mirror of
https://github.com/cozystack/cozystack.git
synced 2026-10-29 08:07:33 +09:00
feat/cozystack-controller-gpu-only
80
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Aleksei Sviridkin
|
da28ea7ee8 |
ci(e2e): run nightly and release-tag e2e on the container lane
Nightly and the rc gate still booted the management nodes as QEMU guests, the substrate the merge-gating lanes left behind. On it the tenant Kubernetes suites have missed their node-join deadline on most nights since 2026-10-02, each failure spends the whole deadline, and the job reaches its timeout before the alphabetically later suites run. Over the same days those suites passed on every PR run that selected them, where the nodes are Talos containers and a tenant worker sits one virtualization level shallower. Both lanes now use the PR lanes' container substrate. Nightly's e2e no longer waits for the nocloud disk, which it does not boot; the disk is still built and published. A manual e2e-tag dispatch against a tag that predates the container lane fails early and names the ref to dispatch from instead. DRBD, the replicated StorageClass and the Cozystack Talos node image lose their last e2e coverage with this. The docs and comments that pointed at nightly and the rc gate as that coverage now say so. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
8cb1023788 |
fix(e2e): fail on a failed release of cilium and csi as on the parent
The history guard covered only the parent Kubernetes HelmRelease. A failed action on one of the addon releases it installs left no mark the suite inspected, so a run surfaced only the downstream symptom, typically the node-join deadline, with nothing naming the release behind it. The tenant CNI and CSI are the sharp cases, since they carry a cluster-wide precondition. The guard now reads every addon release of the parent, selected from the cluster by name prefix so a release added later is covered too. A "failed" or "uninstalled" Snapshot fails the run on a release whose install strategy is RetryOnFailure: the parent, which the aggregated apiserver stamps that way, and cilium and csi. The same footprint on the other addons, which pair the default strategy with retries: -1, is printed as a note and leaves the run green. The footprint is not split into a teardown and a retried failure. In helm-controller v1.5.0, the version the platform installs, and unchanged in v1.5.1, a release with an active retry strategy never remediates: atomic_release.go returns an upgrade for a failed release before it reaches the remediation switch (lines 465-475), and the two other paths that uninstall, a changed release target and deletion, both call ClearHistory afterwards. On a RetryOnFailure release an "uninstalled" Snapshot is therefore unreachable, and the "failed" one is the only signal. A retry that later recovered still fails the run, by design, with one blind spot the statuses cannot close: a failed install that the very next attempt recovers is marked superseded by Helm, because there is no deployed release to supersede instead, and its history reads like an ordinary upgrade. The guard runs after the assertions, while every deadline a failed addon breaks ends the script earlier, so the same classification is also printed without a verdict from the node-join collector and from the exit handler. That handler is now armed at the tenant CR apply, so the management-cluster waits before the tenant answers are covered too. Every read is bounded and never folded into "no history", and the parent must read back RetryOnFailure: an empty value there means the jsonpath stopped resolving, and any other value means the apiserver stamp is gone, either of which would turn its footprint into a note. The history read takes its jsonpath from HELMRELEASE_HISTORY_JSONPATH, and the upstream-type pin accepts the bounded wrapper the read goes through. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
cb7a579065 |
fix(e2e): compare LINSTOR pool datasets, not free bytes, at cleanup
For a node with only thick ZFS pools, `linstor sp l` reports the free capacity cached at the satellite's last report rather than asking the satellite, and that report can land while `zfs destroy` is still freeing blocks in the background. A cleanup that had removed every volume it created could then wait out its whole budget against a figure that never moved. That residue has reached about 0.9 GiB, with no known upper bound, against 1 GiB for the smallest volume a suite is known to have leaked, so a byte tolerance has no safe value between the two. Record the datasets under each node's pool before the suite, read with `zfs list` on that node's satellite, and wait until no dataset outside that set remains. A leaked volume is then reported by name and node, and a destroy still freeing space in the background does not count. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
64b3c9619f |
fix(e2e): hold replicated-lane LINSTOR cleanup to a pre-suite baseline
The kubernetes suites' cleanup in the replicated lane demanded 90 GiB free on every LINSTOR node. Volumes that earlier suites and the platform legitimately keep can hold a node under that figure even when the suite reclaimed everything it used, so the cleanup failed suites that had leaked nothing. Both lanes now record the per-node baseline before the suite creates anything and compare against it on cleanup. The 90 GiB figure is what a replicated suite needs to place its workers and CDI scratch volumes, so it stays as a precondition checked before the baseline, and a shortfall there is reported as the sandbox's state rather than the suite's. A run that fails before recording its baseline used to leave the previous suite's file in place for its cleanup to compare against; the file is now removed when the suite starts. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
312409ecd2 |
test(e2e): fail the suite whose pods outlive its cleanup (#4492)
## What this PR does A Chainsaw suite could pass while the workloads it created kept running after its own cleanup, and nothing noticed until a later suite failed. On one E2E run the brokers of four Kafka releases from suites that had passed and deleted their applications (#4488), plus a TenantControlPlane sitting on its finalizer with two apiserver pods (#4489), pushed pod memory requests to the node limit. Harbor then failed on `Insufficient memory` two suites later, under the wrong name. This makes the suite that leaked fail, and name what it left behind. - Chainsaw v0.2.15 has no hook that runs after a test's cleanup: its `Configuration` carries `error.catch`, which runs only on failure. So `make test-chainsaw` now runs `hack/e2e-chainsaw-run.sh`, which invokes `chainsaw test` once per suite directory and checks the cluster between invocations with `hack/e2e-chainsaw/_lib/leak-guard.sh`. - Attribution: a pod is charged to a suite when all three hold. It is not terminal. Its UID was not in the pod list taken before the suite. The root of its ownerReferences chain was created at or after the suite started. A platform pod replaced mid-suite by a rollout, an eviction or a crash traces back to an object the install created, so it is not charged. A pre-existing pod adopted by a controller recreated mid-suite is excluded by the UID baseline. A leak is charged once, to its own suite. - The suite start is read off an apiserver, from the `creationTimestamp` of a server-side dry-run create, rather than off the sandbox. With several control-plane nodes the reading and a root's timestamp can come from different apiservers, so the rule assumes their clocks agree more closely than the gap between the reading and the suite's first object; no skew margin is added. Source for that behaviour: the generic registry `Store.create` stamps it through `rest.FillObjectMetaSystemFields`, then `DryRunnableStorage.Create` copies it back without persisting (`staging/src/k8s.io/apiserver/pkg/registry/generic/registry/store.go` and `dryrun.go` at Kubernetes v1.33.0). - Budget: charged pods get 300s to disappear, the default `timeouts.cleanup` in `.chainsaw.yaml`; a unit test keeps the two equal. A suite that overrides its own cleanup timeout (kafka sets 8m) is covered the same way, because the check runs only after `chainsaw test` has returned, which is after every cleanup in the suite has finished under its own timeout; the 300s start from there. The leaks this targets sat about 110 minutes (Kafka brokers) and 30+ minutes (the TenantControlPlane) past their suites. Every green suite logs `leak-guard: suite=<dir> clear after Ns`, so the number can be checked against later runs. - Past the budget the run prints `LEAKED-PODS suite=<dir> budget=300s count=N`, then one line per pod with its phase, deletion state, root and `app.kubernetes.io/instance`. It raises a `leaked-pods` error annotation and adds a failed `leak-guard` test case to `chainsaw-report.xml`. The run continues through the remaining suites and exits non-zero at the end, as the single invocation did on a red test. - A suite that already failed on its own is checked once without the wait. Its leftovers are reported as a `::warning` with `LEAKED-PODS suite=<dir> checked-once=chainsaw-failed`, since some may still be tearing down. - Each pod gets one `kubectl wait` inside the remaining budget, and nothing is retried. A timeout or a spent budget gives the verdict from the last good pod list, also when the re-list after it fails, so a leak fails the suite. A wait that fails only after running for its whole `--timeout` counts as a timeout whatever it printed: kubectl maps an expired context to "timed out waiting for the condition" only inside the watch, so a deadline that passes during its initial lookup surfaces as "context deadline exceeded" (`wait.go` and `delete.go` in kubectl v1.33.0). A spent budget is caught before calling kubectl and never passed as `--timeout=0`, which means "check once". - Any other error on a wait or a re-list, before the budget is spent and before any timeout, stops the waiting and is reported as `LEAKED-PODS suite=<dir> count=N inconclusive="<error>"`. That is a warning, not a leak failure, so an apiserver blip cannot become a leak red. - A failed first reading after a suite, including a pod list that cannot be parsed, leaves the suite unchecked. It fails with a `leak-guard` annotation and a report case saying the cluster could not be read, not that it leaked. Errors that are not leaks carry the annotation title `leak-guard`, not `leaked-pods`. - Known leaks are listed in `hack/e2e-chainsaw/_lib/leak-allowlist.txt` as `<suite-dir> <issue-ref>`: `kubernetes-latest` and `kubernetes-previous` against #4489 (both suites create a TenantControlPlane through `_lib/run-kubernetes.sh`). `kafka` and `kafka-metadata` were listed against #4488 until #4490 fixed it; those entries are gone. An allowlisted suite prints its leak as a `::warning` and does not fail. An entry that names a directory that is not a suite stops the run before the first suite, so a rename cannot keep an exemption alive. The PR fixing each issue deletes its lines. Suite selection is unchanged. The runner runs the named directories when there are any, and otherwise every directory directly under `hack/e2e-chainsaw/`, each handed to Chainsaw's own discovery; it refuses a tree with a test at its top level, where that split would stop being exact. I verified this with Chainsaw v0.2.15's own `discovery.DiscoverTests`, built from source at that tag. For the full tree, for `kafka redis harbor` and for `kubernetes-latest`, the union of the per-suite discoveries equals what a single `chainsaw test [suites]` discovers: 57, 7 and 2 tests, identical lists. 57 matches the test count in an earlier full run's report. A unit test pins that the runner passes chainsaw the suites and flags it was given, in order. The order changes too: suites used to run in Chainsaw's own order and now run one directory at a time in alphabetical order, which moves the kubernetes suites after the kafka ones. Before #4490 fixed #4488 this starved them (see the E2E criterion below). With the fix merged the order has no effect, and a future leak that lands before a memory-heavy suite is named at the suite that leaked. One behaviour does change: a selected name that is not a directory used to fail the single invocation before any test ran; now only that suite fails and the others run. `chainsaw-report.xml` keeps its path and shape: one `<testsuites>` root with the summed counts, one `<testsuite>` per suite directory. Its consumers are the `Collect chainsaw report` / `Upload chainsaw report` steps in `pull-requests.yaml`, `e2e-fork.yaml`, `nightly.yaml` and `e2e-tag.yaml`, which `docker cp` the file and upload it as an artifact. Nothing in the repository parses it. The merge is unit-tested on two per-suite reports in Chainsaw's real shape, one of them failed, and the result is checked to parse as XML. Tests: `hack/chainsaw-leak-guard_test.bats` (31 tests, in `make unit-tests`; named without the `e2e-` prefix, which the unit lane excludes) covers the attribution rule, the budget, the allowlist, the merge and the runner, all against stub `kubectl` and `chainsaw`. Every rule was mutation-checked: removing the guard call, having the Makefile bypass the runner, inverting the allowlist match, flipping or tightening the root-timestamp comparison, removing the UID baseline, removing the terminal-phase filter, skipping the budget wait, changing the budget, lowering or raising the depth cap, accepting a stale allowlist entry, reading an unreadable pod list as empty, dropping the leak case from the report, skipping directories in the selection, treating any failed owner read as a missing owner, dropping the separate report for a suite that already failed, dropping the not-run reason from the report, reading any wait error as a timeout, removing the inconclusive branch, letting a failed re-list after a timeout turn a leak inconclusive, ignoring a failed re-list, reporting an unreadable cluster as a leak, reading an unparsable pod list as clean, reporting a kubectl warning instead of the error, passing the report name without its extension, wording a failed re-list as a failed wait, dropping an allowlist's last line when it has no newline, dropping the check that `--set-string` carries a value, and reading a wait that ran its whole timeout but printed a context error as inconclusive. Each one turns its named test red, and restoring turns it green. E2E criterion, registered before the run: in the `Run E2E tests` log of this PR, every suite that ran prints `leak-guard: suite=<dir> clear after`, a `LEAKED-PODS suite=<dir>` line carrying `allowlisted=`, or one carrying `checked-once=` (a suite that failed on its own) or `inconclusive=` (waiting stopped on a failed read), and no line matches `LEAKED-PODS suite=.