Generated at build time from charts/racora-controller/README.md (helm-docs: values.yaml + README.md.gotmpl); edit the source, not this page.

racora-controller

The Racora NRCell controller — a kopf-based K8s operator that reconciles NRCell CRDs into:

• DU Deployments + ConfigMaps + Services in distributed-unit
• CU-CP mobility config overlay (nrcell-cu-cp-config ConfigMap in centralized-unit)
• NRCell status (controller-assigned identity — gnbId, sectorId, gnbDuId, identitySource — plus NCI, pciSource, lastRetune, conditions)

Requires

• racora-crds (NRCell CRD)
• racora-base (racora-system and distributed-unit namespaces)

Install

helm install racora-controller ./charts/racora-controller

Values

Key	Type	Default	Description
namespace	string	"racora-system"	Namespace this controller is deployed into. Must already exist (racora-base creates it).
duNamespace	string	"distributed-unit"	Where the controller writes generated DU resources (Deployments, ConfigMaps, Services). Must already exist (racora-base creates it).
cuNamespace	string	"centralized-unit"	Where the controller writes the cu-cp mobility overlay ConfigMap (nrcell-cu-cp-config). This MUST match the namespace where cu-cp is deployed — the DU and CU are separate planes with separate knobs.
crdNamespace	string	"racora-system"	Where the controller watches for NRCell CRs. Today the controller watches a single namespace; this aligns with how kopf is invoked in main.py (kopf run ... --namespace=racora-system).
coreNamespace	string	"5g-core"	Namespace where the bundled 5G core (Open5GS) runs. The controller derives the CU-CP's default AMF address from it (amf..svc.cluster.local). Must match racora-core's namespace. A customer core is addressed with the umbrella's global.core.amfAddr instead (env RACORA_AMF_ADDR), which overrides this.
ueNamespace	string	"user-equipment"	Namespace where UE simulator pods (srsUE) run. Used as the default ZMQ rx_port host when an NRCell with ruType=zmq doesn't override spec.radioBackend.zmq.rxPort.
gnbId	int	411	CU-wide gNodeB ID. One gNB per CU in the current software, so every cell shares this value — the controller stamps it into each cell's status and uses it for every NCI (NCI = gnbId << 14
image.pullPolicy	string	"IfNotPresent"	Image pull policy — IfNotPresent, because the images ship with the distribution. The image itself is the umbrella's global.components.controller.
mplaneHttpPort	int	5055	TCP port of the mplane sidecar's HTTP endpoints (its /ru/ready is the DU's startup gate). The DU pod runs on the host network, so the port is the node's: keep it clear of anything else on the radio's node.
o1.enabled	bool	true	O1 on every element: a netconf sidecar (the O1 NETCONF server, global.components.netconf, seeded from the config the controller renders) and an o1 sidecar (the O1 adapter in the element's role) next to each gNB container, which reads the config the adapter renders from the datastore. Off: the gNB reads the controller's ConfigMap directly.
o1.httpPort	int	5056	TCP port of the o1 sidecar's HTTP endpoints (/status, /config-healthy, the gNB container's liveness) in every element pod; on an ofh DU's host network it is the node's.
o1.netconfPort	int	830	The NETCONF port every element serves (SSH), on the element's Service; on an ofh DU's host network it is the node's.
o1.accountSecret	string	"o1-netconf-account"	The Secret (username, password) with the NETCONF account every element serves, generated by this chart into the DU and CU namespaces and the controller's own, where the controller mounts it.
o1.accountUsername	string	"racora"	The username of that account.
o1.hostkeySecret	string	"o1-hostkey"	The Secret (ssh_host_ecdsa_key) with the SSH host key every element's NETCONF server presents, generated by this chart into the DU and CU namespaces.
o1.checkSeconds	int	30	Seconds between the controller's readings of every element's O1 datastore: a cell's O1Ready and O1Config conditions, the snapshot it keeps in o1-snapshot-<element>, and the merge-back of what was added over O1 after a pod starts again.
o1.ves.url	string	"http://o1-collector.racora-smo.svc.cluster.local:8080"	The VES collector's base URL for the adapters the controller launches when this chart is installed on its own; under the umbrella global.o1.ves.url wins.
o1.pm.streamTarget	string	"http://o1-collector.racora-smo.svc.cluster.local:8080/pm/stream"	Where the racora-default PerfMetricJob streams performance data when this chart is installed on its own; under the umbrella global.o1.pm.streamTarget wins.
oruResource	string	"racora.io/oru"	The device-plugin resource the fronthaul agent (racora-node) advertises per O-RU; an ofh DU with an M-plane address requests one, so it schedules only where its O-RU is up. Empty when the agent is not deployed (nodePlugins.fronthaul.enabled=false), or no ofh DU schedules.
ofhUnreachableGraceSeconds	int	90	Seconds the O-RU may be unhealthy under a running DU (RadioReady=False/Unreachable) before the controller replaces the DU pod; the replacement waits until the agent reports the O-RU healthy, then provisions it again.
duServedCheckSeconds	int	30	Seconds between the controller's checks of which DUs the CU-CP serves over F1 (its metrics stream), written to every cell as DUServed.
duUnservedGraceSeconds	int	180	Seconds a ready DU may stay unserved by the CU-CP (DUServed=False/NotServed) before the controller replaces its pod; a fresh DU attaches over F1 like a new cell.
