- change the Test metallb-system namespace task to run
k3s kubectl get namespace metallb-system instead of the bare
-n metallb-system, which printed kubectl usage and always exited 0
- add a regression test that asserts the task uses an explicit get
and fails if it ever regresses to the usage-only form
- wire the new test into pre-commit
* feat(kube-vip): add endpoint override for the internal listening address
Add a kube_vip_endpoint variable so the address kube-vip binds and listens on
can differ from the announced apiserver_endpoint. This is useful for complex
routing and site-to-site tunnels where the VIP kube-vip advertises over ARP
differs from the address it listens on internally.
- roles/k3s_server/templates/vip.yaml.j2: use kube_vip_endpoint (defaulting to
apiserver_endpoint) for the `address` env and for deriving `vip_subnet`
- roles/k3s_server/defaults/main.yml: add kube_vip_endpoint default (null)
- roles/k3s_server/meta/main.yml: add kube_vip_endpoint argument_spec
- inventory/sample/group_vars/all.yml: document the new sample variable
- README.md: document the kube_vip_endpoint option
- .github/scripts/test-kube-vip-manifest.py: extend regression test to cover the
default (apiserver_endpoint) and the override case
Closes#221
* chore(ci): extend molecule job timeout to 3 hours
The default scenario occasionally takes longer than 150 minutes on the shared
nested-virt runner (k3s agent notify-wait can exceed the limit under load), and
a single timeout aborts the whole run before the other four scenarios execute.
Raise timeout-minutes from 150 to 180 so a slow-but-progressing run completes
instead of aborting.
The default scenario remains first in the matrix so a failure surfaces fastest.
* fix(kube-vip): fall back on null kube_vip_endpoint and cover it in the test
- vip_subnet and address use default(apiserver_endpoint, true) so the null
role default falls back to the apiserver endpoint instead of rendering an
empty/invalid address and subnet
- change the manifest regression test default case to pass kube_vip_endpoint
as None so it pins the real runtime null condition and fails fast on this
regression rather than timing out in CI
k3s recommends swap be disabled on every node. Add a disable_swap toggle
(default true) to the prereq role that turns swap off now and comments out the
swap entries in /etc/fstab so swap stays off across reboots. The change is
applied uniformly across all k3s_cluster hosts (all-or-nothing) since leaving
swap on for some nodes but not others creates uneven scheduling and latency.
- roles/prereq/defaults/main.yml: add disable_swap: true default
- roles/prereq/tasks/main.yml: idempotent swapoff -a and fstab comment-out block
gated on disable_swap
- inventory/sample/group_vars/all.yml: document the new sample variable
- README.md: document the disable_swap option
- .github/scripts/test-disable-swap.sh: regression test
- .pre-commit-config.yaml: wire the test into pre-commit
Closes#670
Add a preflight assertion to site.yml that confirms every host in the
k3s_cluster group reports a unique hostname. k3s registers each node keyed by
its hostname, so duplicate hostnames prevent nodes from joining the cluster
and are hard to troubleshoot. The check fails fast in the Pre tasks instead of
surfacing as a cryptic registration failure later.
- site.yml: assert (cluster_hostnames | unique | length) == length for the
k3s_cluster group, skipping hosts outside that group
- README.md: document the unique-hostname requirement
- .github/scripts/test-unique-hostname-precheck.sh: regression test
- .pre-commit-config.yaml: wire the test into pre-commit
Closes#636
- roles/k3s_server_post/templates/metallb.crs.j2: render spec.interfaces in the
L2Advertisement when metal_lb_interfaces is a non-empty list, so MetalLB only
announces on the configured interfaces. Empty list (default) keeps announcing
on all interfaces, preserving existing behavior.
- roles/k3s_server_post/defaults/main.yml: add metal_lb_interfaces: [] default
- roles/k3s_server_post/meta/main.yml: add metal_lb_interfaces argument_spec
- inventory/sample/group_vars/all.yml: document the new sample variable
- .github/scripts/test-metallb-interfaces.py: regression test rendering the
template for empty/single/multiple interfaces and confirming the BGP path is
unaffected
- .pre-commit-config.yaml: wire the new test into pre-commit
Co-authored-by: Leo <leo@kuboschek.me>
Add a `cilium_envoy` variable (default true, matching upstream Cilium 1.20
which installs Envoy by default) that controls whether the Envoy proxy is
deployed for Cilium L7 policies. Pass it to Helm as `envoy.enabled` so users
with no L7 policies can skip Envoy to save resources.