* budget=`. With #4490 merged, `kafka` and `kafka-metadata` are no longer allowlisted and must print `leak-guard: suite=kafka clear after` and `leak-guard: suite=kafka-metadata clear after`, which is the live check that #4490 fixed #4488. The first run of this PR (before #4490, run 36223896793) is the positive control: it printed `LEAKED-PODS suite=kafka count=15` and `suite=kafka-metadata count=8`, allowlisted, and every other suite cleared in 0-37s. That run also went red on both kubernetes suites, because those 23 leaked brokers left no memory for the second tenant worker VM (`0/3 nodes are available: 3 Insufficient memory`), which the order change described above made possible. The guard has not yet run against a real cluster, so the first run of this PR should carry the `full-e2e` label; its `clear after` lines are the first measured teardown times, and any unexpected `LEAKED-PODS` line is either a real leak or a false positive to understand before merge. Follow-ups, known and out of scope here: - The allowlist exempts a whole suite, so a new leak in an allowlisted suite stays hidden until the issue's fix removes the line. - Attribution is per suite directory, not per Chainsaw `Test`; the leak lines carry the release instance, which usually names the test. - Only pods are checked. A leaked PVC, LoadBalancer IP or bucket holds no node memory and is not seen. - The 300s budget rests on Chainsaw's cleanup timeout, not on a measured pod-gone latency; the `clear after` lines are where that measurement will come from. - A platform object recreated mid-suite (for example a HelmRelease uninstall/reinstall remediation, or a force-replace on an immutable field) has a root newer than the suite start, and its pods would be charged to whichever suite was running as a leak; the first CI runs are where to look for this. - The per-pod wait recomputes the remaining budget before each pod, and no unit test pins that. - `kubernetes-previous` is allowlisted for #4489 although the stuck finalizer was seen only in `kubernetes-latest`; both suites create a TenantControlPlane the same way, and the exemption hides any new leak there until #4489 is fixed. - Each clean suite now waits for its own pods to finish terminating before the next starts, where the tail used to overlap the next suite. The first full run's `clear after` lines say what that costs against the E2E job's 215-minute cap. - After a timeout followed by a failed re-list, the leak lines come from the list taken before the wait, so a pod whose own wait had already succeeded is still named and the count can be too high; the verdict is unaffected. - Likewise, when every wait succeeded and the re-list then fails, the inconclusive report still lists the pods whose deletion was just confirmed. In the boundary case where the budget runs out during the last successful wait and the re-list then fails, the result is inconclusive rather than a verdict from the last good list, because the guard only treats the budget as spent when it finds it exhausted at a check. - A mistyped `COZY_LEAK_ALLOWLIST` path passes allowlist validation; it fails closed, since no suite is then exempt. - An allowlisted suite is not waited on, so `kubernetes-latest` and `kubernetes-previous` will print `LEAKED-PODS ... allowlisted=` on most runs even for pods still in normal teardown; waiting out the budget first and applying the allowlist only to what remains would stop that noise. - A slow teardown in each of several suites can add up to 300s apiece, against the E2E job's 215-minute cap. On the first full run of this PR (run 36223896793) the `clear after` values of the suites that were waited on summed to 210s; the last full run before this change took 158 of the 215 minutes. - A pod whose owner's resource type is no longer served gets "the server doesn't have a resource type" rather than NotFound, so its suite is reported unchecked instead of leaked; it still fails. - `kubectl wait --for=delete` waits for deletion, while the guard skips pods in `Succeeded` or `Failed`; a pod that finishes but is never deleted (a Job under a leaked root) spends the rest of the budget before the suite shows `clear`. The verdict is right; only job time is lost. - The unit tests do not pin the one-line empty report Chainsaw writes for a directory with no tests (`_lib`), nor the branch that catches a spent budget before calling `kubectl wait`; both are verified against the Chainsaw and kubectl sources rather than by a test. - The runner's per-suite report directory from `mktemp -d` is not removed; the sandbox is discarded after the job. - Each non-suite directory under `hack/e2e-chainsaw/` (today `_lib`) costs one empty Chainsaw invocation, because the runner leaves the decision of what is a test to Chainsaw. ### Screenshots Not a UI change. ### Downstream repositories - [x] No downstream repository is affected by this change - [ ] [cozystack/website](https://github.com/cozystack/website) - follow-up: - [ ] [cozystack/terraform-provider-cozystack](https://github.com/cozystack/terraform-provider-cozystack) - follow-up: - [ ] [cozystack/ansible-cozystack](https://github.com/cozystack/ansible-cozystack) - follow-up: - [ ] [cozystack/ccp](https://github.com/cozystack/ccp) - follow-up: - [ ] [cozystack/talm](https://github.com/cozystack/talm) - follow-up: - [ ] [cozystack/cozyhr](https://github.com/cozystack/cozyhr) - follow-up: - [ ] [cozystack/cozy-proxy](https://github.com/cozystack/cozy-proxy) - follow-up: - [ ] [cozystack/cozystack-telemetry-server](https://github.com/cozystack/cozystack-telemetry-server) - follow-up: - [ ] [cozystack/external-apps-example](https://github.com/cozystack/external-apps-example) - follow-up: - [ ] [cozystack/examples](https://github.com/cozystack/examples) - follow-up: - [ ] [cozystack/community](https://github.com/cozystack/community) - follow-up: The diff adds files under `hack/`, changes the `test-chainsaw` target, and rewords one comment and the failure line in the four E2E workflows. The trigger map's `hack/` entry is for cozystack/ccp, whose skills anchor on `hack/package.mk`, `hack/common-envs.mk` and `make generate`, none of which change here. A GitHub code search for `chainsaw` across the cozystack organisation finds hits only in this repository. ### Release note ```release-note test(e2e): the E2E app suite now fails the suite whose pods outlive its own cleanup by more than 300s, naming each leftover pod and its owner; suites with a filed leak are listed in hack/e2e-chainsaw/_lib/leak-allowlist.txt and only warn. ``` |
||
|
Aleksei Sviridkin
|
5b4380dd3b |
fix(e2e): delete the failed S3 preflight object through a live tunnel
The exit trap stopped the port-forward before running mc rm, so the delete went to a closed port and failed every time, leaving the object behind and printing a warning that said nothing about why. The delete now runs first and the warning carries mc's own error, or says the delete timed out when mc was killed before it could write one. A PUT that timed out is now also treated as uploaded: it can land and still outlive its time bound, and the trap never tried to remove such an object. A PUT the server refused is not, so the common case of access that never went live does not end in a delete doomed to the same refusal. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
b6cd6ab682 |
fix(e2e): print the port-forward log when the S3 preflight client fails
mc alias set reaches the endpoint through the port-forward, so a tunnel that came up and then broke is a likely cause of that failure, and its log is the only record of it. The readiness and retry-timeout exits already print it; this one threw it away. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Marian Koreniuk
|
7500a9fbae |
test(kubernetes): add the Proxmox e2e suite, parked disabled
Upstream CI has no hypervisor, so this suite cannot run there; it ships
`.disabled` and is run by a lane that owns a Proxmox stand.
Named for what the chart actually creates. The resource definition prefixes the
release, so a Kubernetes CR called k8s-e2e produces
ProxmoxCluster/kubernetes-k8s-e2e in the tenant's own namespace, and the admin
kubeconfig Secret is kubernetes-k8s-e2e-admin-kubeconfig.
Stand-specific values are variables rather than hardcoded. Every one of them is
a bare ${COZY_PVE_*} that run-proxmox.sh expands with envsubst, supplying the
defaults itself — GNU envsubst leaves a ${VAR:-default} in the text exactly as
it is, set or not, and that reaches the apiserver as a string where the schema
wants an integer — and refusing a fixture with a variable still in it. The
StorageClass and the PVC probe read the same COZY_PROXMOX_STORAGE, so the volume
lands where the probe looks. The runner reads the tenant kubeconfig's .conf key,
the address a runner outside the management cluster can reach.
The stand's resource pool comes in through COZY_PVE_POOL, empty by default.
hack/proxmox-e2e-fixture_test.bats pins the shape of the fixture and the render.
The PVC probe reads the storage again after the claim is deleted: a deleted
claim whose disk is still there is the leak the pre-delete hook's Step 1b
exists to prevent. keep_tenant is honoured by the suite's own teardown, not
only by the workflow's. The workflow carries no fallback for the stand's
variables: an unset one fails the job before a key is sent anywhere. The
schedule runs only where vars.PROXMOX_NIGHTLY is true: the job targets a
self-hosted runner with the `proxmox` label, and a scheduled run on a repository
without one would sit in the queue until GitHub cancels it.
The suite also asserts the half of the deployment that lives in the management
cluster — the CSI controller and the cloud-controller-manager — because that is
the property the pair is designed around and a tenant-only assertion would pass
with the token in the wrong place.
Signed-off-by: Marian Koreniuk <moriarti@cp.if.ua>
Assisted-by: LLM
|
||
|
Aleksei Sviridkin
|
c5cd21af96 |
test(e2e): fail the suite whose pods outlive its cleanup
A Chainsaw suite could pass while its workloads kept running after its own cleanup: brokers of Kafka releases whose application was deleted, a TenantControlPlane stuck on its finalizer holding its apiserver pods. Nothing noticed until a later suite failed on Insufficient memory, hours on and under the wrong name. Chainsaw v0.2.15 has no hook after a test's cleanup, so the suite now runs one `chainsaw test` per suite directory and checks between them. A pod is charged to the suite when it is not terminal, did not exist before the suite, and the root of its ownerReferences chain was created at or after the suite started on the apiserver's clock. That leaves alone a platform pod a rollout, an eviction or a crash replaces mid-suite, though not one whose platform object is itself recreated mid-suite. Charged pods get the default cleanup timeout of the chainsaw configuration, five minutes, counted from the moment chainsaw returns; past that the run names the suite, each pod and its root, and fails. A timeout or a spent budget takes the verdict from the last good pod list; any other read failure before then stops the wait and is reported as inconclusive rather than as a leak; a failed first reading after the suite leaves it unchecked, and it fails. The runner hands chainsaw the tests a single invocation would discover and merges the per-suite JUnit reports into the chainsaw-report.xml the workflows already collect. Suites with a filed leak are exempt by an allowlist that must name existing suite directories, so a rename cannot keep an exemption alive. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
a92e89e969 |
docs(kubernetes): describe the OIDC UID header from v1.33
The tenant OIDC guide, the version field description and the e2e helper comment still described a --feature-gates term rendered on v1.32. Left as they were, they would tell an operator that the chart switches that alpha gate on, which it no longer does. The guide now says what a v1.32 tenant has to write itself to get the UID, and which --emulated-version values fail the render, so an operator who meets that refusal finds it explained. The breaking-changes note in the README said the chart refuses the mixed gate spelling while OIDC is on; with that refusal gone it would promise a render error that no longer comes. The version description returns to its earlier wording in values.yaml and in the update-versions.sh template, because that template rewrites the line on every version bump and would otherwise bring the old text back. make generate carries it into the schema, README, API types and ApplicationDefinition. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
1071802e3f |
test(e2e): pin the release-history read against the upstream type (#3607)
## What this PR does
`hack/remediation-guard.bats` carried a test called "status.history
extraction pins HR v2 status.history shape". It wrote a HelmRelease
document containing the field names it expected, then read that same
document back with the same jsonpath through `yq`. What it established
is that `yq` can read a fixture the test itself composed.
Its own comment named the failure it was there to catch: a future Flux
release renames the field, the jsonpath matches nothing, the guard
reports no remediation cycle, and real cycles pass the e2e assertion. A
rename moves neither the fixture nor the expression, so the test stays
green through exactly that event. The comment also named where the truth
lives, `github.com/fluxcd/helm-controller/api/v2`, and then pinned a
copy of the author's expectations instead.
`internal/fluxcontract` supplies neither half. The expression is read
out of `hack/e2e-chainsaw/_lib/remediation-guard.sh` at test time, so it
is the one that ships rather than a copy. The object is a
`helmv2.HelmRelease` built from the upstream type at the version
`go.mod` holds, serialized and matched through
`k8s.io/client-go/util/jsonpath`, which is the engine kubectl's `-o
jsonpath` uses. A renamed or retyped field is then a compile error or a
failed match, and a guard that stops reading history fails the lookup
with a message saying so.
That is one test. Two more exist so that it reads the expression kubectl
actually receives, because the expression now lives in a named
assignment the test sources and the caller expands: between the value
checked here and the value that runs there is a shell variable, and each
step of that indirection can fail while the suite stays green. The first
follows the caller's own call: it reads the local handed to
`helmrelease_has_remediation_cycle`, requires that local to be assigned
exactly once, and requires that assignment to take the expansion of the
shared name rather than the bare name, which matters because kubectl
echoes a brace-less template back verbatim, so a read that lost its `$`
returns its own text, clears the script's empty-history check, and
matches no status. Starting from the call rather than from the text is
what makes the pin describe the read that runs: asking only whether a
well formed read exists somewhere in the file is satisfied by a second
assignment overwriting the same local, and by a read filling a local the
guard never consumes. The second holds that no shell library but the one
that owns the name assigns it, which does not follow from the first: an
assignment beside the read satisfies the first and still replaces the
value after this file has approved the one the library exported. It
covers every library in that directory rather than the caller alone,
because the caller sources three of them and the owning one first, so an
assignment in any of the others lands later and wins.
Those two match text with a pattern, and each carries both halves of its
contract. A refuse table lists the forms the pattern must reject, and an
accept table lists the forms it must keep accepting, one case per
freedom the code claims: for the call, the `if` prefix, the negation, an
expansion with or without braces; for the read, the spelling of `-o`,
the quoting of the substitution, an expansion with or without braces,
and the local's name, which stays free because the call and the read are
renamed together. Without the accept side those freedoms survive only
until someone tightens the pattern, and nothing reddens when they go.