mobility.triggerHandoverFromMeasurements	bool	true	Automatic A3/measurement-driven handover at the CU-CP. CU-wide and boot-only in OCUDU: the controller bakes it into the CU-CP overlay, so a change takes effect at the next cu-cp restart — which is why it lives here and not on the NRCell spec.
mobility.resyncIntervalS	int	300	Cadence (seconds) of the idempotent upsert-only mobility re-push to the running CU-CP — belt-and-braces convergence for edge cases the per-reconcile runtime sync can miss.
cucpMetricsPeriodMs	int	5000	CU-CP metrics report period (ms), baked into the CU-CP overlay. Without it the CU-CP emits no periodic RRC/NGAP counter lines (handover requested/successful etc.) and the monitoring dashboards starve. Boot-only: takes effect at the next cu-cp restart.
decisions.pollIntervalS	int	1	Seconds between /decisions/active polls (per NRCell timer).
decisions.minConfidence	float	0.95	Minimum decision confidence to act on (deterministic functions emit 1.0; load-bearing once learned models emit decisions).
decisions.minStabilizationS	int	30	Seconds a cell's last applied change (per function) must age before the next decision for that function may apply.
decisions.perFunction	object	{}	Per-function overrides of the two knobs above, keyed by function name ('pci', 'anr').
replicaCount	int	1	Single replica — kopf-based reconciler, no HA. Scale-out would require leader election which kopf supports but the controller isn't wired for it yet.
livenessProbe.enabled	bool	true	Serve kopf's health endpoint from the operator's event loop and probe it: a controller whose event loop stops (a deadlock, a hung call) stops answering and the kubelet restarts it.
livenessProbe.port	int	8080	Port of the health endpoint (/healthz), inside the controller pod.
livenessProbe.initialDelaySeconds	int	30	Seconds after start before the first probe.
livenessProbe.periodSeconds	int	20	Seconds between probes.
livenessProbe.timeoutSeconds	int	5	Seconds a probe may wait for the answer.
livenessProbe.failureThreshold	int	3	Failed probes in a row before the kubelet restarts the controller.
resources	object	{"limits":{"memory":"512Mi"},"requests":{"cpu":"100m","memory":"256Mi"}}	Resource requests/limits. The controller is light: kopf event loop + small in-memory state per NRCell. Defaults sized for ~50 cells.
extraEnv	list	[]	Extra env vars (e.g., for the CU-IP Flight URL override). Each entry is rendered into the container's env list verbatim.
controlPlanePin	bool	true	Pin both reconcilers (racora-controller, racora-core-controller) to the control node, like the CU planes and the in-cluster core: "first node is the brain", additional nodes are DU/RF capacity that comes and goes with the radios. Adds node-role.kubernetes.io/control-plane=true to the nodeSelector; platforms with tainted control planes set it false and pin with nodeSelector instead (platforms/kubernetes/values.yaml).
nodeSelector	object	{}	Node selector for the controller pod.
tolerations	list	[]	Tolerations for the controller pod.
affinity	object	{}	Affinity rules for the controller pod.
serviceAccount.create	bool	true	Create the ServiceAccount, ClusterRole and binding the controller needs (cross-namespace ConfigMaps, Deployments and Services; NRCell status).
serviceAccount.name	string	""	ServiceAccount name; empty = the chart-generated default.
coreController.enabled	bool	true	Deploy racora-core-controller (the Subscriber reconciler and core provider adapter).
coreController.image.pullPolicy	string	"IfNotPresent"	Image pull policy — IfNotPresent, because the images ship with the distribution. The image itself is the umbrella's global.components.coreController.
coreController.resyncSeconds	int	300	Seconds between resyncs that re-apply every Subscriber to the core (the provider's apply is an upsert, so a wiped core database heals).
coreController.retrySeconds	int	30	Seconds before a Subscriber whose provider was not ready (or whose adapter failed) is retried.
coreController.resources	object	{"limits":{"memory":"256Mi"},"requests":{"cpu":"50m","memory":"128Mi"}}	Resources for the core controller: a light kopf loop.

Verify

kubectl get pods -n racora-system -l app=racora-controller
kubectl logs -n racora-system -l app=racora-controller --tail=20

# Apply a test NRCell to confirm reconciliation (see the repo README
# for a copy-pasteable example cell):
kubectl get nrcell -n racora-system
kubectl get deployment -n distributed-unit

RBAC

The controller's ClusterRole grants cross-namespace writes for ConfigMaps + Deployments + Services. Today scoped cluster-wide for simplicity; in a tighter security posture this could be narrowed to just distributed-unit + centralized-unit via RoleBindings.

The core controller

The same chart deploys racora-core-controller, the core-side reconciler: a second, small kopf service owning the Subscriber CRD. Each Subscriber is provisioned into the selected core provider through the adapter the provider declares (the racora-core-provider ConfigMap in coreNamespace; see cores/README.md), retried while the provider is not ready, and re-applied on a cadence (coreController.resyncSeconds) so a wiped core database heals. Its RBAC is separate: Subscribers, the credentials Secrets in crdNamespace, and pods/exec in coreNamespace — the RAN controller holds none of those.