INFRASTRUCTURE LEARNINGINFRASTRUCTURE LEARNING
インフラ大全インフラ大全IT INFRASTRUCTURE / LEARNING
Infrastructure learning guide
Find a lesson or lab/

Esc to close · Search terms stay on your device

LEARNING日本語

Topic lessonTopic: Kubernetes and EKS: structure, networking and upgrades

EKS upgrades and recovery: support deadlines, nodes and workloads

An EKS control-plane upgrade does not establish that nodes, add-ons and applications all upgraded together. Identify each change, then align support deadlines, compatibility, availability during the change and recovery conditions. An update status alone does not prove successful user operations.

What you will learn

Separate update targets and ownership; explain standard/extended support, PDB boundaries, conditional rollback and data recovery.

Prerequisite

The control plane coordinates Kubernetes; nodes run Pods. An AMI is a machine image, kubelet manages Pods on a node, and add-ons provide capabilities such as DNS/networking. No live account is required.

Separate the target and owner of each update

Four layers to record before and after a change
TargetExample changeSeparate verification
ApplicationDeployment Pod template or image.Readiness, user responses and data compatibility.
Control planeKubernetes minor version.Node/add-on/client compatibility and observed versions.
NodeAMI, kubelet or execution infrastructure.Eviction destinations, placement, startup and readiness.
Add-onVPC CNI, CoreDNS and related components.Supported versions, configuration, dependencies and behavior.
Version relationships, not a simultaneous update sequence
Version relationships, not a simultaneous update sequence1: Control plane — Compare compatibility and actual versions — Nodes / kubelet. 2: Control plane — Review supported combinations — Add-ons / clients. 3: Nodes / kubelet — Observe placement and execution — Application / data123Control planeNodes / kubeletAdd-ons / clientsApplication / dataVersion relationships, not a simultaneous update sequence1: Control plane — Compare compatibility and actual versions — Nodes / kubelet. 2: Control plane — Review supported combinations — Add-ons / clients. 3: Nodes / kubelet — Observe placement and execution — Application / data123Control planeNodes / kubeletAdd-ons /clientsApplication /data
  1. Control plane → Nodes / kubeletCompare compatibility and actual versions
  2. Control plane → Add-ons / clientsReview supported combinations
  3. Nodes / kubelet → Application / dataObserve placement and execution
Control plane
Kubernetes control-plane version
Nodes / kubelet
Check updates for the compute approach
Add-ons / clients
Check compatible versions and settings
Application / data
Verify user operations

Record current version, target, owner, verification method and stop conditions for each layer. An application Pod-template change and a Kubernetes cluster-version change are distinct operations.

Separate support durations from the actual deadline

Support measured from availability on EKS
PhaseDocumented duration/conditionPlanning implication
Standard support14 months from EKS availability.Do not substitute the upstream Kubernetes release date.
Extended support12 months after standard support; additional cluster-hour charges.Review upgrade policy and costs.
After extended supportControl plane becomes subject to automatic upgrade.Plan proactively rather than relying on an exact automatic-update time.

Knowing the duration does not establish a particular cluster’s deadline. Match its version and upgrade policy to the official support calendar. Calendar dates use UTC; month-only dates are approximate. References were checked on 2026-10-26; no version is labelled permanently latest.

Distinguish EXTENDED from a policy that avoids extension and triggers automatic upgrading at standard-support end. Managed/self-managed nodes and existing Fargate Pods still need separate version checks. Review compute, storage, networking and optional features as well as support charges.

Check compatibility and advance one minor version at a time

EKS control-plane minor upgrades proceed one step at a time. Review deprecated APIs, cluster insights and node/add-on/kubectl compatibility first. Allowed version skew does not authorize arbitrary combinations or sequencing.

Check the compute approach after control-plane success
ApproachUpdate verification
Managed node groupPlan node updates separately; observe versions and Pods afterwards.
Self-managed nodesCustomer-managed node versions, maintenance and compatibility.
Auto ModeInspect AWS-managed update behavior and workload disruption/outcomes.
FargateDo not assume existing Pod kubelets updated; plan replacement and validation.

For application rollout, match Deployment configuration with old/new Pod observations and actual user operations. Do not reuse ECS minimumHealthyPercent/maximumPercent as Kubernetes settings. Similar terminology does not make the rules identical.

Distinguish disruptions constrained by a PDB

A PodDisruptionBudget (PDB) constrains unavailable replicas during supported voluntary eviction. Draining a node evicts workloads for relocation. A PDB does not block every direct deletion, node failure or workload-controller rollout.

Distinguish managed-node-group update operations
OperationPDB behaviorDecision evidence
Rolling updateRespects PDBs; inability to drain can fail the update.Check readiness, spare resources and placement constraints.
Force updateDoes not respect PDBs.Do not treat ignored disruption constraints as a routine shortcut.
Scaling down desiredSizeNot protected like the ordinary update/drain workflow.Plan node termination and application effects separately.

Seven days is only one control-plane rollback condition

EKS documentation checked on 2026-10-26 describes rollback to the previous minor version within seven days of an in-place upgrade completing. A cluster created at its current version is ineligible. Support/policy, cluster state, incompatible features, readiness insights and node/add-on compatibility also matter.

Two fixed records with time information only
RecordWhat is establishedWhat remains unknown
48 hours since completion; previous minor proposed.Within the time window.Whether other eligibility conditions hold.
192 hours since completion.Exceeds seven days: 168 hours.Feasibility and conditions of other recovery approaches.

Auto Mode nodes are managed within rollback; managed node groups and other compute need separate preparation. Do not assume equivalent Fargate-node rollback. Control-plane rollback preserves customer data/workloads and does not revert add-ons. Force is not a universal override of prerequisites such as the time window.

Define recovery targets and post-recovery verification

  1. Stop further changes and align targets, versions, times and user impact.
  2. Identify application, node, control-plane and add-on recovery separately.
  3. Check deadlines, support, insights, data compatibility and ownership.
  4. Assess disruption and persistent writes/external notifications for the chosen approach.
  5. Verify actual versions, Pods, connectivity and user operations, not only operation status.

Check update targets and recovery conditions

Choose an option to see the result and explanation.

The control-plane upgrade succeeded; node/add-on versions are uncollected. Has everything upgraded?
Read the answer and reason

Answer: No; verify nodes and add-ons separately · Each target needs observed versions and compatibility checks.

Does a PDB guarantee that direct Pod deletion is blocked?
Read the answer and reason

Answer: No; direct deletion bypasses the PDB · Distinguish supported eviction constraints from direct deletion.

It is 48 hours after an in-place upgrade completed. Is rollback to the previous minor guaranteed?
Read the answer and reason

Answer: No; other eligibility conditions still matter · Time alone does not establish support, state or compatibility.

Control-plane rollback succeeded. Did application writes and add-ons revert too?
Read the answer and reason

Answer: No; inspect data, workloads and add-ons separately · Separate preserved state from components whose versions change.

Check, then explain in your own words

  1. Identify four update targets and their owners.
  2. Separate support duration from actual version/policy/deadline.
  3. Explain PDB and rollback boundaries, then define post-recovery observations.
Check your understanding: A report says only “cluster upgrade successful.” What would you verify next?

The changed target/version, support deadline, node/add-on/client compatibility and actual versions, Pod/connectivity/user outcomes, rollback conditions and data state.

One operation’s success does not establish synchronized component updates or application/data recovery.

Sources and conditions checked

The official references used in this lesson are collected here with their scope, conditions and corresponding sections. Independent expert review and live AWS comparison have not been performed.