The hub's VMUser (vmauth.yaml) references simple-user-secret via passwordRef,
but the Secret was never added to the hub's manifests. Without this Secret,
the VM operator cannot reconcile the VMUser into the vmauth config, causing
ALL requests to fall through to the unauthorizedUser catch-all (vmsingle).
Result: Vector log shipping to VictoriaLogs was broken — vmauth routed
/insert/elasticsearch/_bulk to vmsingle instead of vlogs-victorialogs.
- Remove all ${DS_VICTORIAMETRICS} uid refs from platform-overview; use
type-only datasource so grafana-operator resolves default prometheus DS
- Replace grafanaCom id:14279 cronjob dashboard with inline custom version
supporting cluster_environment variable (dev/edp/observability)
- Add new GARM runners dashboard (edp-garm) ready for when GARM metrics
are scraped; uses or vector(0) guards so panels show 0 not empty
Note: cluster_environment values confirmed as dev/edp/observability (no benchmark).
GARM metrics not yet present in VictoriaMetrics (0 series found).
Hub defaultRules groups kubernetesSystemControllerManager, kubeScheduler, and
kubernetesSystemScheduler used wrong key 'enabled: false' — chart expects 'create: false'.
This caused KubeControllerManagerDown/KubeSchedulerDown to fire as false positives
because OTC CCE managed k8s does not expose control plane for scraping.
Dev local vmagent had empty externalLabels, so backup-alert rules evaluated by local
vmalert had no cluster_environment label on kube_job_status_failed metrics. Added
cluster_environment=dev to match what the vm-client-stack vmagent adds for hub shipping.
Previous 22.5m deadline caused DeadlineExceeded on 2026-06-19 when
rclone sync took >22m (vs 13-16s prior days). Likely triggered by
significant new data in OBS bucket. 2h window accommodates large
incremental syncs while BackupJobTooSlow alert still fires at 5m.
- Enable VMAgent (was disabled → no metrics scraped)
- Remove disable_login from Grafana config; add security block so operator can auth via API
- Disable VMAuth (invalid trailing-dot hostname o12y.observability.; not needed on dev)
Chart 0.48.1 / operator v0.58.0 uses VLogs CRD for VictoriaLogs, not
VLSingle. The VLSingle kind was introduced in a newer operator version
and is not registered in this chart release. Changing to VLogs which
has identical spec fields (retentionPeriod, removePvcAfterDelete,
storage, storageMetadata, resources all supported).
Pre-upgrade cleanup hook uses bitnami/kubectl and spawns on every ArgoCD
sync. Dev cluster nodes are at 99% CPU / pod limit — hook pod cannot be
scheduled, blocking the entire sync indefinitely.
Disabling cleanup.enabled prevents the hook Job from being created.
CRD cleanup is safe to skip on a fresh bootstrap where no old CRDs exist.
VLSingle CRD missing at sync time — ArgoCD pre-validates all resources
before applying any, causing 'synchronization tasks not valid' on CRs
whose CRDs are created by the operator in the same sync wave.
SkipDryRunOnMissingResource=true bypasses dry-run for missing CRDs,
unblocking the CRD bootstrap deadlock.
Fix CRD bootstrap deadlock on victoria-metrics-k8s-stack ArgoCD app.
Adds prune: true and RespectIgnoreDifferences=true to prevent sync
failures when CRs are applied before CRDs are established.
Remove client-side openssl encryption. OBS SSE-KMS handles encryption at rest.
Updated: no apk add openssl, no openssl enc step, no secrets-backup-config Secret,
upload .tar.gz directly. Image tag bumped to 1.0.1 (built without openssl).
Ref: IPCEICIS-9317
Automated upload (95deeef) overwrote 5 manually-pinned values:
- forgejo-server: restore workflow-webhook-20260305 (DB has v15a/v15b
migrations; rolling back to 14.0.2-edp1-rootless WILL break the DB)
- garm: restore v0.1.7-forgejo-22 (v0.1.7-forgejo-23 has exec format
error — wrong arch build, crashes on OTC CCE amd64 nodes)
- sizer-receiver/secret.yaml: re-add sizer-oidc-client secret (deleted
by upload; causes OIDC auth failure on every sizer-receiver login)
- dex/manifests/dex-sizer-client.yaml: re-add (deleted by upload;
dex cannot resolve sizer OIDC client without this secret)
- dex.yaml: restore manifests source block (removed by upload;
without it ArgoCD never deploys the dex/manifests/ directory)
backup-alerts.yaml (new VMRule from automated upload) is kept as-is.
alpine/k8s:1.28.0 does not ship openssl. Script calls openssl enc
on line 116 causing exit 127 on every run since initial deploy.
Fix:
- apk add --no-cache openssl at script start (defensive, idempotent)
- upgrade image 1.28.0 -> 1.32.0 (kubectl client was 5 minor versions
behind cluster v1.33, outside supported skew of +/-1)
GoReleaser config uses 'dockers_v2' (invalid key, should be 'dockers')
so versioned container images were never pushed. Only :latest exists.
Reverting to :latest until CI pipeline is fixed to publish version tags.
Refs: IPCEICIS-9326
v0.8.2 does not exist — tags go v0.8.1 → v0.8.3.
v0.8.3 introduced RequireOrgMatch middleware that breaks dev env where
repos are under giteaAdmin but OIDC org resolves differently.
Pin to v0.8.1 until IPCEICIS-9326 fixes multi-env org support.
v0.8.3 introduced RequireOrgMatch middleware that breaks dev env where
repos are under giteaAdmin but OIDC org resolves differently.
Pin to v0.8.2 until IPCEICIS-9326 fixes multi-env org support.
The deploy hydration created dex-sizer-client with wrong value.
Reverting to the original shared secret that sizer expects
(73eda906... - active for 81 days before hydration overwrote it).
Changes:
- sizer-oidc-client: restore correct shared secret
- dex-sizer-client: add managed manifest to prevent future drift
- dex.yaml: add manifests source for ArgoCD to sync the secret
Broken by stacks rehydration pipeline run.
Secret mismatch caused infinite login loop on sizer.dev.t09.de.
Added sizer-oidc-client secret manifest to GitOps so ArgoCD manages it.
Value now matches dex-runner-sizer-client (dex side).
DB was migrated to v15 schema by this image in March.
The 14.0.2-edp1-rootless image cannot start against it.
Today's automated pipeline sync triggered pod restart, exposing the mismatch.
