k3sm

Conformance


⚠️ This is NOT a CNCF Certified-Kubernetes badge#

k3sm has not run the upstream [Conformance] suite, and by design cannot pass it. That suite (CNCF Certified Kubernetes) assumes Linux containers, cgroups, CNI, and network namespaces; k3sm runs Pods as native Darwin processes on Apple Silicon with none of those substrates, so it does not pass [Conformance], and this document does not claim it does.

This document maps k3sm’s targeted feature classes to a synthetic-conformance criterion (a k3sm-authored test, NOT the CNCF suite) or a documented ceiling. A criterion here means “k3sm’s own test asserts this behavior,” not “upstream’s e2e passed.”

Why k3sm Cannot Pass [Conformance]#

This profile is k3sm’s whole public statement of which upstream areas it cannot conform to and why. User-facing tradeoff guidance (what an operator should do about each ceiling) is owned by Limitations.

Terms#

  • criterion: a k3sm-authored Go test that asserts the behavior. It is either a unit test, run by go test ./..., or an end-to-end test in e2e/ (build tag e2e) that k3sm’s acceptance scripts run against a live server.
  • live criterion owed: the behavior ships and its unit tests pass, but the end-to-end test against a live cluster is not written yet.
  • ceiling: a behavior the macOS substrate forbids, or one that routes to the vm RuntimeClass for correctness. It is documented rather than chased.
  • 🟡 planned: the behavior does not ship yet.

Feature classes → criterion or ceiling#

feature classk3sm posturecriterion or ceiling
API machinery (apiserver, RBAC, VAP admission, APF, CRD/aggregated-API, SSA, webhook delivery)embedded real kube-apiservercriterion (Node,RBAC; admission and conversion webhook delivery through the Service proxy)
Admission config (audit logging, PSA cluster default, default LimitRange)argv + config written at provisioncriterion (audit-log level, PSA warn default and its enforce cutover, memory-only default LimitRange)
Pod lifecycle (phase/conditions, probes, graceful stop, OOMKill, restartPolicy, native sidecars, hooks)provider reconstructs the kubelet surfacecriterion; live criterion owed for native sidecars and CrashLoopBackOff
Workload controllers (Deployment/RS/StatefulSet/DaemonSet/Job/CronJob, GC/finalizers, preemption)embedded real KCM + schedulercriterion (the controllers are upstream’s), with DaemonSet toleration injection and Job completion and backoff accounting unit-tested; live criterion owed for Job completion
Services (ClusterIP/NodePort TCP, EndpointSlices, sessionAffinity, internalTrafficPolicy: Local)userspace Service proxycriterion; externalTrafficPolicy: Local is a ceiling (admitted, behaves as Cluster)
DNS (search/ndots/A, ExternalName, in-pod wiring, headless, SRV, PTR, pod A)in-process resolver + getaddrinfo shimcriterion (headless, SRV, PTR and per-pod A answers unit-tested in the resolver)
Per-pod network identity (per-pod IP, headless, StatefulSet identity)a distinct /32 per Pod on the node’s rangecriterion (distinct pod IP per Pod, unit-tested); live criterion owed. Addressing only, never isolation
Ingress / IngressClass / LoadBalancerown userspace L7 proxy; LoadBalancer Services hosted on every node by node poolcriterion (unit-tested ingress hosting, IngressClass status and LoadBalancer bind reporting); live criterion owed
NetworkPolicyService-VIP-mediated ingress allow/deny at the proxypolicy hint on the server node; ceiling as tenant isolation → vm
Scheduling (nodeSelector/affinity/taints, topology labels)real upstream schedulercriterion
Resource model (QoS, memory→OOMKill)proc_pid_rusage samplercriterion memory / ceiling CPU CFS-millicore limits
Observability (/stats/summary, /metrics/resource, lifecycle Events)runtimed usage samples; provider EventRecordercriterion (summary, /metrics/resource joint CPU and memory emission, Pulled/Created/Started/Killing Events)
Config & storage (ConfigMap/Secret, emptyDir/projected, PVC/PV local-path)apiserver-served + runtimed materializationcriterion + 🟡 planned resize/snapshot/generic ephemeral volumes
Auth & supply chain (bootstrap triad, apiserver↔node TLS, code-signing)k3s-style join token pinned to the cluster CA; release binaries are Developer ID signed and notarizedcriterion
Datastore (kine→SQLite WAL, KCM scoping)bundled kine (child process); embedded etcd on an experimental two-server control planecriterion

Hard ceilings (documented, never chased)#

  • No Linux containers / cgroups / CNI / netns / device-plugins / hugepages.
  • No per-pod uid isolation, because same-node pods share one _k3sm trust domain. Untrusted tenancy is destined for vm (linux/arm64, single-node).
  • Absolute-path volume mounts resolve for native workloads (a DYLD_INSERT path-rebase shim; no chroot), not SIP platform binaries. A /bin/sh script can’t read a mounted file at its absolute path, because macOS strips DYLD_INSERT_LIBRARIES from platform binaries.
  • externalTrafficPolicy: Local, CFS millicore CPU limits, HPA-on-CPU-limit, hostPath / terminationMessagePath bind mounts, node-pressure eviction as a hard guarantee, NetworkPolicy as a tenant boundary.

Scope of these claims#

  • Criteria are k3sm’s own tests. A criterion row is “asserted by a k3sm test,” never a [Conformance] pass. Re-derive any row against current main.
  • A 🟡 planned row names behavior that does not ship yet, and a “live criterion owed” row names an end-to-end test still to be written.
  • A live [Conformance] run against the adapted subset remains a future goal, out of scope for this profile.