- roles/k3s_server_post/defaults/main.yml: add cilium_envoy: true default
- roles/k3s_server_post/tasks/cilium.yml: add --helm-set envoy.enabled to the
install/upgrade command, driven by the cilium_envoy conditional
- inventory/sample/group_vars/all.yml: document cilium_envoy sample var
- .github/scripts/test-cilium-envoy-toggle.py: regression test asserting the
install command carries the envoy.enabled helm-set and renders true/false
- .pre-commit-config.yaml: wire the new test into pre-commit
Co-authored-by: Léo Nonnenmacher <leo@nonnenmacher-logel.fr>
* chore(deps): apply dependency updates in one combined change
- Bump ansible-core to 2.19.11 and jmespath to 1.1.0 in requirements.in
- Regenerate the Python 3.11 pip-compile lock in requirements.txt
- Bump molecule-plugins to 23.6.0 while keeping molecule on the stable 6.x
series (avoids the molecule-plugins 26 major jump that broke vagrant module
resolution in CI)
- Bump ruamel-yaml-clib to 0.2.15
- Bump the zgosalvez/github-actions-ensure-sha-pinned-actions action to 5.0.6
(SHA-pinned) in lint.yml
* fix(server): make log_destination conditional boolean for ansible-core 2.19
- The always block's 'Save logs of k3s-init.service' task used when: log_destination
where log_destination is a path string derived from an env var
- ansible-core 2.19 rejects string-derived conditionals; evaluate the path as a
real boolean (non-empty) check so the conditional is a true boolean
- Required to keep the k3s_server role working with ansible-core 2.19.11 (the
dependency bump in this change)
* fix(verify): coerce regex_search assertions to bool for ansible-core 2.19
- ansible-core 2.19 requires assert conditionals to be boolean; regex_search
returns a string, which is now rejected
- Wrap all regex_search results used in assert.that with | bool so the calico,
cilium, metallb, and kube-vip image-tag checks produce boolean results
* fix(verify): use boolean is regex_search test instead of | bool
- | bool on a regex_search result coerces a tag string like v0.16.0 to False
in ansible-core 2.19, failing the image-tag assertions
- Use the is regex_search test which returns a real boolean without string
coercion for the calico, cilium, metallb, and kube-vip image assertions
* fix(verify): use is not none for regex_search assertions
- ansible-core 2.19 has no "is regex_search" test and rejects bool string
coercion, so use the regex_search filter with an "is not none" comparison,
which yields a real boolean for the image-tag assertions
- Applies to calico, cilium, metallb, and kube-vip image checks
* fix(metallb): retry transient apiserver resets in config tests
- The Layer 2 and BGP final configuration checks ran a kubectl get per
resource with no retry, so a transient connection refused from the kube
API could abort converge
- Mirror the download_retries/download_delay retry pattern used by the
'Wait for MetalLB resources' task so these checks survive api server
resets on slow runners
- Correct the Deploy metallb manifest/pool when condition so MetalLB is
installed whenever kube-vip does not own the VIP range and Cilium BGP
is disabled
- The previous guard (cilium_bgp is not defined or cilium_iface is not
defined) skipped MetalLB whenever cilium_iface was set, breaking the
cilium + MetalLB scenario
- Use cilium_bgp | default(false) | bool to stay safe when Cilium vars are
not in scope (#644) while still deploying MetalLB for non-BGP Cilium
- Retry the converge-side MetalLB resource wait so a transient kube API
ServiceUnavailable does not abort the converge play
- Add a regression test that evaluates both when conditions across flannel,
calico, non-BGP cilium, BGP cilium, and kube-vip scenarios
ci: skip CI for Dependabot pull requests
- Add an actor guard to the CI workflow jobs so automatic Dependabot PRs
do not consume the shared self-hosted runner
- Dependabot CI runs need maintainer approval instead of auto-running
* fix(flannel): default the interface to each host's default IPv4 interface
- Replace the hardcoded flannel_iface: eth0 with a per-host default derived
from ansible_facts.default_ipv4.interface
- KVM/cloud hosts that are not named eth0 (e.g. enp1s0, ens3) now resolve the
interface automatically instead of failing the k3s_node_ip lookup
- Update the commented calico_iface / cilium_iface examples to match
- Co-authored-by: Fritz Dunkel <677609+FinalDoom@users.noreply.github.com>
- Fixes#621
* test(flannel): add regression test for per-host interface default
- Add a focused test that renders the sample inventory's flannel_iface
expression against fake ansible facts
- Assert a non-eth0 host (enp1s0, ens3) resolves its own interface and that
the expression defaults from ansible_facts.default_ipv4.interface
- Wire it as a local pre-commit hook (default-interface-test)
* test(molecule): assert nodes register a non-loopback InternalIP
- Add a flannel-scenario verify assertion that every node reports an
InternalIP derived from its configured flannel_iface
- Guards against k3s binding to 127.0.0.1 instead of the cluster interface
- Complements the pre-commit default-interface test with an end-to-end check
- Add canonical agent and contributor documentation.
