Technical explanation

A is clearly the intended answer because its filters correspond to the displayed source namespace ( endor ), destination port ( 443 ), and verdict ( DROPPED ). However, it contains example.con , whereas the observed flow names example.com . Therefore, none of the options would literally return this exact flow if the FQDN filter is matched as written. The corrected command is:

hubble observe --to-fqdn example.com --from-namespace endor --to-port 443 --verdict DROPPED

Option B is malformed in two additional places and filters port 80 rather than 443. Option C selects the wrong FQDN and source namespace. Option D requests forwarded flows, directly contradicting the Policy denied DROPPED verdict; its wildcard also does not repair the verdict mismatch.

Hubble’s observe filters are cumulative: a returned flow must satisfy the specified destination FQDN, originating namespace, destination port, and verdict. Layer 7 visibility is enabled with a Cilium network policy containing the relevant L7 rules, which redirects selected traffic through the proxy so that application-level details can be reported.

Official references

Inspecting Network Flows with the Hubble CLI , Layer 7 Protocol Visibility

Study Guide topic: Hubble flow filtering, FQDN visibility, namespaces, ports, and verdicts.

Technical explanation

Cilium can implement Layer 3 and Layer 4 forwarding exclusively through its eBPF datapath. IP, TCP, and UDP traffic can be routed, load-balanced, filtered, and redirected through eBPF programs attached to Linux networking hooks. These operations depend on network-layer addresses, protocols, ports, identities, and connection state; they do not require application-protocol parsing by a userspace proxy.

Application-layer operations are different. Cilium’s official Service Mesh architecture uses a proxy such as Envoy to parse HTTP, gRPC, and DNS when Layer 7 policy, observability, or traffic management requires understanding individual requests. The eBPF datapath transparently redirects selected traffic to the node-local proxy and retains identity context, while Envoy performs the protocol-aware operation.

Kafka parsing is also an application-layer activity rather than ordinary Layer 3 or Layer 4 forwarding. Moreover, current Cilium releases removed the former Envoy Go extension mechanism used for Kafka and generic proxylib rules, further reinforcing that it is not an eBPF-only forwarding case.

Therefore, D accurately identifies the scenario in which the Service Mesh datapath can remain entirely in the kernel without requiring Envoy.

Official references

Cilium Service Mesh ; eBPF Datapath Introduction .

Study Guide topic: Service Mesh.

Technical explanation

The official IPAM documentation expressly states that the safest way to change IPAM mode is to install a fresh Kubernetes cluster using the required new IPAM configuration. A live cluster already contains allocated workload addresses, node routing state, Cilium endpoint state, service state, and potentially cloud-provider resources that depend on the active allocator. Replacing the allocator in place can invalidate these assumptions and cause persistent connectivity disruption rather than merely a short agent restart.

Updating a Kubernetes Node object does not generally convert Cilium’s IPAM mode. The responsible resource and allocation behavior depend on the selected mode: Kubernetes host-scope IPAM, cluster-pool IPAM, multi-pool IPAM, CRD-backed modes, and cloud-specific allocators manage address information differently. Restarting agents after changing a configuration value likewise does not constitute a safe migration plan, because existing endpoints and routes may retain state created under the former allocator.

Cilium advises against changing the IPAM mode of an existing cluster unless a specifically documented migration procedure applies. The currently documented exception is migration from cluster-pool IPAM to multi-pool IPAM. As the question provides no such constrained scenario, D is the unambiguously safest answer.

Official references

IP Address Management .

Study Guide topic: Installation and Configuration.

Technical explanation

C is not a valid description. A service-mesh sidecar externalizes networking functions into a proxy container running beside the application; it does not force the instrumentation logic into the application’s source code. In fact, a core service-mesh objective is to provide connectivity, security, traffic management, and observability transparently without requiring application-code changes.

The operational concerns in the other choices are characteristic of sidecar implementations. A proxy must be injected into each workload pod, increasing container count and potentially affecting pod initialization, resource consumption, ordering, and readiness. Adding a sidecar to an existing pod template normally requires the pods to be recreated because Kubernetes cannot dynamically add a new container to an already-running pod.

Traffic interception also directs application traffic through the sidecar proxy and its network namespace paths, adding network-stack traversal and proxy-processing overhead. Cilium’s service-mesh design can instead use node-level Envoy proxies together with eBPF traffic redirection, avoiding one proxy container in every application pod. This preserves transparent application behavior while reducing the per-workload operational burden.

Official references

Cilium Service Mesh , Cilium Ingress and Network Policy Example

Study Guide topic: Sidecar-based and sidecar-free service-mesh architectures.

