Ninth phase of Arc 3. Mediation as a platform service.

A service mesh adds mTLS, observability, traffic shaping, retry/timeout policies, and identity-aware NetworkPolicy to service-to-service traffic without application changes. By phase end every pod on basecamp talks to every other pod over mTLS; NetworkPolicies enforce zero-trust between services; the mesh dashboard shows L7 traffic flow without sidecar overhead.

This phase deepens Phase 18 (Networking Deep) and reinforces Phase 17 (mediation). Cilium’s eBPF mesh layer is also what makes Phase 28 (Observability at Platform Depth) cheap and fast.

Prerequisites

• Phase 24 complete; multi-cloud basecamp working

Why this phase exists

Before a mesh, every service handles its own mTLS, retries, observability, traffic shaping. Each service does it slightly differently; the inconsistency is a security and operational hazard. The mesh moves these concerns to a uniform layer, freeing services to do business logic.

The pattern is mediation: interpose a control point between services so cross-cutting concerns become uniform. The same pattern Kubernetes Services implement at L4; the mesh extends it to L7.

The pattern-first frame

Same eight steps.

1. PROBLEM

Service-to-service traffic on basecamp needs: encryption (mTLS), authentication (workload identity), authorization (who can call who), observability (latency, error rate, throughput per service pair), traffic shaping (retries, timeouts, circuit breakers), and routing (canary deploys, A/B).

Each service implementing these independently produces inconsistency + bugs. The mesh provides them uniformly.

2. PRINCIPLES

2.1 The mesh as universal mediation

A mesh interposes between every service-to-service call. Sidecar-based meshes (Istio classic, Linkerd) inject a proxy per pod. eBPF-based meshes (Cilium) use the kernel.

→ Pattern: service-mesh; reinforces mediation

Investigate:

• What does an Envoy sidecar add to each request hop in resource cost?
• How does Cilium’s eBPF mesh achieve the same outcome without sidecars?
• When does the sidecar model win (rich L7 features, easier debugging)?

2.2 mTLS as the encryption + identity baseline

Mutual TLS authenticates both ends of every connection via certificates. The mesh issues short-lived certificates; rotation is automatic.

Investigate:

• What does the SPIFFE/SPIRE identity model give you?
• Why is “long-lived certificate” the wrong model for service-to-service?
• How does mTLS compose with NetworkPolicy?

2.3 Zero-trust networking

Never trust based on network location. Every connection authenticated; every authorization explicit; every flow auditable.

→ Pattern: zero-trust-networking

Investigate:

• Why is “the perimeter firewall” insufficient in containerized worlds?
• What does workload identity mean (SPIFFE/SPIRE, IRSA, GCP workload identity)?
• How does Cilium implement zero-trust via identity-aware NetworkPolicies?

2.4 L7 vs L4 mediation

The mesh can mediate at L4 (TCP) or L7 (HTTP/gRPC). L4 is cheaper; L7 enables retries, header-based routing, observability.

Investigate:

• When is L4-only enough? (Hint: opaque TCP services.)
• What does L7 routing buy that L4 doesn’t?
• What’s the latency cost of L7 mediation vs L4?

2.5 Traffic shaping (retries, timeouts, circuit breakers at the mesh)

Arc 2 Phase 16 added retries/timeouts/circuit breakers in app code. The mesh can do them uniformly without app changes.

Investigate:

• Why is mesh-level retry better than app-level? (Hint: consistency.)
• When does the mesh’s retry conflict with app retries? (Both fire → amplification.)
• How do you reason about the combined retry budget?

2.6 Mesh observability

The mesh sees every service-to-service call. RED metrics per service pair come for free. Distributed tracing (Phase 28) hooks into the mesh.

Investigate:

• What does the Cilium Hubble UI show that’s hard to get without a mesh?
• How does mesh observability compose with OTel (Phase 28)?
• When does mesh observability replace app-level instrumentation? (It usually augments.)

3. TRADE-OFFS

4. TOOLS (as of 2026-06)

• Cilium 1.16+ with Mesh + Hubble enabled
• cilium CLI
• Linkerd as the simpler compare
• Istio Ambient mode as the sidecar-light option

Reading

• Cilium docs: Service Mesh and NetworkPolicy
• “Istio in Action” (Posta + Maloku)
• “Mastering Service Mesh” (Pugazhenthi)
• The Envoy proxy docs (background; even for non-Envoy meshes)

5. MASTERY: Cilium Mesh on basecamp

[ ] Install Cilium with Mesh enabled on basecamp K3s
[ ] Verify mTLS between two services automatically
[ ] Install Hubble; observe service map
[ ] Add identity-aware NetworkPolicy: only beacon talks to Postgres
[ ] Trigger an mTLS handshake failure deliberately; observe + diagnose
[ ] Configure L7 traffic shaping: 3 retries with backoff for one HTTP service
[ ] Configure canary deploy: 90% to v1, 10% to v2; observe Hubble L7 metrics
[ ] Extend mesh to EKS + GKE clusters (cluster mesh / federation)
[ ] Verify cross-cluster mTLS works
[ ] Add custom CiliumNetworkPolicy with L7 rules (HTTP method-based)

6. COMPARE: Linkerd or Istio

Pick one as the parallel installation. Replicate one slice of basecamp’s traffic. Reflect on what each gets right.

400-word reflection.

7. OPERATE

• 3-4 runbooks: mTLS handshake failing, NetworkPolicy blocking expected traffic, Hubble not showing traffic, cluster-mesh broken
• 1-2 ADRs (Cilium over Istio, sidecar vs ambient)
• Weekly log

8. CONTRIBUTE

• Cilium docs
• Linkerd docs
• A blog post on a real mesh debugging session

What ships from this phase

• Cilium Mesh on basecamp with mTLS + NetworkPolicy + Hubble
• Cross-cluster mesh across K3s + EKS + GKE
• Mesh runbooks

Validation criteria

[ ] Cilium Mesh installed; mTLS verified between two services
[ ] Identity-aware NetworkPolicies enforced
[ ] Hubble service map visible
[ ] Cross-cluster mesh working (or documented why deferred)
[ ] All 10 operational depth checks
[ ] Compare reflection (400 words)
[ ] 3-4 mesh runbooks
[ ] 1-2 ADRs
[ ] Pattern entries:
    - service-mesh → OUTLINE
    - zero-trust-networking → OUTLINE
    - mediation → DEEP
[ ] Exit Test passed

Exit Test

Time: 2.5 hours.

Part 1: Build (75 min)

Add a new service to basecamp. Apply identity-aware NetworkPolicy: only this service can talk to Postgres. Verify with curl from authorized vs unauthorized pod.

Part 2: Diagnose (45 min)

Cilium mTLS handshake failing between two services that should talk. Possible: identity mismatch, NetworkPolicy denying, certificate not propagated.

Part 3: Articulate (30 min)

~600 words: “Defend Cilium Mesh’s eBPF approach over Istio’s sidecar approach for basecamp. Cite resource cost, debugging, L7 feature differences.”

Anti-patterns

Patterns touched this phase

• service-mesh: OUTLINE
• zero-trust-networking: OUTLINE
• mediation: DEEP

→ Next: Phase 26: Platform Engineering, ship ascent