Two boundaries, named rather than left to be found. The pin holds the
shape of the read, not the vocabulary of the answer: the guard greps for
`failed` and `uninstalled`, and the test supplies those two strings
itself, so a release status renamed upstream leaves the jsonpath
returning strings, the test green because it wrote them, and the shell
grep silently matching nothing. Those values come from Helm's release
status rather than the Flux type, so pinning them reaches past what this
change binds, and there is more than one way to do it. Second, a literal
history read that fills a local the guard never consumes is not caught,
and on its own that is harmless: the guard still runs the shared
expression, and the moment it stops doing so the read filling the
consumed local no longer matches and the caller test fails.
The caller-side half is a heuristic and is meant to be read as one. It
decides whether a piece of text is the read by matching it, and a
matcher that reasons about shell is wrong at some spelling by
construction, which is the same objection the file raises against
reading the library's assignment as text instead of sourcing it.
Sourcing is available for the library and not for the caller, so what is
left is to make the imprecision run one way: the count is of
occurrences, so an extra one reddens and a missed one cannot pass, and a
form outside the accepted set fails with the pattern printed rather than
quietly stopping asserting.
That direction is the whole argument for keeping it. A false red is a
broken build on a legitimate rewrite, and the failure names what it did
not recognise, so it costs a minute. A silent green is a guard that
asserts nothing for as long as nobody looks, which is the state this
branch exists to end. The ways the caller can break without anyone
noticing are worth that trade. A read that lost its `$` hands kubectl a
template with no braces, which comes back as the variable's own name; a
hand-written path that is valid but names the wrong field comes back as
that field's values. Both are non-empty, so the script's own
empty-history backstop stays quiet and the cycle grep matches nothing. A
deleted call is quieter still, since there is then no assertion to fail.
None of the three is visible to the library-side pin, which judges the
value the library hands out and never learns whether it reached kubectl.
What does fail loudly is an expression kubectl cannot use at all, and
that is the one case needing no pin.
The shape the fixed test now has already exists in this tree, in
`hack/seaweedfs-naming-audit.bats`, under "the audit's reconstruction
agrees with the chart helper it mirrors". That one also writes a
heredoc, which is what makes the pair worth comparing: it copies the
real `_naming.tpl` out of the chart, renders it through `helm`, and
checks the audit script's output against the render. The heredoc is a
harness for calling the real helper, not a record of what the author
expected it to say, and the two sides have different authors. A test
that writes both sides proves only that it is self-consistent.
Nothing is lost by the deletion. The removed test's tail fed extracted
statuses into `helmrelease_has_remediation_cycle`, and the seven tests
remaining in the bats file `hack/remediation-guard.bats` still cover
that, including `uninstalled then deployed still returns detected`. The
extraction half is now pinned more strictly than before: the old test
asserted only that the output was non-empty and contained one line, the
new one asserts the exact two bare tokens, so dropping the `{"\n"}`
separator also goes red.
CI runs it. `make test-controllers` is its own step in
`.github/workflows/pull-requests.yaml` and runs `go test ./internal/...
-count=1`, which picks the new package up without registration.
### Screenshots
Not a UI change.
### Downstream repositories
Walked the trigger map against the diff. It adds a Go test package under
`internal/`, deletes one test from `hack/remediation-guard.bats`, adds a
variable to `hack/e2e-chainsaw/_lib/remediation-guard.sh`, points the
kubectl invocation in `hack/e2e-chainsaw/_lib/run-kubernetes.sh` at that
variable, adds a sentence to `docs/agents/e2e-testing.md`, and records
`make test-controllers` in `docs/agents/overview.md`. No package is
added, renamed or removed, and nothing touches a `values.schema.json`, a
version enum, a `values.yaml` default, an `ApplicationDefinition`, a
CRD, a namespace, an annotation or a telemetry metric.
`hack/e2e-prepare-cluster.bats` is untouched, so the node contract that
`talm` and `ansible-cozystack` restate by hand does not move.
The closest call is `ccp`, whose skills anchor on files under `hack/`.
Its trigger is a move, a rename, or a make target that does something
different. This changes the contents of three files, not their names,
and no target behaves differently; `hack/package.mk` and
`hack/common-envs.mk` are untouched.
- [x] No downstream repository is affected by this change
- [ ] [cozystack/website](https://github.com/cozystack/website) -
follow-up:
- [ ]
[cozystack/terraform-provider-cozystack](https://github.com/cozystack/terraform-provider-cozystack)
- follow-up:
- [ ]
[cozystack/ansible-cozystack](https://github.com/cozystack/ansible-cozystack)
- follow-up:
- [ ] [cozystack/ccp](https://github.com/cozystack/ccp) - follow-up:
- [ ] [cozystack/talm](https://github.com/cozystack/talm) - follow-up:
- [ ] [cozystack/cozyhr](https://github.com/cozystack/cozyhr) -
follow-up:
- [ ] [cozystack/cozy-proxy](https://github.com/cozystack/cozy-proxy) -
follow-up:
- [ ]
[cozystack/cozystack-telemetry-server](https://github.com/cozystack/cozystack-telemetry-server)
- follow-up:
- [ ]
[cozystack/external-apps-example](https://github.com/cozystack/external-apps-example)
- follow-up:
- [ ] [cozystack/examples](https://github.com/cozystack/examples) -
follow-up:
### Release note
```release-note
NONE
```
|
||
|
Aleksei Sviridkin
|
97cb2781db |
test(e2e): pin the release-history read against the upstream type
The test that claimed to pin the HelmRelease v2 history shape wrote a HelmRelease document containing the field names it expected, then read that same document back with the same jsonpath. It established that yq can read a fixture the test itself composed. Its own comment described the failure it was meant to catch: a future Flux release renames the field, the jsonpath matches nothing, the guard reports no remediation cycle, and real cycles pass the e2e assertion. A rename moves neither the fixture nor the expression, so the test stays green through exactly that event. The comment also named where the truth lives, github.com/fluxcd/helm-controller/api/v2, and pinned a copy of the author's expectations instead. Replace it with a Go test that supplies neither half. The expression is read out of the shell library that ships it, and the object is built from the upstream type at the version go.mod holds, then serialized and matched the way kubectl does. A renamed or retyped field is a compile error or a failed match; a guard that stops reading history fails the lookup with a message saying so. Both directions were confirmed by mutation. Locating the expression by searching the file is not enough on its own. Inline in a long script, locating it means deciding which occurrences are code and which are prose, and every way of getting that wrong fails green: an example quoted near a broken read satisfies the pin while the guard reads nothing. So the expression moves into a named assignment in remediation-guard.sh, which the guard already sources, and the test sources that library and takes the value the shell is left holding. Deciding which text is the assignment is then not the test's problem: an `export`, an indent, a second assignment later in the file all resolve the way the guard will resolve them, because the thing resolving them is the shell. Sourcing borrows the environment along with the shell, so the snippet unsets the name before it sources. `go test` hands the child its own environment, and a shell that had exported this name would answer in place of a library that stopped assigning it: the value comes back non-empty, the empty check stays quiet, and the pin passes over a guard that reads nothing. That is this file's own subject reaching it through the environment rather than through the source. Moving the expression out of the call site puts a step between what is pinned and what runs, so the call site is checked too, and the check starts from the call rather than from the text: the local the guard is handed must be assigned exactly once, by an anchored assignment whose jsonpath is the expansion of the shared name. That a well-formed read exists somewhere in the file is a weaker claim than it reads as, since a second assignment to the same local, or a read filling another one, satisfies it while the guard runs a literal nothing here pins. The expansion and not the name: kubectl echoes a braceless template back verbatim, so a read that dropped the `$` would return its own text, clear the script's empty-history backstop, and carry no status the cycle check matches. The library is checked by sourcing rather than by reading, so an assignment the shell would run is resolved by the shell instead of by a guess about which text the shell would pick. Each rule is pinned by its own test, so a regression names which one it opened rather than only that extraction broke. How kubectl's printer treats an absent path is its own configuration rather than part of this contract, so it is left unasserted: guessing at it would put the expectation and the fixture back in the same hands. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
4475fb4c17 |
test(e2e): assert the System-mode OIDC uid claim mapping
The System lane read the rendered AuthenticationConfiguration for the issuer URL and the audience but not for the mapping that gives the user a UID. A configuration missing that mapping is still valid, so the tenant control plane starts and the suite passes while every OIDC user authenticates with an empty UID. The unit twin mocks the Secret without the mapping, so dropping the grep from the helper reddens it rather than leaving it unpinned. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
187e4ce12b |
test(e2e): assert the tenant OIDC UID header on the control plane
The render matrix proves the chart emits --requestheader-uid-headers, but nothing checked that the flag survives the path a real tenant takes: aggregated API admission of spec.oidc, the HelmRelease, and the rendered KamajiControlPlane. Assert it in both OIDC helpers, next to the existing --authentication-config check on the same resource. Read the arg list once into a variable and grep the capture twice. The existing check consumed a pipeline, which is single-use, and a second kubectl call would double the management-cluster reads these helpers already make. Only the header flag is asserted. The chart also renders --feature-gates=RemoteRequestHeaderUID=true, but only on v1.32 where the gate is still alpha, while this lane takes the highest version in the map, which is above that. The header flag is the part that holds wherever the flag exists at all. The bats case for the missing flag runs the helper as a standalone subshell under set -e rather than as an if condition. Shells suppress errexit for a command whose status is being tested, and that reaches inside the subshell, so the if spelling reports success and pins nothing. The binding query next to these assertions piped kubectl straight into sort and so returned sort's status. Its callers capture it in a command substitution, where errexit does not reach, so a failed query read as an empty binding list. It now returns kubectl's failure itself. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
65c183e311 |
fix(e2e): keep backup preflight failure evidence across attempts
The backup access preflight printed its port-forward capture only when the port-forward never came up, and truncated its mc log at the top of every attempt. On the path that fails in practice, a healthy port-forward and an S3 round trip that never completes, the output was the tail of the last attempt with nothing to say which stage it died at, which reads as a successful upload followed by a hung download. Print the port-forward capture alongside the mc log on the timeout path, append the mc log across attempts, and record which of PUT, GET, compare or DELETE each attempt reached. Assisted-by: LLM Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Myasnikov Daniil
|
b30c7efaf6 |
fix(e2e): make the OIDC teardown assertions fail closed
The three checks that System-mode artifacts are gone after the mode switch were `if kubectl get ... >/dev/null 2>&1`, which reads every non-zero exit as "the object is gone". An RBAC denial, a timeout or a missing CRD produces the same exit, and stderr went to /dev/null, so the suite reported a clean System teardown for never having observed one. --ignore-not-found is what separates absent from unaskable: absent is exit 0 with empty output, everything else stays non-zero. Preferred over matching NotFound in the message, which depends on wording and, under a runner with tracing on, captures the trace instead of the error. Each probe is pinned individually rather than as a group: the new case breaks one query per iteration while the other two answer absent, so reverting any single probe to the bare form reddens the suite. Verified by doing exactly that three times. A test that broke all three at once would have passed on any one surviving probe, which is how the first attempt at this test fooled me. Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com> Assisted-By: LLM |
||
|
Myasnikov Daniil
|
b131bd56a4 |
fix(e2e): restore the default StorageClass when the fallback render fails
verify_storageclass_fallback_default deletes the `replicated` class, renders with --dry-run=server, and restores it inline "before any assertion exit" -- the comment's words. Under the caller's `set -eu` the render is a plain assignment, so a failed render exits the function there: before rc is read, and before the restore runs. The management cluster is then left with no default StorageClass for every suite after it, which is the opposite of what the comment promises. Captured with `|| rc=$?`. Pre-existing shape, but the restore block is edited on this branch anyway, and the QEMU lanes are where it bites. Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com> Assisted-By: LLM |
||
|
Myasnikov Daniil
|
186547d835 |
test(e2e): run the merge-gating lanes on Talos containers
The srv1-srv3 nodes of both lanes that gate a merge -- the same-repo one in pull-requests.yaml and the fork one in e2e-fork.yaml -- run as Talos containers instead of QEMU guests, so a tenant worker sits at L2 rather than L3. Measured in CI on commits where both substrates ran the same tree: whole-job wall clock down 24-49 minutes, essentially all of it in Chainsaw, while the two tenant-Kubernetes suites go from 1143-2040s to 466-809s and stop depending on the runner's vgif flag. The node-join soft-red gate goes with the substrate, which is what it was for. It existed because nested virtualisation on the shared runners degraded far enough that a worker registering in minutes elsewhere could miss any deadline this test could afford; #3513 established that the discriminator was the host's kvm_amd vgif flag, and removing a nesting level removes what it acted on. So hack/e2e-node-join-soft-red.sh, the SOFT-RED-node-join.txt marker and the soft_red job outputs are gone from all four workflows, and a missed deadline is an ordinary red again. The ::warning annotation survives, because the reason a run went red should be legible without opening the log, and so does the 124-versus-everything -else test, because "the workers were slow" and "the wait never ran" send a reader to different places. Done by swapping the substrate inside the existing e2e jobs rather than by promoting a separate container job. Those jobs carry more than a substrate -- the GitHub App token, the report-overrun guard, the images list, the SSH breakpoint -- and promoting a job that never had them would have dropped four features without saying so. Three properties of container mode drive the design, and each fails quietly rather than loudly. machine.kernel.modules is a silent no-op -- kernel_module_spec.go returns early on ModeContainer with no error and no event -- so the host loads openvswitch and zfs before the nodes start and the workflow step asserts both rather than trusting them; a missing module surfaces an hour later looking like a CNI or a storage regression. Docker gives a privileged container its own tmpfs /dev seeded once at creation, so /dev/kvm works but ZFS zvols created later never appear and linstor-csi fails ControllerPublishVolume; the compose file binds the host devtmpfs, which is also why talosctl cluster create docker can never carry LINSTOR. And there is no ARP VIP, so 192.168.123.11 stands in for the QEMU lane's .10. The QEMU machinery stays in the tree and keeps running: nightly.yaml and e2e-tag.yaml still call `make prepare-env`, so DRBD, the replicated StorageClass, its Immediate binding mode, the tenant StorageClass propagation that only applies to remotely-accessible classes, and the Cozystack Talos node image with its extensions all keep nightly and release-candidate coverage. None of them is exercised per pull request any more. That is the trade. Nothing on the PR path builds or downloads a nocloud disk now, so build-talos drops the talos-nocloud step and the talos-image artifact. build-talos itself stays: finalize consumes its digest fragment. One property this does not fix, recorded beside the runner class because it will be read off a red run eventually: a container node reports the HOST's CPU and memory as capacity, and hack/e2e-container-up.sh counters that with kubelet systemReserved. That corrects the scheduler and not what a pod reads from /proc, so a workload sizing itself from /proc/cpuinfo or /proc/meminfo is configured for the whole runner. Also carried here because the same files carry them: the ghcr.io pull-through mirror removal (see #4007 -- whichever lands first makes the other a no-op), the authoritative HelmRelease readiness gate, the backup-credential preflight, the prepull guard against an image-less container fragment, and folding the two standalone OIDC suites into the latest-version tenant cluster. Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com> |