- Modernize issue and pull request templates.
- Skip CI for documentation and issue-template-only changes.
* fix(calico): support split CRDs for current releases
- Download the v1_crd_projectcalico_org.yaml bundle before the operator
- Apply both files with server-side apply and force-conflicts per the
upstream upgrade procedure
- Wait for the operator Deployment and for the managed CRDs to be
Established after the operator starts
- Replace the create/rescue/replace flow with an idempotent apply that
no longer conceals partial failures
- Verify TigeraStatus for calico and apiserver is Available, not just
that Pods exist
* feat(dependencies): upgrade supported cluster components
- Bump K3s to v1.36.2+k3s1, Calico to v3.32.1, Cilium to v1.20.0,
kube-vip to v1.2.2, kube-vip cloud provider to v0.0.12, and MetalLB to
v0.16.0 across sample inventory, role defaults, and argument specs
- Pin the Cilium CLI with a new cilium_cli_tag (v0.19.7) instead of the
floating stable.txt lookup
- Replace the CiliumBGPPeeringPolicy v2alpha1 BGP template with the
v2 CiliumBGPClusterConfig, CiliumBGPPeerConfig, CiliumBGPAdvertisement,
and CiliumLoadBalancerIPPool resource set
- Move Cilium load balancer Helm keys from bpf.loadBalancer to the valid
top-level loadBalancer path
- Add preflight schema validation and remove the deprecated policy after
the v2 objects are accepted
- Wait for cilium status after installation
- Pin kube-vip RBAC in a repository template instead of fetching a
mutable URL, and include EndpointSlice permissions
- Fix the kube-vip bgppeers format to address:ASN comma-separated peers
- Fail clearly when the MetalLB speaker tag replacement does not apply
- Drop the obsolete MetalLB webhook service name version branch
* test(molecule): verify upgraded cluster components
- Assert every node reports the expected K3s kubelet version
- Verify the active CNI (Flannel / Calico / Cilium) is Ready and runs
the expected image tag, including Calico TigeraStatus Available
- Verify the active load balancer (MetalLB / kube-vip) runs the expected
image tags and that MetalLB is absent when kube-vip is active
- Assert no Flannel DaemonSet remains when Calico or Cilium is enabled
- Assert the example LoadBalancer address falls inside the configured
pool range
- Add a manifest-only Cilium BGP regression test that renders the v2
template with zero, one, and multiple neighbors and rejects any v2alpha1
or CiliumBGPPeeringPolicy output
* fix(dependencies): correct dependency version pins
- Set the sample kube-vip image to v1.2.2 and repair the damaged comment
- Pin the kube-vip cloud provider default to v0.0.12 in the task URL
- Set the MetalLB controller argument-spec default to v0.16.0
- Restore the MetalLB available timeout default to 240s
* docs(dependencies): document current cluster versions
- Update kube-vip, kube-vip cloud provider, and MetalLB defaults
- Add cilium_tag and cilium_cli_tag rows
- Explain that MetalLB v0.16.0 is the application image target even though
a newer chart-only tag (metallb-chart-0.16.1) exists
- Add an existing-cluster upgrade warning covering the K3s etcd 3.5.26
bridge and one-minor-at-a-time rule, consecutive Cilium minor upgrades,
Calico v3 resource UID handling, and MetalLB app vs chart tags
* fix(dependencies): address PR review findings
- Read the MetalLB speaker tag check from the managed host with slurp
instead of a controller-side file lookup, and match the full image
reference
- Restore the tigera-operator namespace on the Calico operator Deployment
wait while keeping the managed CRD waits cluster-scoped
- Make Molecule verify inputs durable and scenario-specific via a
per-scenario verify-vars.yml, driven by explicit verify_cni/verify_lb
values instead of non-persisted converge facts
- Rename the kube-vip multi-peer BGP env var from bgppeers to bgp_peers