Technical explanation

Kubernetes kube-proxy conventionally implements Service translation and load balancing through either iptables or IPVS. With Cilium’s kube-proxy replacement enabled, eBPF programs and maps perform Kubernetes Service handling directly in the kernel, including ClusterIP, NodePort, LoadBalancer, ExternalIP, and related service translation functions. C is therefore correct.

The replacement can operate at socket hooks and packet-processing hooks. Service and backend information is stored in eBPF maps, allowing the datapath to select backends and perform address translation without traversing the kube-proxy-generated iptables or IPVS rules normally used for Kubernetes Services.

BGP is not replaced. Cilium’s BGP Control Plane is a separate feature used to advertise routes and service addresses to external routers. Routing itself is also not eliminated; Cilium can implement and accelerate routing decisions with eBPF, but packets still require a valid forwarding model. firewalld is a host firewall-management service and is not the underlying Kubernetes Service implementation replaced by kube-proxy replacement.

Relevant installations must satisfy the kernel and device requirements for Cilium’s eBPF service load-balancer functionality.

Official references

Kubernetes Without kube-proxy

Study Guide topic: kube-proxy replacement, eBPF service maps, iptables, and IPVS.

Technical explanation

Cilium uses Envoy as its userspace Layer 7 proxy. When a Cilium policy contains HTTP or another supported application-layer rule, Cilium’s eBPF datapath identifies matching traffic and redirects it to a node-local Envoy instance. Envoy evaluates the application-layer attributes—such as HTTP method, path, or headers—against the generated policy configuration and then forwards or rejects the request.

The proxy can operate in embedded mode as a separate process inside the Cilium agent pod or as the independently life-cycled cilium-envoy DaemonSet. Both deployment forms use Cilium’s optimized Envoy distribution and custom policy-enforcement filters. Communication between the agent and Envoy uses local UNIX-domain sockets for configuration, access logs, and administrative operations.

HAProxy is a capable general-purpose load balancer, but it is not Cilium’s L7 policy proxy. Squid primarily serves forward and caching proxy use cases. linkerd2-proxy belongs to the Linkerd service mesh and is not used by Cilium for network-policy enforcement. Consequently, only D identifies the proxy integrated into Cilium’s L7 datapath.

Official references

Cilium Envoy , Cilium eBPF Datapath Introduction

Study Guide topic: Envoy proxy integration and Layer 7 policy enforcement.

Technical explanation

If a global Service has no healthy local endpoints matching its selector, every available backend can be remote. Cluster Mesh synchronizes remote service and endpoint information, allowing the local Cilium datapath to load-balance requests to backend pods in connected clusters. The absence of local endpoints therefore provides a direct explanation for the observed behavior.

If the local cluster were not part of the Cluster Mesh, its Cilium agents would not normally receive the remote endpoint state needed to route traffic through the global Service, so B does not explain successful remote-only selection. An affinity value of none is the default behavior and expresses no preference between local and remote endpoints. When both categories exist and are healthy, this permits load balancing across both; it does not require every connection to use remote backends.

Setting service.cilium.io/shared: "false" prevents the local Service’s backends from being shared with remote clusters. It does not instruct the local cluster to direct all requests toward remote endpoints.

A separate possible cause, not presented among the choices, would be service.cilium.io/affinity: "remote" . Among the supplied answers, however, A is the valid explanation.

Official references

Service Affinity ; Cluster Mesh .

Study Guide topic: Cluster Mesh.

Technical explanation

For every pod managed by Cilium, the responsible node-local Cilium agent creates a CiliumEndpoint object with the same name and namespace. The object exposes endpoint state such as labels, the allocated security identity, addressing information, and effective policy. The agents also create endpoint objects for their inter-agent health endpoints. B is therefore correct.

The other responsibilities belong primarily to the Cilium Operator. The operator normally registers Cilium’s CustomResourceDefinitions, manages addresses for LoadBalancer Services when LB IPAM is active, and performs configured cluster-wide synchronization or shared-state operations. It also garbage-collects orphaned CiliumEndpoint resources when their associated pods no longer exist or have permanently completed.

The Cilium agent itself runs on each node. It responds to workload lifecycle events, creates and manages local endpoints, loads eBPF programs, applies policies, and maintains the node’s datapath. This division ensures that latency-sensitive forwarding and endpoint operations remain node-local while logically cluster-wide activities are consolidated in the operator.

Official references

CiliumEndpoint CRD , Cilium Operator Responsibilities

Study Guide topic: Cilium agent, Cilium Operator, and CiliumEndpoint lifecycle.