celld-operator

Operator configuration

Every operator flag, its Helm value, and what happens when an optional integration is absent. Flags configure the operator once; per-app behavior lives in the WorkerApp resource.

Trust and scope

These flags decide which WorkerApps the operator accepts. A WorkerApp outside them reports SpecValid: False and nothing is created for it.

FlagHelm valueDefaultPurpose
--allowed-image-prefixesoperator.allowedImagePrefixesghcr.io/denoland/celld:Comma-separated prefixes a celld.image must start with.
--allowed-iam-rolesoperator.allowedIAMRolesemptyExact role ARNs allowed in credentialsFrom.iamRole.
--allowed-azure-client-idsoperator.allowedAzureClientIDsemptyExact Azure workload identity client IDs allowed in credentialsFrom.azureClientID.
--allowed-bucket-endpointsoperator.allowedBucketEndpointsemptyExact HTTPS endpoints allowed in bucket.endpoint; empty permits AWS S3 only.
--watch-namespaceoperator.watchNamespaceempty (cluster-wide)Restrict reconciliation, the cache, and the chart's workload Role to one namespace.

Ingress and routing

FlagHelm valueDefaultPurpose
--ingress-modeoperator.ingressModehttproutehttproute (Gateway API), virtualservice (classic Istio), ingress (networking.k8s.io/v1), or none. See ingress modes.
--gateway-nameoperator.gatewayNameedgeShared Gateway that HTTPRoutes attach to.
--gateway-namespaceoperator.gatewayNamespaceinfraNamespace of that Gateway.
--httproute-retriesoperator.httpRouteRetriesfalseAdd a retry on 503 to HTTPRoutes. Requires experimental-channel Gateway API CRDs and a gateway that implements it.
--istio-gatewaysoperator.istioGatewaysunsetPre-existing networking.istio.io Gateways (namespace/name, comma-separated) for virtualservice mode.
--ingress-classoperator.ingressClasscluster defaultIngressClass for ingress mode.
--cluster-issueroperator.clusterIssuerunsetcert-manager ClusterIssuer for ingress mode; each app's Ingress then requests its own TLS certificate.

Fleet access and polling

FlagHelm valueDefaultPurpose
--operator-namespacerelease namespacecelld-operator-systemNamespace whose manager pods fleet NetworkPolicies admit to :8081.
--operator-principaloperator.operatorPrincipalderivedOperator identity in Istio AuthorizationPolicies; empty derives cluster.local/ns/<release ns>/sa/<service account>. Set it when your trust domain is not cluster.local.
--state-poll-intervaloperator.statePollInterval15s/state polling cadence for metrics export. Keep it well below 60s, the freshness limit on scaling queries.
--deploy-poll-intervaloperator.deployPollInterval60sBucket deploy/current.json cadence for appVersion: auto and drift checks.
--prometheus-urloperator.prometheusURLhttp://prometheus-operated.monitoring.svc:9090Where KEDA queries the operator's celld_* metrics.

The operator exits at startup on an unknown ingress mode or a non-positive poll interval.

Other chart values

ValueDefaultPurpose
controllerManager.replicas1Manager replicas; leader election is on (--leader-elect), and only the leader polls fleets.
controllerManager.container.image.repositoryghcr.io/ezgamehost/celld-operatorOperator image.
controllerManager.container.image.tag""Empty uses the chart's appVersion, the operator build the chart was released with.
controllerManager.container.resources10m / 128Mi requests, 500m / 512Mi limitsA starting point, not a capacity guarantee: size it with a load test for your fleet count.
metrics.enable / metrics.port / metrics.insecuretrue / 8443 / falseThe manager's metrics endpoint. Secure mode serves HTTPS with authn/authz; insecure mode serves HTTP.
prometheus.enablefalseCreate a ServiceMonitor (with honorLabels, so the namespace/pod labels survive).
certmanager.enablefalsecert-manager certificates for the metrics endpoint, with CA-verified scraping. Do not combine with insecure metrics.
networkPolicy.enablefalseA NetworkPolicy that admits metrics scrapes only from namespaces labelled metrics: enabled.
crd.enable / crd.keeptrue / trueInclude the CRD, and keep it on uninstall.
rbac.enabletrueInclude RBAC objects.

The manager runs non-root with a restricted security context and exposes its own health probes on :8081 (not celld's). Keep the control-plane: controller-manager and app.kubernetes.io/name: celld-operator labels on manager pods: fleet NetworkPolicies select them.

Graceful degradation

The operator requires none of its optional integrations to start reconciling. Absence is reported per WorkerApp, not failed on:

MissingEffectCondition
Gateway API CRDs (httproute mode)Hostnames are not routed; the fleet still serves in-clusterIngressReady: False
Istio CRDsNetworkPolicy alone guards :8081; no mesh mTLSMeshPolicyReady: False
KEDA (autoscaled apps only)spec.autoscaling has no effectAutoscalingReady: False
PrometheusKEDA scalers error; no replica changes are made on missing datanone: alert on KEDA

Store qualification tooling

The repository ships hack/cas-hammer, the concurrency probe for qualifying object stores:

$ go run ./hack/cas-hammer --bucket <bucket> --endpoint <url> --writers 8 --rounds 32

Development

$ make test     # envtest suite + unit tests
$ make lint     # golangci-lint
$ make run      # run against the current kubeconfig

Contributions are welcome; see CONTRIBUTING.md.

Edit this page on GitHubUpdated 2026-10-18