||
|
Timofei Larkin
|
f5b957ee27 |
feat(kubernetes)!: manage worker pools as KubernetesNodes, adopt on upgrade (Phase 2, breaking) (#3315)
## What this PR does Phase 2 of the Kubernetes-app split ([cozystack/community#8](https://github.com/cozystack/community/pull/8)) — the **breaking** half. Stacked on #3314 (review that first). Removes worker node pools from the `kubernetes` chart (they moved to `kubernetes-nodes` in #3314) and adopts existing pools on upgrade: - `kubernetes` chart is now **control-plane only**: drops the pool `range` block, the `talos-reconcile` Job + RBAC (deleted; moved to `kubernetes-nodes`), the preserve-old-KMT block, the `kubernetes.nodeGroups` helper, and the per-group dashboard RBAC. - **`spec.nodeGroups` is removed from the `Kubernetes` CR** (BREAKING). Worker pools are now separate `KubernetesNodes` resources. - **Migration 54** (bumps `targetVersion` to 55): pre-upgrade hook that, for every `kubernetes-<cluster>` release, creates one `kubernetes-nodes-<cluster>-<pool>` HelmRelease and re-annotates the existing worker objects onto it with `helm.sh/resource-policy=keep`, so the now control-plane-only parent upgrade does not prune running workers. The default `md0` (implicit when `nodeGroups` was empty) is materialised explicitly. Idempotent. Genuine infrastructure/read failures fail closed (`exit 1`, retried on the next upgrade); hostile stored shapes (a pool name over Helm's 53-char release-name cap, an RFC-1123-invalid or whitespace/slash map key, a non-object `nodeGroups`) are warned, pinned, and skipped rather than deadlocking the entire fleet's upgrade. - The `ingressNginx` addon no longer auto-provisions a node (the implicit `md0` default is gone). If enabled without a pool carrying `roles: [ingress-nginx]`, the controller DaemonSet simply schedules nothing (degraded, not failed) — the requirement is documented in the chart NOTES.txt rather than enforced at render time. Because the child chart renders the pool objects **byte-identically** to the old parent render (golden parity in #3314; the KubevirtMachineTemplate content-hash name is preserved), adoption does not churn live worker VMs. ### Migration number `54` is contiguous with main (latest is `53`) and `targetVersion` is `55`. The `52`/#3201 collision noted in earlier revisions has been resolved: this branch was rebased and merged past it, so the numbering is contiguous and `run-migrations.sh` has no gap to hard-fail on. ### Not yet done (blocking draft → ready) - [ ] **dev e2e upgrade validation** (`cozystack-pr-test`): upgrade a custom-`nodeGroups` cluster and a default-`md0` cluster; assert zero worker-VM churn (KMT hash, Machine uid/creationTimestamp stable) and that the adopted child HRs reconcile. - [ ] downstream trigger-map walk (website docs for the `nodeGroups` → `KubernetesNodes` change; terraform-provider for the removed API field). ### Downstream repositories Deliberately unticked (draft) — the breaking API change affects docs and the terraform provider; follow-ups to be filed before ready. - [ ] No downstream repository is affected by this change - [ ] [cozystack/website](https://github.com/cozystack/website) - follow-up: - [ ] [cozystack/terraform-provider-cozystack](https://github.com/cozystack/terraform-provider-cozystack) - follow-up: - [ ] [cozystack/ansible-cozystack](https://github.com/cozystack/ansible-cozystack) - follow-up: ```release-note Worker node pools of a managed Kubernetes cluster are now managed as separate `KubernetesNodes` resources. `spec.nodeGroups` on the `Kubernetes` CR is removed; existing pools are adopted automatically on upgrade. ``` <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * Worker pools are now separate `KubernetesNodes` resources with their own lifecycle. * Added a render-time compatibility guard to prevent worker pools from running ahead of the parent control-plane minor version. * Improved upgrade/uninstall behavior to better preserve already-adopted worker objects. * **Breaking Changes** * Removed per-pool `nodeGroups` and worker node health settings from the `Kubernetes` chart/CR; pool configuration now lives in `KubernetesNodes`. * Ingress requires at least one `KubernetesNodes` pool with `roles: [ingress-nginx]`. * `cluster` and `storageClass` are immutable after creation; pool `version` may lag but must not be ahead of the parent minor. * **Documentation** * Updated READMEs, notes, and schemas to reflect the `KubernetesNodes` split and the immutability/version rules. <!-- end of auto-generated comment: release notes by coderabbit.ai --> |
||
|
Aleksei Sviridkin
|
ad62a8d62b |
test(e2e): stop blocking merges on the tenant node-join deadline
The tenant node-join wait no longer discriminates. Nested virtualisation on the shared runners degrades far enough that a guest which registers in minutes on real hardware can miss any deadline this test can afford, so a red on that one wait says as much about the machine the run landed on as about the product, and it teaches a reader to re-run the job. The suite still fails on it. That is what keeps Chainsaw's catch running: the previous-instance container logs, the host snapshot, the data-plane capture and the event dump are all keyed on the test failing, and the report stays honest. What changes is one layer up. The wait writes a marker beside its diagnostics, and a new lane-side gate reads that marker to decide whether the job blocks. It opens only when every failed test in the run carries one, and blocks on anything it cannot attribute: a report it could not read, a failed test the report does not name, a failure with no marker. Only the lanes that block a merge ask it. The two pull-request lanes do; nightly and the tag lane run the same suites and stay hard. A soft pass on the tag lane would ship a release candidate whose tenant clusters brought up no worker nodes, and nightly is the only place an e2e failure reaches anyone who did not open a pull request, since no lane in this tree notifies anywhere. A softened red finishes the job green, so the job's result stops answering whether the suite passed, and both kinds of consumer are handed the decision instead of inferring it. A step gated on the job failing would silently stop running: the SSH breakpoint is that case, and it still opens for a pull request marked for debugging. A job that reports the result would silently start reporting a pass: the required commit status now says the deadline was missed and that the merge is not blocked on it, rather than that the suite passed. The deadline goes to 29m. That is the test's own budget and nothing more: its clock starts after the machinedeployment wait, the LoadBalancer assignment and the healthz probe, while the chart's autoscaler and health-check timers have been running since the Machine was created, so no ordering against them is claimed. The chart's 20m and 30m give a scale to pick against, and a budget over the first leaves room for a remediation cycle without promising that one fits. The operation ceiling both suites give the script rises with the deadline, and the diagnostics phase budget rises inside it, because the collectors gated ahead of the serial console cost more between them than the old budget covered. Both budget guards read their terms from the sources that set them. The e2e job cap rises with them, and its guard holds what a guard can hold: the pair's operation and catch ceilings, and one figure across every lane that runs the suite. It also records that the suites' full summed ceilings have never fit any cap in that file, so the margin above the bound is taken from a measured run instead. The prose that restates the cap elsewhere moves with it, and the sweep that holds the deadline's restatements now reads the figure spelled in words as well as in minutes. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
IvanHunters
|
f035da8a0f |
Merge branch 'main' into feat/kubernetes-nodes-phase2b
Resolve conflicts introduced by main: - Drop the FluxCD tenant addon (API type, deepcopy, values schema, RD) in favour of the GPU type main added; main removed the fluxcd addon wholesale. - Renumber the Phase 2 worker-pool adoption migration 54 -> 56, behind main's redis-operator (54) and fluxcd-removal (55) migrations; targetVersion 55 -> 57. - Accept the branch's removal of the in-chart talos-reconcile Job; it moved to the kubernetes-nodes chart, main only touched its comment. - Regenerate kubernetes values.schema.json / README and the kubernetes-rd ApplicationDefinition schema. Signed-off-by: IvanHunters <xorokhotnikov@gmail.com> |
||
|
Andrei Kvapil
|
7cca3d75e7 |
feat(kubernetes)!: remove the FluxCD tenant addon
The optional `addons.fluxcd` toggle of the Kubernetes app deployed a Flux Operator + FluxInstance into a tenant cluster. It was served by two vendored charts, packages/system/fluxcd-operator and packages/system/ fluxcd, which were the last AGPL-3.0 content in this Apache-2.0 repository and a governance blocker for CNCF Incubation. The management cluster does not use these charts (it runs classic Flux via internal/fluxinstall; migration 21 already removed the operator from it), and the addon defaulted to disabled. Rather than maintain an AGPL payload for an unused feature, drop it entirely: - delete both system packages and their platform source entries - remove the addons.fluxcd HelmReleases, values and generated schema from the Kubernetes app - drop the addons.fluxcd block still submitted by the e2e harness - drop the removed fluxcd package entries from CODEOWNERS BREAKING CHANGE: the addons.fluxcd option of the Kubernetes app is removed. Tenant clusters that had it enabled keep their running Flux release; see the companion migration for the upgrade path. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Andrei Kvapil <andrei.kvapil@aenix.io> |
||
|
Aleksei Sviridkin
|
7ec5dff93a |
test(e2e): add a runner-layer fixed-work canary around the node-join wait
The counters the node-join failure path collects are read in whatever unit their source publishes: time in the CPU rows, events in the KVM exits, instantaneous values and running maxima in the files beside those. Not one of them carries a unit of work of its own, so a machine that spent every tick and got a fraction of the work done reads healthy in all of them. Add a collector that carries its own unit of work. It runs two fixed amounts of work at the runner layer and times each: a loop over a handful of integer-valued scalars, which pressure on the shared cache or the memory controller moves little, and a store stream through blocks larger than the last-level cache one core can allocate into, which is dominated by it. Fixed work gives each reading a scale of its own, so a red run can be read with no green one beside it, and the capture states the expected range and where the estimate comes from. Sampled where the three counter pairs already are, before the wait on every run and after it on both paths, but outside their brackets: it occupies a core rather than reading a file, and a burn inside one of those intervals would land in the interval each pair divides by: runner-kernel CPU time charged to the join window, and for the two counting the guests, a core taken away from what they count. Two arms, each under a twenty-second ceiling with the five-second kill grace behind it, so fifty seconds at worst against the 420s diagnostics budget. The budget guard derives that from the ceiling the source declares, and the phase-warning guard names the canary among the collectors that keep running when timeout is missing. An exit shaped like the ceiling is attributed to it only when the whole bound was on the clock: an outside kill produces the same status, and read by status alone it would publish exactly the finding this collector exists to detect. A reading below the range the legend states, and an arm the ceiling stopped before it could produce one, reach the job log from all three reporting sites, each naming where it was taken, the way a missing reading does from the two outside the diagnostics block. On the passing path the report that holds the figure is the artifact nobody downloads. Each floor is printed in the legend it comes from and sits where a bounded arm crosses it before the ceiling stops the arm, because a floor the ceiling reaches first never fires: the arm is cut off before its rate can fall that far, and every alert then arrives as a bound rather than as the slowdown the floor was meant to name. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
IvanHunters
|
cd72d81bb4 |
Merge branch 'main' into feat/kubernetes-nodes-phase2b
Reconcile the phase-2b split with main's configurable worker-health feature (MHC condition timeouts, maxNodeProvisionTime, podCpuLimit/ podCpuRequest), which landed on both charts after this branch forked: - kubernetes (parent): pool rendering stays removed; main's parent worker-pool additions do not apply post-split. - kubernetes-nodes (pool): absorb the feature; keep the directional parent-version skew guard and the immutable cluster/storageClass rules. KubevirtMachineTemplate content hashes are unchanged (render snapshot). - migration 54: carry podCpuLimit/podCpuRequest (KMT-hash inputs) and the cluster-wide readyUnknownTimeout/readyFalseTimeout/maxNodeProvisionTime into adopted pools, so adoption neither rolls nor resets a tuned pool. Covered by a new adopt-values bats case. - drop the parent chart's obsolete MHC-timeout suites; the kubernetes-nodes twins cover the feature. Signed-off-by: IvanHunters <xorokhotnikov@gmail.com> |
||
|
Aleksei Sviridkin
|
b1b8cbecc2 |
test(e2e): price the layers above the sandbox on the node-join window
Every CPU reading on this failure path has a subject inside the sandbox: a cgroup, a tenant worker's threads, a sandbox node's own /proc/stat. Each of them can say a guest was given its ticks while getting no work done, and none can say what the layers above spent. That is where this suite's own tax is paid: the sandbox nodes' guest time is charged to the runner kernel as user time, and whatever the runner VM loses to the machine hosting it is charged to nobody the sandbox can see. Two readings close that. /proc/stat is not namespaced, so reading it from the sandbox container reads the runner VM's kernel -- the only place steal on that layer appears, which is what separates a slow suite from a machine sharing a core with somebody else. The sandbox VMs are QEMU processes in this container's own PID namespace, so their threads are a local read too, and they say which node spent the kernel's user time and whether the three spent it evenly. Both are taken at the same two points as the KVM counters and for the same reason: the values are cumulative, so a rate needs a pair, and the interval worth measuring is the window the guest is failing to boot in. The thread reading runs the probe the worker capture already uses rather than a copy of it, since a second copy would drift exactly at the failure paths. Each carries the layer it describes in its own legend, because three kernels in this report publish a /proc/stat and two captures read QEMU threads, one layer apart in each direction -- and a reader who takes one for the other compares a hypervisor against its own guest. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
