Cilium Operator

This document provides a technical overview of the Cilium Operator and describes the cluster-wide operations it is responsible for.

Highly Available Cilium Operator

The Cilium Operator uses Kubernetes leader election library in conjunction with lease locks to provide HA functionality. The capability is supported on Kubernetes versions 1.14 and above. It is Cilium’s default behavior since the 1.9 release.

The number of replicas for the HA deployment can be configured using Helm option operator.replicas.

helm install cilium cilium/cilium --version 1.15.4 \
  --namespace kube-system \
  --set operator.replicas=3
$ kubectl get deployment cilium-operator -n kube-system
cilium-operator   3/3     3            3           46s

The operator is an integral part of Cilium installations in Kubernetes environments and is tasked to perform the following operations:

CRD Registration

The default behavior of the Cilium Operator is to register the CRDs used by Cilium. The following custom resources are registered by the Cilium Operator:

  • CiliumBGPAdvertisement

  • CiliumBGPClusterConfig

  • CiliumBGPNodeConfig

  • CiliumBGPNodeConfigOverride

  • CiliumBGPPeerConfig

  • CiliumBGPPeeringPolicy

  • CiliumCIDRGroup

  • CiliumClusterwideEnvoyConfig

  • CiliumClusterwideNetworkPolicy

  • CiliumEgressGatewayPolicy

  • CiliumEndpoint

  • CiliumEndpointSlice

  • CiliumEnvoyConfig

  • CiliumExternalWorkload

  • CiliumIdentity

  • CiliumL2AnnouncementPolicy

  • CiliumLoadBalancerIPPool

  • CiliumLocalRedirectPolicy

  • CiliumNetworkPolicy

  • CiliumNode

  • CiliumNodeConfig

  • CiliumPodIPPool


Cilium Operator is responsible for IP address management when running in the following modes:

When running in IPAM mode Kubernetes Host Scope, the allocation CIDRs used by cilium-agent is derived from the fields podCIDR and podCIDRs populated by Kubernetes in the Kubernetes Node resource.

For CRD-Backed IPAM allocation mode, it is the job of Cloud-specific operator to populate the required information about CIDRs in the CiliumNode resource.

Cilium currently has native support for the following Cloud providers in CRD IPAM mode:

  • Azure - cilium-operator-azure

  • AWS - cilium-operator-aws

For more information on IPAM visit IP Address Management (IPAM).

Load Balancer IP Address Management

When LoadBalancer IP Address Management (LB IPAM) is used, Cilium Operator manages IP address for type: LoadBalancer services.

KVStore operations

These operations are performed only when KVStore is enabled for the Cilium Operator. In addition, KVStore operations are only required when cilium-operator is running with any of the below options:

  • --synchronize-k8s-services

  • --synchronize-k8s-nodes

  • --identity-allocation-mode=kvstore

K8s Services synchronization

Cilium Operator performs the job of synchronizing Kubernetes services to external KVStore configured for the Cilium Operator if running with --synchronize-k8s-services flag.

The Cilium Operator performs this operation only for shared services (services that have annotation set to true). This is meaningful when running Cilium to setup a ClusterMesh.

K8s Nodes synchronization

Similar to K8s services, Cilium Operator also synchronizes Kubernetes nodes information to the shared KVStore.

When a Node object is deleted it is not possible to reliably cleanup the corresponding CiliumNode object from the Agent itself. The Cilium Operator holds the responsibility to garbage collect orphaned CiliumNodes.

Heartbeat update

The Cilium Operator periodically updates the Cilium’s heartbeat path key with the current time. The default key for this heartbeat is cilium/.heartbeat in the KVStore. It is used by Cilium Agents to validate that KVStore updates can be received.

Identity garbage collection

Each workload in Kubernetes is assigned a security identity that is used for policy decision making. This identity is based on common workload markers like labels. Cilium supports two identity allocation mechanisms:

  • CRD Identity allocation

  • KVStore Identity allocation

Both the mechanisms of identity allocation require the Cilium Operator to perform the garbage collection of stale identities. This garbage collection is necessary because a 16-bit unsigned integer represents the security identity, and thus we can only have a maximum of 65536 identities in the cluster.

CRD Identity garbage collection

CRD identity allocation uses Kubernetes custom resource CiliumIdentity to represent a security identity. This is the default behavior of Cilium and works out of the box in any K8s environment without any external dependency.

The Cilium Operator maintains a local cache for CiliumIdentities with the last time they were seen active. A controller runs in the background periodically which scans this local cache and deletes identities that have not had their heartbeat life sign updated since identity-heartbeat-timeout.

One thing to note here is that an Identity is always assumed to be live if it has an endpoint associated with it.

KVStore Identity garbage collection

While the CRD allocation mode for identities is more common, it is limited in terms of scale. When running in a very large environment, a saner choice is to use the KVStore allocation mode. This mode stores the identities in an external store like etcd.

For more information on Cilium’s scalability visit Scalability report.

The garbage collection mechanism involves scanning the KVStore of all the identities. For each identity, the Cilium Operator search in the KVStore if there are any active users of that identity. The entry is deleted from the KVStore if there are no active users.

CiliumEndpoint garbage collection

CiliumEndpoint object is created by the cilium-agent for each Pod in the cluster. The Cilium Operator manages a controller to handle the garbage collection of orphaned CiliumEndpoint objects. An orphaned CiliumEndpoint object means that the owner of the endpoint object is not active anymore in the cluster. CiliumEndpoints are also considered orphaned if the owner is an existing Pod in PodFailed or PodSucceeded state. This controller is run periodically if the endpoint-gc-interval option is specified and only once during startup if the option is unspecified.

Derivative network policy creation

When using Cloud-provider-specific constructs like toGroups in the network policy spec, the Cilium Operator performs the job of converting these constructs to derivative CNP/CCNP objects without these fields.

For more information, see how Cilium network policies incorporate the use of toGroups to lock down external access using AWS security groups.

Ingress and Gateway API Support

When Ingress or Gateway API support is enabled, the Cilium Operator performs the task of parsing Ingress or Gateway API objects and converting them into CiliumEnvoyConfig objects used for configuring the per-node Envoy proxy.

Additionally, Secrets used by Ingress or Gateway API objects will be synced to a Cilium-managed namespace that the Cilium Agent is then granted access to. This reduces the permissions required of the Cilium Agent.

Mutual Authentication Support

When Cilium’s Mutual Authentication Support is enabled, the Cilium Operator is responsible for ensuring that each Cilium Identity has an associated identity in the certificate management system. It will create and delete identity registrations in the configured certificate management section as required. The Cilium Operator does not, however have any to the key material in the identities.

That information is only shared with the Cilium Agent via other channels.