and vip_cidr to vip_subnet so v1.2.2 actually reads them
- Map the legacy Cilium routed mode to tunnel and stop passing the alias
directly to the chart
- Use return-code based failed_when on apply and preflight commands so
non-error failures are no longer treated as success
- Clarify the sequential K3s upgrade path and backups in the README
- Add kube-vip and MetalLB regression tests and a Cilium mode mapping unit
* fix(dependencies): resolve re-review findings
- correct the Calico TigeraStatus resource kind\n- document tunnel as the supported Cilium routing mode\n- validate load balancer addresses across range and CIDR pools
* fix(molecule): verify embedded flannel instead of a flannel DaemonSet
- K3s 1.36 runs flannel embedded in the k3s agent rather than as a
kube-flannel-ds DaemonSet, so the flannel verifier queried a workload
that no longer exists and failed the verify step
- For the flannel scenarios, assert every node is Ready and that neither
the Calico nor the Cilium namespace exists
- Drop the now-invalid kube-flannel-ds DaemonSet assertion
* fix(molecule): wait for the LoadBalancer address before asserting reachability
- The nginx LoadBalancer service had no ingress address when the
reachability assertion ran, so status.loadBalancer.ingress[0].ip was
undefined and the ipwrap filter failed during verify
- Poll the service until MetalLB or kube-vip assigns an external IP
- Record the assigned address once and reuse it for the reachability probe
and the pool membership checks
* fix(ci): harden calico apiserver wait and extend molecule job timeout
- Bump calico system resources wait retries 30->60 and delay 7->10 so the
slow-to-reconcile calico-apiserver deployment has enough time under nested-virt
- Raise the molecule step timeout-minutes from 90 to 150 to accommodate
contended 5-node scenarios (cilium, kube-vip) that were hitting the 90-min cap
* fix(calico): treat optional API server as best-effort on converge
- The Calico API server (calico-apiserver) is an optional add-on for managing
Calico policy through the projectcalico.org/v3 Kubernetes API; it is not
required for Calico CNI data plane operation
- With Calico v3.32.1 on K3s 1.36 the tigera-operator never provisions the
calico-apiserver namespace, causing the converge wait to fail deterministically
- Keep the strict wait for core Calico components (typha, kube-controllers,
calico-node, csi-node-driver) and make the API server wait tolerate failure
- Restrict the TigeraStatus Available check to the calico status, matching the
upstream v3.32.1 K3s quickstart which validates without the API server
- Use published Bento boxes for Ubuntu 26.04, Debian 13, and Rocky Linux 10.1.\n- Pin exact VirtualBox amd64 box versions in CI.\n- Remove the obsolete Ubuntu SSH workaround and update fixtures.
- Match the minimum CPU and memory used by Ubuntu Molecule platforms
- Restore Vagrant's normal key insertion after password authentication
- Cover the canonical prewarm configuration in the fixture
- Use the Molecule Ubuntu SSH settings when creating its linked-clone master
- Bound a prewarm SSH failure to ten minutes
- Verify the generated Vagrantfile in the master-preparation fixture
- delegate cgroups for transient K3s server units\n- verify inventory node registration without legacy role labels\n- wait for bootstrap CRDs before replacing the transient service
- Retry transient k3s release downloads with bounded backoff.\n- Bound failure diagnostics and validate guest release connectivity.\n- Discover repository-owned Molecule state under the actual project root.
- Update checkout, setup-python, cache, upload-artifact, and pin enforcement actions\n- Keep every third-party action pinned to an immutable release commit\n- Replace the deprecated cache action reference that blocked CI
Timothy Stewart
dependabot[bot]