0a1d2673d0 |
test(e2e): capture worker block IO and volume state on node-join failure
A worker that burns its whole vCPU while making no progress is either computing without result or waiting on blocks, and every counter this path already collects is blind to the second: the CPU families say what the group was allowed and what it got, the byte counters say what crossed the network, and neither counts a read. Read the Pod's block IO counters from the same kubelet endpoint the CPU ones come from, once rather than twice, because they accumulate from the container starting and a worker container starts with the guest it carries. Read the state of the replicated volumes behind the worker disks as well. A replica that is SyncTarget, Inconsistent or Diskless serves the guest's system disk from a peer over the network rather than locally, and the report collects the same tables only after the suite, by which time cleanup has deleted the volumes a red run would need them for. The guest's own /proc/diskstats would answer this from inside and cannot be reached: the credential minted for the guest carries the os:reader role, which is not allowed to read file contents, and Talos publishes no resource carrying disk statistics. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
827f4e7e10 |
test(e2e): measure the sandbox kernel KVM exit counters across the node-join wait
Every CPU reading on this failure path describes a subject inside the sandbox -- a cgroup, a QEMU thread, a node's own /proc/stat -- and all of them can report a guest that was handed its ticks while the guest gets no work done. None of them can say what a tick cost. That price is paid one layer down: a tenant worker is a guest of a sandbox node, which is itself a guest of the runner, and the runner's kernel absorbs the exits, nested page faults and TLB flushes the arrangement provokes. Read those counters from the debugfs that kernel publishes them in. The suite runs in a container sharing it, so this is a local read rather than a walk over nodes, which is what makes it affordable at all: the phase budget has no room left for a further per-node read. debugfs is mounted in a mount namespace belonging to the call, so nothing is left behind for the twenty minutes of suite that follow it. The pair is taken across the node-join wait rather than inside the diagnostics block: once before the wait on every run, once after it on the failing path and on the passing one. The counters are running totals -- per-VM since that VM was created, kernel-wide summed over the VMs alive at the read -- so the interval a difference divides by becomes the window the guest was failing to boot in, and a passing run leaves a baseline to read the failing ones against. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
IvanHunters
|
5297df5cf3 |
Merge remote-tracking branch 'origin/main' into pr-3315-lgtm
Signed-off-by: IvanHunters <xorokhotnikov@gmail.com> # Conflicts: # api/apps/v1alpha1/kubernetes/types.go # api/apps/v1alpha1/kubernetes/zz_generated.deepcopy.go # hack/e2e-chainsaw/_lib/run-kubernetes.sh # packages/apps/kubernetes-nodes/tests/render-parity.sh # packages/apps/kubernetes/README.md # packages/apps/kubernetes/templates/cluster.yaml # packages/apps/kubernetes/values.schema.json # packages/apps/kubernetes/values.yaml # packages/system/kubernetes-nodes-rd/cozyrds/kubernetes-nodes.yaml # packages/system/kubernetes-rd/cozyrds/kubernetes.yaml |
||
|
Aleksei Sviridkin
|
73b8afc256 |
test(e2e): run tenant workers on one vCPU with CPU headroom above it
Under a two-vCPU worker both vCPU threads burn 1.4 to 1.8 cores between them while the guest kernel never gets past SMP bringup, and raising the CFS ceiling from two cores to three changed neither the burn nor the progress, so the spin is between the two vCPUs, not against the quota. A single-vCPU guest has no sibling to spin on. The ceiling of two keeps the QEMU emulator and IO threads from eating into the one core the guest computes with, and the request keeps scheduling where the workers always were. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
b8964c8233 |
test(e2e): give tenant workers CPU headroom with a request to match
Restores the ceiling the kubernetes suites briefly carried, now that a node group can declare the CPU request that goes with it. The limit on its own was defaulted into the request by KubeVirt, so every worker asked the scheduler for the whole ceiling and sat Pending on Insufficient cpu before a guest existed. The request is set to what KubeVirt derived for these workers before, which leaves scheduling where it was and leaves the ceiling as the only variable the run changes. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
80e2753bf4 |
fix(e2e): drop the worker CPU limit until it can carry its own request
A container that sets a CPU limit and no request is given a request equal to the limit, so each tenant worker asked the scheduler for its whole three-CPU ceiling and every kubernetes suite worker sat Pending on Insufficient cpu across all three sandbox nodes, before any guest booted. The headroom experiment comes back when the chart can pair the limit with an explicit request. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
c2c50b4119 |
test(e2e): give tenant workers CPU headroom above their vCPU count
Workers that fail to join have been measured burning their vCPU threads flat out under a ceiling equal to their vCPU count while the guest kernel recorded no progress at all, with the emulator, IO, vhost and vnc threads flat at zero. That is the shape of a vCPU spinning on a lock whose holder is the other vCPU, preempted out of the quota the two share, and it predicts that a ceiling with room above the vCPU count lets the preempted one back in. What the next runs answer is whether the join failure rate moves. If it does not, the lever is elsewhere and the value goes back. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
cf2d9c6d13 |
test(e2e): split the tenant worker CPU time across QEMU threads
The node-join diagnostics phase samples the tenant worker's CPU counters twice on failure, but it reads them from the compute container's cgroup, and that cgroup holds every thread QEMU runs: the guest's vCPUs, QEMU's main thread, and its IO and worker threads. A container burning most of its quota while the guest reports no progress is two incompatible findings wearing one number, and no container-level counter separates them. Add a collector that execs into each Running worker's compute container, finds the QEMU process and copies every /proc/<pid>/task/<tid>/stat line into the artifact verbatim, with a legend beside them. Nothing is computed here. The legend names where utime and stime sit and why counting fields over whitespace reads the wrong column for exactly the threads the split is about: a KVM vCPU thread is named "CPU 0/KVM", that space is inside the comm parentheses, and every field after it shifts by one. It runs inside the sampling loop the two existing subjects already share, so it costs no second wait: both readings ride the interval already being paid, and the rate comes from the two files plus the stamps each one carries. Every way the read can come up short is a named outcome rather than an empty directory, and the legend and the pairing instruction are withheld from each of them: telling a reader to subtract two files asserts that both hold a whole reading. The probe shells out to nothing but getconf, whose absence it names, because a probe that needed cat would report a container missing coreutils as a container whose QEMU had exited, and it counts what it managed to read so that having read no process at all stays separable from having read them all without finding QEMU. Its stderr is suppressed on the walk rather than on either read: redirections are applied left to right, so a 2>/dev/null after an input redirect is not in place yet when opening the input fails, and a thread that exits between the glob and the read would otherwise put a shell complaint in the artifact beside a healthy capture. It also reports a QEMU that was named and then gone before its task files could be opened, because a process heading alone is enough to make the capture non-empty and would otherwise carry a legend describing columns the file does not have. The Pod listing the cAdvisor captures already share now takes its whole query as a parameter, so this collector reuses that read rather than carrying a third copy of it, and asks for the Running Pods only: a launcher left over from an earlier attempt costs the sibling caller nothing and would cost this one a bounded exec and a slot under its cap. The guard that holds the sampling group against the room the tenant snapshot needs now counts every capped walk in the loop rather than the two spelled the same way, so the new subject is inside the arithmetic rather than beside it. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
959e90c9ab |
test(e2e): time the guest console capture and baseline it on a green run
A tenant worker's guest console is stamped in seconds since that guest's own kernel started, so a log ending mid-boot says how far the guest got and nothing about when. Whether it ends there because the guest started late or because the guest stopped printing decides which failure is being looked at, and neither instant was recorded: the read's own instant was nowhere in the file, surviving only as the artifact's mtime, which is metadata rather than content and so says nothing to anyone reading the log; and the guest's start only inside a describe dump that a capture with no silent console never triggers. Across three collected runs every capture exited zero and every log began at the first kernel line, so the reads were complete at both ends and the logs end where the guests stopped -- and which of the two happened could not be read off any of them. Each console log now carries the instant its own read started and returned, and CAPTURE-TIMELINE.txt beside it carries each Pod's phase, its start time and the start of its compute and console containers, from one bounded read taken once for the whole selector. Those subtract to how long a guest had been alive when it was read, which is what the log's last kernel stamp is then compared against. The console container is asked for in both status lists, since KubeVirt runs it as an init container and reading only one of them would report a container that started fine as one that never ran. A read that failed, a read cut off part way and a read that answered with nothing are recorded as themselves rather than left as a missing file, and the note names the error log only when there is one. The same capture also runs on the passing path, before the tenant is torn down and the Pod holding the stream goes with it. Every console this tree has collected came from a failure, so what a healthy boot looks like has never been measured, while "slower than usual" is the claim the failure keeps resting on. That caller asks for the pool's minimum rather than its maximum, since two workers booting is the same measurement as ten, and it warns in the job log rather than failing a suite that has already proved everything it exists to prove. Both paths write the same directory, so the capture also records which of them produced it. That capture costs real wall clock at the tail of an operation with a fixed ceiling, which is a way for a run that passed to be killed while collecting a debugging aid. So it is declined out loud when the ceiling no longer has room for it, against a reserve derived from its own cost and held in step with the timeout both suites give the operation. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
2b3c163d15 |
fix(e2e): size tenant workers by resources, not by u1.large
Sizing the workers with u1.large doubled their memory along with the vCPUs, and four 8Gi workers do not fit on the sandbox beside the rest of the suite: the second worker of each cluster stayed Pending with Insufficient memory on all three nodes, so the node group could never reach two Ready nodes and node-join failed for a reason that has nothing to do with the CPU ceiling it was meant to test. Setting resources.cpu and resources.memory keeps the two vCPUs at the memory the workers had before. The chart drops the instancetype from the VM when a group sets both, and the values schema still requires the field, so it stays in the values as the sizing fallback it now is. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
3b49be07fa |
test(e2e): give tenant workers two vCPUs
The tenant worker boots Talos and unpacks the kubelet image on a single vCPU whose CFS quota also covers the QEMU emulator and IO threads, and measurements on both quiet and loaded runs show the compute container pinned at 93-98% of that one-core ceiling for the whole node-Ready budget. The join deadline is then decided entirely by how much CPU work the boot happens to need, and under parallel e2e load it does not fit. Doubling the worker to u1.large raises the ceiling instead of the weight, which measurements showed changes nothing since the workers never lose a scheduling fight. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
527d4ea00f |
test(e2e): measure tenant worker CPU as a rate, and the sandbox nodes under it
Two instruments for the node-join failure the kubernetes-* suites keep hitting. Neither changes what the suites assert. The worker CPU capture now reads container_cpu_usage_seconds_total beside the CFS counters, and reads them twice a fixed interval apart instead of once. The throttled counters say a container met its ceiling and cannot say how much CPU it got, which are the two candidate explanations for a stalled worker: a guest held at a one-core ceiling, and a guest whose vCPU thread is rarely scheduled at all. Both leave throttled periods behind. Every counter involved accumulates from container start, so a single reading is an average over the whole uptime and no rate at all, and the profile under investigation is a guest that freezes for tens of seconds and then runs flat out: the average comes out low, the instantaneous figure comes out at the cap, and only the second separates it from a guest that is simply capped. The sandbox nodes are a new subject. The worker captures describe a cgroup, so a guest that is runnable and never scheduled looks from there like a guest that asked for nothing. Whether the sandbox node itself was given its turn is recorded only in the steal column of its own /proc/stat: no kubelet endpoint carries it, and crust-gather cannot reach it because PodSecurity refuses its node-shell Pod here, so the read goes over the Talos API, which is already how the report reaches those nodes for dmesg. The rows are captured verbatim with their column legend beside them rather than reduced to a percentage, because /proc/stat has no header, steal and guest sit next to each other, and guest is large on a node running VMs and already counted inside user. Both subjects are read on either side of one shared wait, so the pair pays it once. The wait is not the interval either subject's counters span, though: each pass walks its own nodes and the other subject's pass falls between a subject's two readings, and on the run this exists for those passes are the slow part. Each capture therefore stamps both ends of its read, so a rate comes from subtracting stamps rather than from assuming the knob, and the width of those brackets is the uncertainty. That wait is a measurement interval rather than a stand-in for a condition nobody wrote a wait for, so the e2e conventions carve it out by name instead of marking it as a fixed timeout to be replaced. The pair is gated below the serial console rather than above it. Its ceiling is sixteen bounded reads plus the interval, which is most of the phase budget, and admission gates only when a collector may start, so above the console a run that hit those bounds would leave the console never started rather than merely cut short. The console is the only capture that survives a worker which never reached apid, which is the shape this failure usually takes. That overrun is inside the term the budget derivation already allows for one late admission, and a guard holds the pair's ceiling against it. No collector here may leave an empty directory, since an empty directory reads as a healthy node: a missing binary, a missing talosconfig, a listing that never answered, a node with no address and a read that was cut off each land as a named note or a COLLECTION-FAILED.txt. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> |
