentation provides the controller vocabulary; the queue example illustrates the application-state questions that remain outside simply declaring a replica count.