Virtualization · Cluster
Orchestration
Scheduling, reconciliation, and rollout.
- 01 The Reconciliation Loop and Level-Triggered Control of Declared State How a controller compares declared state with observed state, why every event is only a hint to re-read, and what a stale cache costs at the write.
- 02 Filtering and Scoring: How a Scheduler Picks One Node Out of Thousands The two-phase placement cycle, why feasibility is computed per node instead of per constraint, and how scoring turns a set of legal nodes into a single binding decision.
- 03 Resource Requests, Limits, and What the Scheduler Actually Counts Why a request is a scheduling reservation rather than a runtime guarantee, how limits are enforced by cgroups, and why the two numbers produce three quality-of-service classes.
- 04 Placement Constraints and the Anatomy of an Unschedulable Workload How taints, tolerations, affinity, and topology spread interact so that a workload can be legal on every node individually and still fit nowhere, and how preemption resolves it.
- 05 Rolling Updates, Surge, and the Arithmetic of Availability How a rolling update walks a replica set down and up without dropping below a floor, why maxSurge and maxUnavailable interact with readiness rather than with process start, and where the arithmetic breaks.
- 06 Heartbeats, Leases, and the Timeout That Declares a Node Dead Why node liveness is a leased lease with a bounded renewal, how the kubelet's lease write differs from its status report, and what a false positive costs when the network partitions.
- 07 Canary Releases and the Statistics of Deciding to Promote How traffic is split between two versions without a shared load balancer, why the promotion decision needs a sample-size calculation, and what a canary cannot detect at any sample size.
- 08 Owner References, Finalizers, and Why Deletion Is a Protocol How the API server tracks which object owns which, why a delete is a two-phase operation with an intermediate state, and what a finalizer that never completes does to a namespace.
- 09 Leader Election, Leases, and What Quorum Protects Against Why a control plane needs a single active writer, how lease-based election bounds the overlap window, and where the underlying consensus store changes the guarantee.
- 10 Autoscaling as a Feedback Loop: Gain, Lag, and Why It Oscillates How a replica controller turns a metric into a desired count, why the control law is proportional rather than predictive, and how sampling lag and stabilisation windows prevent or cause oscillation.