||
|
Aleksei Sviridkin
|
61edf33ef4 |
fix(e2e): capture both ends of the worker byte path on node-join failure
A tenant worker that misses the node-Ready deadline while its kubelet image pull crawls leaves two mechanisms behind one symptom, and the artifact could not tell them apart. Either the link is slow, or the pull keeps restarting and discarding what it fetched; from inside the guest both are a progress line that barely moves. Per-worker network counters, read from the kubelet's cAdvisor endpoint the CPU throttling capture already uses, count every byte that crossed into the Pod including the ones a restart threw away, so the gap between that and the progress the guest reported separates the two. Which row carries it is not obvious and getting it wrong inverts the reading, so the collector states it. cAdvisor publishes one row per interface in the Pod netns, and a bridge-bound worker has four: the one still named eth0 is a dummy holding the Pod IP, carries no traffic, and reports zero -- the row a reader reaches for first, where zero reads as a worker that received nothing. The renamed NIC's receive is the number that answers; the bridge counts the same bytes again and the tap is named from the host's end. No interface predicate is applied, so all four reach the artifact and can be told apart. Two further properties of the family fail by capturing nothing rather than loudly: the dropped-packet series are spelled packets_dropped, and every row carries an empty container label because cAdvisor publishes them on the Pod sandbox, so a filter borrowed from the CPU counters beside it matches nothing at all. Both are stated where the filter is. The fixture's rows are real kubelet output relabelled; its interface names are read from the KubeVirt version this repo ships rather than sampled, since no virt-launcher Pod was available. The guest side gains the service state machine and the link table. Talos creates a service's log buffer when its runner opens the log writer, as it starts the process, so a kubelet still waiting on its volumes, on time sync, or on the container runtime reporting healthy never reached its runner, has no log at all, and the capture recorded an rpc error naming a log that was never registered: true about the log, silent about the service, while the state and the awaited condition were the finding. The link table is read as yaml because MTU lives in the spec rather than among the printed columns, and it is the value that decides whether a path-MTU reading of a stalled transfer survives. Both carry a tighter bound than the two reads beside them, being single tabular RPCs rather than streams, and the walk they sit in is the term the phase budget sizes itself against. A worker whose reads all failed now says so once for the worker. Each log already carried its own error and exit code, which is accurate per file and adds up to a directory indistinguishable from a worker that had little to report. The phase budget drops from 8m to 7m to pay for the two added guest reads, because the inequality keeping the tenant snapshot reachable had ten seconds of slack. A shorter budget declines more, and what it declines first is what is gated last: the mirror capture, whose subject is partly recoverable from its own request counts and the host snapshot, and the image-cache re-probe, which has no substitute and was already the first to go. The second cost lands on the guest captures, which are gated after the new counters and reach their gate with both the shortened budget and a listing plus a read per node already spent -- up to 100s at the read bound and the three-node cap, a ceiling those numbers impose rather than a duration measured on a live cluster. Sharing that read would mean the two captures becoming one walk, which is left out for scope rather than merit: the phase gate is a deadline rather than a per-collector allowance, so one walk admitted at the first of the two gates would cost less and collect more. The cost is stated rather than netted away, and the source says which way the argument runs. Both captures run one shared body, which also settles what the two copies had already broken once: the audit that catches an unbounded read pins each collector by something only it produces, and two collectors issuing byte-identical reads left neither with anything of its own, so stubbing one out of the audit went unnoticed. It is pinned by the report directory each writes. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
829edfadb3 |
test(e2e): warm the ghcr.io mirror before the tenant suites need it
The in-sandbox pull-through cache fetches a blob only when a client asks for it, so the first tenant worker's kubelet pull is a live fetch from ghcr.io. That leg is what misses the node-Ready budget: measured at 40-165 KB/s for a 61 MiB image, with two blob streams dying mid-transfer after 7-8 MiB on an upstream HTTP/2 reset. Talos holds the kubelet service in PreFunc until the pull completes, so the node never registers, and the tenant cilium HelmRelease then fails for want of a schedulable node rather than for any fault of its own. Fill the cache from a Job started in the same install step that deploys the mirror, tens of minutes before the kubernetes suites run. The Job waits for the mirror to serve, then asks it for each kubelet tag's index, walks the manifests it gets back, and discards the blob bodies; distribution populates its blobstore on the way past. It warms one platform per tag -- the one the workers run -- picked out of the index by record, which keeps the walk a grep. Walking both platforms was a true addition to the same throttled leg: arm64 is bytes nothing in the sandbox pulls, and the digest ordering put it first for both tags, so on the slow runs this exists for the deadline expired with the second tag still cold. It warms the two minors the suites select, not the whole version map. The map would be ~300 MiB at the one-platform walk this ships (~61 MiB per tag), three fifths of it for minors no suite can resolve, spent on the same throttled egress the rest of the install is using. The tag list still comes from the chart's map rather than a copy, ordered the way the suites resolve it, so the set is theirs without restating their expression. The readiness wait is a wall clock, not a per-attempt count. Cilium answers a Service with no ready endpoints by rejecting, so a probe costs nothing when the mirror is absent and the full timeout when it is merely slow; a constant was wrong in both directions across two attempts at picking one. Tags are refused by character class before the pattern is applied, so a value cannot pass by having a valid first line: grep matches per line and the pattern's anchor is end-of-line, not end-of-input. The Job exits non-zero on any miss, so backoffLimit is a real bound and a warm-up that reached nothing cannot report success. It names its own requests, because the node-join diagnostic counts kubelet GETs in the access log to tell a worker that pulled through the mirror from one that fell back to public ghcr.io. Naming them is only half of that. The same diagnostic now reports the Job's status and result and the Job outlives the suites, because excluding the warm-up without reporting it would leave a red run unable to distinguish a warm cache from a warm-up that never started. That counter also moved onto the access-line shape: registry:2.8.3 writes three lines carrying the path per request, so matching the path counted one pull as three. It reads the repository from the same variable the warm-up does, so a rename cannot leave the diagnostic counting zero while the workers pull under the new name. GET and HEAD both count: containerd resolves a manifest with a HEAD first, so a GET-anchored count reports a pull that died before its first GET as no request at all. The failure path also prints how long the mirror took to answer the worker. A warm cache and a cold one are otherwise indistinguishable in the artifacts -- the Job reports success either way, because the proxy stores the blob in the background and a completed GET is a different event from a stored blob -- and server-side duration separates them by orders of magnitude. It filters the log in the shell that read it: a single argv element is capped at 131072 bytes on Linux regardless of ARG_MAX, and probe traffic carries this log past that within about twenty minutes of mirror uptime, so handing it to a helper would empty the section on every run worth reading. Nothing is printed when the log could not be read at all, because a heading with no lines under it reads as a finding about the workers. Measured against a local pull-through registry using the script the builder emits: 32m47s cold, 1.1s warm for the same tags, and the blobs it fetches are the same digests the failing runs stalled on. A warm-up that fails, times out or never starts does not fail the step, and the suites behave as they did before it existed. It is not free, though: it pulls across the same throttled egress the install is using, which is why the warm set is two tags and not five. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <3811295@gmail.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
b8a5b9ba1d |
test(e2e): capture worker CPU throttling counters on node-join failure
When a tenant worker misses the node-Ready deadline, the sandbox node it runs on is routinely at half its capacity with a clean kernel log. In that state "the host starved the VM" and "the VM was held at its own CFS ceiling" produce identical evidence: node-level utilisation reports what the node used, never what the container was allowed. The CFS counters are the only place the difference is recorded, and nothing in the tree collected them for the tenant workers, so every diagnosis of this failure so far had to be inferred from outside. Read them from the kubelet's cAdvisor endpoint on the node hosting each worker. Five series carry the whole answer and cAdvisor publishes all of them per container, so nothing is computed here: the scheduled and throttled period counts, the time spent throttled, and the quota and period that form the ceiling. The throttled counters alone say a container hit some ceiling, not which one, and a VM capped at one core and a VM capped at eight are the same number without the quota. Absence of the quota series is itself a reading rather than a gap. cAdvisor emits it only when the quota is non-zero, so a container reporting the period and no quota is one running uncapped - which is one of the two answers this collector exists to distinguish, arriving without being derived. Reading the kubelet rather than the container matters beyond convenience. A reader running inside the container shares the cgroup it measures, so it slows down in proportion to how much the answer matters and would need a budget sized for the pathological case; the kubelet is outside that cgroup. It also drops any dependency on what the container image ships and on how the host lays cgroups out. The namespace is not a worker filter and is not used as one: it also holds the tenant control plane, whose apiserver and etcd carry CPU limits of their own and can be genuinely throttled. The Pod name carries the second half of the filter, so a throttled apiserver cannot answer under a heading that says worker. The walk is over nodes rather than Pods, because the endpoint is per node and one read covers every worker it hosts. It is gated on the diagnostics phase budget, and placed first among the five gated collectors: the budget declines whatever has not started when it runs out, its own ceiling is four bounded reads rather than the minutes the four after it can spend, and it is the only one whose question has no other answer in the tree. That order is a trade, stated here because it is a decision rather than a detail - it puts the guest serial console one place further from the budget, and that capture is the only one describing a worker which never reached apid. Both reads take the block's own read-budget knob rather than a literal equal to it, so lowering that knob lowers these too. Each keeps its own exit status rather than routing through the shared helper, which always returns zero by design: this collector has to tell "the kubelet was not read" from "the kubelet answered and carried no worker", and only the status separates them. The stream is captured whole and filtered after, since a pipeline reports its last command's status and would otherwise hand that decision to grep. Every way a read can come up short is named rather than left as zero bytes: a kubelet that was not read, one that answered with nothing for the namespace, a stream cut short, and exit 124 - which cannot distinguish this collector timing out from the read exiting 124 on its own - each get their own line, and each says whether an error log exists to explain it. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
IvanHunters
|
b4a2f6f791 |
Merge remote-tracking branch 'origin/main' into pr-3315-lgtm
# Conflicts: # api/apps/v1alpha1/kubernetes/types.go # api/apps/v1alpha1/kubernetes/zz_generated.deepcopy.go # hack/e2e-chainsaw/_lib/run-kubernetes.sh # packages/apps/kubernetes-nodes/tests/render-parity.sh # packages/apps/kubernetes/README.md # packages/apps/kubernetes/templates/cluster.yaml # packages/apps/kubernetes/templates/talos/talos-reconcile-job.yaml # packages/apps/kubernetes/tests/talos_templates_test.yaml # packages/apps/kubernetes/values.schema.json # packages/apps/kubernetes/values.yaml # packages/system/kubernetes-nodes-rd/cozyrds/kubernetes-nodes.yaml # packages/system/kubernetes-rd/cozyrds/kubernetes.yaml |
||
|
Aleksei Sviridkin
|
9d4bcd8d66 |
fix(e2e): wire the ghcr.io mirror into the tenant install and node-join flow
Deploy the pull-through mirror from its own install step rather than folding it into the Talos image-cache step, whose early return on a missing values key would otherwise skip it silently. Build the tenant CR spec.talos block from both the OS image cache and the registry mirror, each included only when its in-sandbox backend is up. On a node-join failure dump the mirror's state alongside the Talos image cache, under the same phase gate as its neighbours so it cannot starve the guest captures of budget. Record the mirror's legs in the step's budget inventory. talos_spec_block is covered by a bats suite over all four mirror-up and mirror-down combinations, splicing the result under spec: and reading it back with yq so a block indented into the wrong parent fails, and the install step is pinned to exclude the Cilium policy from its pre-Cilium apply. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
1c6ee7c047 |
fix(e2e): harden the ghcr.io mirror resolver and bound its diagnostics
resolve_ghcr_mirror_endpoint decides once, for the whole shared sandbox, whether tenant workers are pointed at the in-sandbox mirror, so a misclassified transient failure would disable it for the rest of the run. Cache only stable answers: a genuine API NotFound and a definite Available=False; never a transient query error, a missing kubectl, or a failed egress-allow apply. Read readiness from the Deployment's own Available condition when the rollout watch exits non-zero, and commit the endpoint only once the egress allow is in place, surfacing the apply failure reason rather than a generic blip. The node-join diagnostic counts kubelet-image requests across the whole access log rather than a fixed tail, and every read it makes is wall-clock bounded by the caller's COZY_DIAG_READ_TIMEOUT so no single read can hold the failure path open. The bats suite exercises every resolver branch and the diagnostic's filtering, with a kubectl stub that reports errors on stderr as the real binary does and a timeout mock so the bounded reads stay reachable under test. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
IvanHunters
|
58910c0edc |
feat(kubernetes): route tenant worker image pulls through a registry mirror
Add a talos.registryMirrors passthrough to the kubernetes and kubernetes-nodes
charts that renders into the worker TalosConfigTemplate machine.registries.mirrors,
so an operator can point a registry such as ghcr.io at an in-cluster pull-through
mirror and a worker boot no longer depends on live public egress. Default {}
keeps the rendered worker reconcile Job byte-identical to before, content-hash
Job name included. Free-form mirror values are shell-escaped for the unquoted
apply heredoc, covered by an execution-level bats test.
Also add the e2e ghcr-mirror (registry:2 pull-through plus a Cilium egress
policy) and wire it into the kubernetes chainsaw suites, so CI exercises the
mirror path and takes the flaky public ghcr.io kubelet-image pull off live
egress.
Refs #3513 (addresses the egress variant of the tenant-worker node-join
failures; the dominant variant is CI-runner CPU starvation, tracked in the
issue).
Signed-off-by: IvanHunters <xorokhotnikov@gmail.com>
|
||
|
Aleksei Sviridkin
|
5400f165e3 |
fix(e2e): keep a failed mirror lookup from pinning the public factory
resolve_talos_image_factory_url asked whether the in-cluster Talos image mirror exists with a plain `kubectl get deploy`, where a connection refused, an Unauthorized and a real absence all arrive as the same non-zero exit. It then cached that reading for the whole suite, so one unlucky moment sent every tenant worker to the public Image Factory over live egress -- the dependency the mirror exists to remove -- while the run still looked ordinary. Ask with --ignore-not-found -o name instead: a real absence answers exit 0 with no output, which leaves a non-zero exit meaning only that the question failed, and re-ask a couple of times before giving up. Only the empty answer turns the suite toward the public factory. An unanswered lookup warns, falls back for its own caller alone, and is not written to the decision file, so the next tenant test asks again rather than inheriting a guess. Each attempt is bounded twice, and the two bounds are deliberately not equal. kubectl defaults --request-timeout to 0, so a connection that establishes and then stalls has no deadline of its own; the flag alone is not a wall-clock bound either, because kubectl retries discovery several times before giving up. Measured against a stub apiserver, the three ways this call fails behave differently: a refused port answers in under a second and names the refusal, one that stalls mid-TLS gets kubectl's own handshake deadline and still produces text, and a blackholed address produces nothing until some deadline expires. With both bounds at the same value the expiry that arrives first is always the wall-clock kill, so --request-timeout can never fire and that last case stays silent. The wrapper therefore sits ten seconds above it: the request deadline ends the attempt and kubectl says why, and the wrapper remains the backstop for a client that overruns its own deadline. Where timeout is not installed the read runs unwrapped rather than through it: its exit 127 would otherwise count as an attempt that did not answer, three times over, and a missing local binary would pin the whole suite to the public factory while every message blamed the apiserver. kubectl's stderr is left alone, because it is the only thing separating refused from unauthorized from stalled. It is still not guaranteed, since a kill can land before the client has said anything, so the lookup also names the last attempt's exit status before giving up, in the vocabulary the diagnostic reads use for 124 and 137. The diagnostic path asks the same question and keeps its own answer to it. That one gives up after a single attempt, under a phase budget where re-asking spends what the snapshot behind it needs; this one runs on the happy path, where a re-ask costs a suite nothing and buys it the mirror. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
04494555c9 |
test(e2e): pin that an unanswered mirror lookup is not a mirror absence
The Talos image factory decision asks the apiserver once whether the in-cluster mirror Deployment exists, reads any non-zero exit as an absence, and caches that answer for the whole suite. A refused, throttled or timed-out lookup therefore sends every tenant worker in the run to the public Image Factory over live egress -- the dependency the mirror exists to remove -- and the run reads as ordinary environment noise rather than as a decision that was never actually made. Route that question through one helper and pin what the helper owes its caller: the three outcomes it must tell apart, the flags that make the three-way answer possible at all, a deadline on every attempt, the caching rule -- a real absence is resolved once and cached, an unanswered lookup falls back for its own caller alone -- and that a missing local timeout binary leaves the lookup asking the apiserver rather than answering unknown three times over. Two of those need saying carefully, because both concern what the caller is told when the lookup fails. A killed attempt gets its own case, driven by a wrapper that exits without running kubectl at all, so the lookup has to account for why it gave up rather than relay someone else's words: pinning the reason on a kubectl that failed loudly, as the sibling case does, pins it only where it was never in doubt. And the wall-clock bound is pinned as strictly greater than the request deadline beside it, which is the property that lets kubectl reach its own deadline and say what happened; level with it, the kill always lands first. The helper is a placeholder in this commit and claims present unconditionally, so the cases that need the distinction fail. The stubs are files on PATH rather than shell functions because the lookup runs under an external timeout binary, which execs a real file and would never see a shell override. xtrace stays off for the rest of a test that captures stderr: turned back on, a failing assertion wedges bats instead of reporting it. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
0d7beaf630 |
fix(e2e): bound the tenant failure diagnostics per read and as a phase
The node-join diagnostics issued about a dozen kubectl reads with neither an outer wall-clock bound nor --request-timeout, starting with a describe nodes through the tenant kubeconfig. That read targets the component least likely to answer in a node-join failure, and an unbounded read there does not lose only itself: it holds the Chainsaw op until the op is killed, so everything after it is lost rather than truncated, the tenant crust-gather snapshot the failing exit triggers above all. Each read now carries the per-call bound the newer collectors on this path already use, sized lower because this block issues a dozen back to back where those bound one. A read that fails or is cut off says so instead of passing in silence, and says only what was observed: a read that did not finish is silence about the cluster, not a finding about it. The importer walk gains a cap for the same reason its listing is now read on its own, since a listing that never answered must not read as a namespace with no importer in it. Per-read bounds alone do not end the phase, and the collectors here can together spend more than what is left of the op by the time they run. Much of that is not ceiling-bounded at all: on the guest-Talos path the applies, waits and deletes carry --request-timeout or --timeout, which bound one HTTP request rather than a client retrying against a wedged apiserver. So the phase is bounded as a phase. Once its budget is spent the collectors that have not started are declined out loud. That is not a worst case and is not written as one, but it does stop the snapshot being queued behind work begun past the deadline, which is how the snapshot was being lost. Raising the op ceiling was the alternative and it is not one: a larger ceiling times a sum that grows with every collector added is the same problem later. Because what runs last is what gets declined, the block is reordered so the cheap reads come first. The CSR list and the signer log are the discriminator for the failure mode a follow-up fix has to target; behind a serial-console walk and a guest-Talos capture that can spend the whole budget between them, they were exactly what a tight run would drop. The comments carry invariants rather than arithmetic. A comment that states a sum has to be kept true forever in every copy of it, and any edit to the code or the order falsifies one of them. The numbers live in the code as values and in the guard that pins them; the prose keeps why the budget comes from an inequality, why admission gates the start rather than the finish, why the cheap reads run first, and which residuals the guard's literals do not cover. The sibling scheduling-gate branch ends in the same exit and depends on the same snapshot, so its two reads are bounded too. So are the reads in the talos-image-cache diagnosis, whose gate also gains --ignore-not-found: exit 1 from kubectl get means a refused connection or Unauthorized as readily as NotFound, and reading it as absent announced the mirror as never deployed on the strength of a call that never reached the apiserver. Bounding those dumps needed a note of its own first, since timeout prints nothing when it fires and a cut-off dump would otherwise leave a bare section header that reads as a cache with no Pod, no Service and no EndpointSlice. The knobs those bounds come from are validated rather than only documented. Each is pasted somewhere that takes digits and nothing else, and each failed differently and quietly: a suffixed or leading-zero budget reached the arithmetic in the block's first statement and unwound the whole function before its own headline, so a failing run produced no diagnostics at all; a suffixed read bound became a millisecond request timeout and made every note blame the cluster for a value; zero disabled the bound outright; a suffixed importer cap made the comparison exit non-zero, so the cap stopped existing; and a non-numeric grace makes timeout exit before running the command at all. A rejected value falls back and says so, which is what the rest of this tree does with its own knobs. All four are re-checked where they are used, since a value set after sourcing never passed the assignment, and the read bound is checked again before the per-request string is composed from it so the two halves cannot disagree. Each default is declared once as a constant, and the re-checks read the bare variable so an unset knob is named rather than silently defaulted. When timeout itself is absent the reads run unbounded rather than being bounded with a binary that is not there: the latter turns every read into an exit 127 and every note into the cluster refusing, and collects nothing, where unbounded is what this path did before any of this. The phase says so once, and says what it costs rather than reassuring. The reachability re-probe in the image-cache helper stays unbounded and its comment names all seven of the calls it makes, the fire-and-forget delete after the probe Pod included: the same function decides the mirror on the happy path, so changing when it gives up is not a diagnostic's decision to make. Its bounds are spelled at each call site rather than precomputed, since a prefix built at source time cannot follow a knob a test lowers afterwards while the --request-timeout beside it does. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
c1071dab9b |
feat(kubernetes): let a node group log the guest serial console (#3637)
## What this PR does Both tenant Kubernetes e2e suites keep missing the node-Ready budget with zero Nodes registered, and the time goes inside the guest, between the worker VMI reaching Running and Talos asking for its apid certificate. Nothing collected today reaches that window. The in-guest capture the harness already has drives `talosctl`, so it needs the apid the worker never got to, and the management side has neither a Node nor a certificate request to read. Four attempts at this failure have been designed from outside the guest; the most recent raised the budget from 12m to 18m and the same fingerprint came back at the new deadline. KubeVirt can stream a VM's guest serial console into a `guest-console-log` container beside virt-launcher, readable with `kubectl logs` and needing nothing from inside the guest — no talosctl, no client certificate, no helper Pod, no reachable apid. That is the one property this failure class demands, so this turns it on for the suite's worker node group and collects it into the diagnostic bundle when node-join fails, ahead of the `talosctl` capture that cannot describe the same runs. Reads start at the beginning of the stream rather than tailing it: a guest that repaints its console would push the boot output out of any `--tail` window, while `--limit-bytes` truncates from the far end instead. Every read is bounded, and the walk stops at a cap that records both counts, so the collector cannot become an unbounded term in a failure path that already runs several others under one deadline. A read that returns no console is not filed as a quiet guest. Three outcomes get three labels, because only one of them says anything about the guest: the read failed and returned nothing, the read succeeded and returned nothing, or it broke off part way. Each read's stderr is kept beside the capture rather than inside it — kubectl's own `container guest-console-log is not valid for pod` would otherwise fill the file, and a console that never started would read as one that had spoken. That case is the headline one here, so the Pod state and events that explain it are collected alongside, once for the whole selector. The switch is a new opt-in `logSerialConsole` on a node group, off by default. It renders nothing unless set to true, including when set to false, because the `KubevirtMachineTemplate` is content-hash named and a byte emitted for a group that did not ask for it renames the template and rolls every live worker VM on upgrade. I verified that by rendering the base tree and this branch: absent and explicit-false produce a byte-identical template name on both charts. Worker VMs come from two charts, `kubernetes` and `kubernetes-nodes`; both carry the field, and the golden-parity check between them gains a case that keeps their output identical. The passing path checks separately that the container actually attached, and runs that check last, after everything the suite exists to prove — a debugging aid should not preempt the assertions that cover the cluster. Without it the suite would first learn the setting was inert in the one run that needed the console, when it can no longer be recovered. Relates to #3513. ### The risk this carries, stated up front The platform turns serial console logging off cluster-wide in the KubeVirt CR. That setting is a mitigation added in February 2026 for kubevirt/kubevirt#15989, where the `guest-console-log` container kept virt-launcher from starting and left VMs in `PodInitializing`. A per-VM `logSerialConsole` takes precedence over it — `isSerialConsoleLogEnabled` reads the VMI field first and only then falls back to the cluster setting — which is what makes this work, and also makes this the first thing on the platform to go through that mitigation. The mitigation was written against KubeVirt v1.6.x; the platform now ships v1.8.4, and the upstream issue is still open with no reproduction reported either way on the newer version. So the first run of these suites is also the experiment, and it doubles as a revalidation of that February mitigation on the version actually shipping, which nobody has done. The two outcomes are distinguishable rather than ambiguous: kubevirt/kubevirt#15989 shows up as virt-launcher wedged in `PodInitializing` with the VMI never reaching Running, while the failure this instruments shows the VMI Running and zero certificate requests. A red run with the first signature is an answer, not noise — and the collector is built so that it says which one happened instead of reporting silence. The blast radius is small: the product field is absent by default and only the e2e node group sets it, so backing this out is one line. The revert criterion, stated before the run rather than after it. If worker virt-launcher Pods sit in `Init` on `guest-console-log` and no Node registers, kubevirt/kubevirt#15989 still reproduces on v1.8.4: drop the `logSerialConsole: true` line from the e2e node group, keep the chart field, and the upstream issue gains the reproduction on a current version that its thread does not have. If the Pods reach Running and the console logs carry boot output, the mitigation is a candidate for removal platform-wide, which is its own change. Either way the run answers something that was open before it. Dropping `virtualMachineOptions.disableSerialConsoleLog` outright was the alternative. It is a smaller diff and needs no values key, but it exposes every VM on the platform to a race nobody has re-tested, instead of one node group in one test. That is its own revalidation exercise and belongs in its own change, on evidence this one produces. ### Verification `make unit-tests` passes. The new bats suite runs green under both `bash` and a `dash` `/bin/sh`, which matters because `hack/cozytest.sh` is `#!/bin/sh` and an earlier revision of this branch carried a bash-only construct that failed silently there. Every test in it is killed by a specific mutation of the code it covers. Both charts' helm-unittest suites pass, the golden-parity script reports byte-identical output across all four cases including the new one, and `make generate` is idempotent at the repo root. Reviewed by Claude across several rounds and by an independent second reader; the review loop found and fixed real defects in this branch, including a green-path check that had become unable to report the one thing it existed for. ### Screenshots Not a UI change. ### Downstream repositories - [ ] No downstream repository is affected by this change - [ ] [cozystack/website](https://github.com/cozystack/website) - follow-up: - [x] [cozystack/terraform-provider-cozystack](https://github.com/cozystack/terraform-provider-cozystack) - follow-up: https://github.com/cozystack/terraform-provider-cozystack/issues/26 - [ ] [cozystack/ansible-cozystack](https://github.com/cozystack/ansible-cozystack) - follow-up: - [ ] [cozystack/ccp](https://github.com/cozystack/ccp) - follow-up: - [ ] [cozystack/talm](https://github.com/cozystack/talm) - follow-up: - [ ] [cozystack/cozyhr](https://github.com/cozystack/cozyhr) - follow-up: - [ ] [cozystack/cozy-proxy](https://github.com/cozystack/cozy-proxy) - follow-up: - [ ] [cozystack/cozystack-telemetry-server](https://github.com/cozystack/cozystack-telemetry-server) - follow-up: - [ ] [cozystack/external-apps-example](https://github.com/cozystack/external-apps-example) - follow-up: - [ ] [cozystack/examples](https://github.com/cozystack/examples) - follow-up: ### Release note ```release-note feat(kubernetes): worker node groups accept `logSerialConsole`, which streams a worker VM's guest serial console into a `guest-console-log` container readable with `kubectl logs`. Off by default. It is the only way to observe a worker that stalls during boot before Talos apid answers, when no Node and no certificate request exist to diagnose from. Enabling it rolls the node group and overrides the platform's cluster-wide `disableSerialConsoleLog`. ``` <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * Added an optional `logSerialConsole` setting for Kubernetes worker node groups. * Streams worker VM guest serial-console output to a dedicated log container when enabled. * Configuration is disabled by default and available through Kubernetes application settings. * **Bug Fixes** * Improved worker failure diagnostics with bounded serial-console output and pod-state information. * Validates that enabled worker pools attach the expected console-log container. * **Documentation** * Documented configuration behavior, rollout effects, defaults, and timeout considerations. * **Tests** * Added coverage for rendering, attachment validation, capture limits, and failure scenarios. <!-- end of auto-generated comment: release notes by coderabbit.ai --> |
||
|
Aleksei Sviridkin
|
50c5e2c14c |
test(e2e): capture the tenant worker guest serial console
The tenant Kubernetes suites fail the node-Ready budget with zero Nodes registered, and the time is lost inside the guest between the worker VMI reaching Running and Talos requesting its apid certificate. Nothing collected today covers that window: the existing in-guest capture drives talosctl, so it needs the very apid the worker never reached, and on the management side there is neither a Node nor a certificate request to read. Turn on the guest console log for the suite's node group and collect each virt-launcher's guest-console-log container from the management cluster. Reads start at the beginning of the stream rather than tailing it, so a guest that repaints its console cannot push the boot output out of the window; every read is bounded, and the walk stops at a cap that records both counts, so the collector cannot become the unbounded term in a failure path that already runs several others under one deadline. A read that returns no console is not filed as a quiet guest. Three outcomes get three labels, because only one of them says anything about the guest: the read failed and returned nothing, the read succeeded and returned nothing, or it broke off part way. Each read's stderr is kept beside the capture rather than inside it, since kubectl's own "container is not valid for pod" would otherwise fill the file and a console that never started would read as one that had spoken. That case is the headline one here, so the Pod state and events that explain it are collected alongside, once for the whole selector. Enabling the console overrides a cluster-wide setting the platform holds for an open upstream defect, so the run is also an experiment, and its own failure mode is the one that hides best: the container wedges virt-launcher in Init, no worker boots, and the suite reports "fewer than 2 tenant nodes Ready" -- the same sentence the failure under study produces. Read as the known signature, the experiment's answer would be filed as the flake it was meant to explain. So the failure path names that case first, before any other diagnostic, and only when it is true. The passing path checks separately that the container attached at all, and reports the ratio it saw rather than a bare success, since the check is deliberately satisfied by any one Pod. It keeps three outcomes apart too: a read that did not answer, a selector that matched no Pod, and Pods that matched and carry nothing. It reads the Pod name alongside the container names so the last two cannot collapse, and both container lists so a future move of the container cannot fail a run over a diagnostic. It runs last, after everything the suite exists to prove, because a debugging aid must not preempt the assertions that cover the cluster. The capture runs before the talosctl one because it is the only of the two that can describe a worker which never reached apid, the shape this failure keeps taking. kubectl reports a warning and an empty result alike, on stderr with a zero status, so the exit status decides which artifact its output lands in. Assisted-By: Claude <noreply@anthropic.com> Signed-off-by: Aleksei Sviridkin <f@lex.la> |
||
|
Aleksei Sviridkin
|
bc6d5e67b7 |
fix(e2e): keep failure snapshots complete and gate tenant patch on aggregated-API auth (#3444)
## What this PR does Two unrelated defects made the sandbox install job's failure handling unreliable. This fixes both. The first hit every red run. When a test fails, the diagnostic snapshot runs `crust-gather collect`, and its host-log leg spawns a privileged debug pod per node. The sandbox runs baseline PodSecurity, which rejects each of those pods with a 403, so crust-gather keeps retrying until it burns its whole collection budget and exits 1. The snapshot verdict then printed `FAILED` on every failed run even though the namespaced object state was already captured, so the line that says whether the uploaded snapshot is trustworthy was always wrong. The host-log leg collects nothing usable here anyway, so this passes `--disable-additional-logs` at all three collect sites: the host snapshot from the BATS runner, the host snapshot from the Chainsaw error handler, and the nested tenant snapshot. Object collection now finishes and the verdict again separates a real collection failure from a complete snapshot. The second is a race in the install test. The "Configure Tenant" step issues its first `kubectl patch` of the root tenant as soon as an earlier test sees the aggregated APIService go `Available`. That condition only tracks endpoint availability, not the serving pod's auth state. A freshly rolled cozystack-api pod can start answering before it has loaded the requestheader client-CA, and until it does it drops the front-proxy identity and answers as `system:anonymous`, so the patch is rejected with a 403 on the test's very first command. This adds a short authenticated GET against the same resource before the patch, so the step proceeds only once the aggregated API both serves and authenticates the request. If the aggregation layer is actually wedged past two minutes the gate fails loud instead of hiding a real regression. Both were traced from a recent sandbox failure (run 30004030331, during PR #3441's CI). Out of scope: the broader convergence flakiness of the install step is not touched here. That covers the long reconcile windows for the storage and platform releases and the app-level races in the Chainsaw suite. The durable product-side fix for the second defect, where cozystack-api should not report ready before the requestheader client-CA is loaded, is a separate follow-up. This partially addresses #3368, its crust-gather diagnostics item, and does not close it. ### Screenshots <!-- REQUIRED for UI changes: attach screenshots or screen recordings demonstrating the visual impact of your changes. PRs with UI changes without screenshots will not be merged. --> No UI changes. ### Downstream repositories <!-- Cozystack is upstream for repositories that are NOT kept in sync with it automatically, and nothing in CI compares the two sides, so a change here can break them silently. Which change forces what, and which file to touch, is the trigger map in docs/agents/contributing.md under "Downstream Repositories". Walk it before you tick anything. IF YOU ARE AN AI AGENT OR AN LLM, THIS IS PART OF THE TASK, NOT A FORM TO FILL IN: 1. Walk the trigger map against the actual diff of this PR, file by file. Do not decide from the PR title or from your memory of what you changed. 2. For every repository you tick, open the follow-up PR there and link it on the same line. A ticked box with no link claims work that does not exist, which is worse than an empty box. Search that repository for an open PR or issue covering it first, and link that instead of filing a duplicate. 3. When the follow-up is out of scope, or needs a decision that is not yours to make, open an issue in that repository instead, link that, and say so in the PR body. Do not push a speculative PR into another repository just to fill a line here. 4. Do not tick "No downstream repository is affected" to make the checklist go away. If you are unsure, leave every box empty and say so in the PR body, so a human decides. --> The diff touches only the e2e test harness (`hack/cozytest.sh`, `hack/e2e-chainsaw/.chainsaw.yaml`, `hack/e2e-chainsaw/_lib/run-kubernetes.sh`, `hack/e2e-install-cozystack.bats`). The trigger map's `hack/` couplings are `hack/e2e-prepare-cluster.bats` (node prerequisites for ansible-cozystack and talm) and `hack/package.mk`, `hack/common-envs.mk`, `hack/update-crd.sh` plus the package layout (ccp, external-apps-example, cozyhr). None of those are moved, renamed, or edited here, and nothing else in the map matches. - [x] No downstream repository is affected by this change - [ ] [cozystack/website](https://github.com/cozystack/website) - follow-up: - [ ] [cozystack/terraform-provider-cozystack](https://github.com/cozystack/terraform-provider-cozystack) - follow-up: - [ ] [cozystack/ansible-cozystack](https://github.com/cozystack/ansible-cozystack) - follow-up: - [ ] [cozystack/ccp](https://github.com/cozystack/ccp) - follow-up: - [ ] [cozystack/talm](https://github.com/cozystack/talm) - follow-up: - [ ] [cozystack/cozyhr](https://github.com/cozystack/cozyhr) - follow-up: - [ ] [cozystack/cozy-proxy](https://github.com/cozystack/cozy-proxy) - follow-up: - [ ] [cozystack/cozystack-telemetry-server](https://github.com/cozystack/cozystack-telemetry-server) - follow-up: - [ ] [cozystack/external-apps-example](https://github.com/cozystack/external-apps-example) - follow-up: - [ ] [cozystack/examples](https://github.com/cozystack/examples) - follow-up: ### Release note <!-- Write a release note: - Explain what has changed internally and for users. - Start with the same `type(scope):` prefix as in the PR title - Follow the guidelines at https://github.com/kubernetes/community/blob/master/contributors/guide/release-notes.md. --> ```release-note fix(e2e): keep failure snapshots complete when PodSecurity blocks crust-gather's host-log leg, and wait for the aggregated API to authenticate before patching the root tenant so the install step no longer flakes on a system:anonymous 403 ``` <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit - **Bug Fixes** - Improved failure diagnostics during automated Kubernetes and tenant tests by avoiding unsupported privileged log collection steps. - Ensured failure snapshots complete more reliably instead of timing out or returning incomplete results. - Added a readiness check before configuring the root tenant, preventing intermittent authorization failures during installation tests. - **Tests** - Increased reliability of installation and end-to-end test workflows under baseline security policies. <!-- end of auto-generated comment: release notes by coderabbit.ai --> |