cropalato/proxy-relay-control repository overview

⁠proxy-relay-control

An identity-aware egress relay for multi-tenant Kubernetes clusters that reach the internet through a corporate forward proxy.

⁠The problem

On-prem clusters usually have exactly one way out: a corporate proxy that authenticates its clients. That leaves two unappealing options.

• Allow the node IPs on the proxy. Every pod on every node gets internet access, and the proxy's logs attribute all of it to a node rather than a team.
• Give each tenant a proxy username and password. Unmanageable at namespace scale, and the credential ends up in manifests the tenant controls.

Host-level allowlists do not rescue the first option either. Once raw.githubusercontent.com or a shared artifact host is allowed, it is effectively the whole internet.

⁠What this does

Tenants point http_proxy at the relay. For each connection the relay:

1. identifies the calling workload from its pod IP via the Kubernetes API,
2. authorizes it against a cluster-scoped EgressPolicy — host, port, and optionally URL path and HTTP method,
3. relays the traffic to the corporate proxy using the credentials of the policy that allowed it.

The corporate proxy keeps its existing authentication model. It sees one client presenting a different account per tenant, so its logs and quotas stay attributable, and no tenant ever holds a proxy credential.

tenant pod ──http_proxy──▶ relay ──CONNECT + Proxy-Authorization──▶ corporate proxy ──▶ internet
                             │
                             ├─ identity: pod IP → namespace, pod, service account
                             ├─ policy:   EgressPolicy (host, port, path, method)
                             └─ credential: UpstreamProxy → Secret, per tenant

⁠Path rules and TLS

A CONNECT tunnel shows the relay a host and a port and nothing else, so path rules require terminating TLS. That is opt-in per destination:

• tlsMode: tunnel (default) — bytes are spliced opaquely. Authorization is host and port only. Cert-pinned clients and client-certificate workloads keep working.
• tlsMode: inspect — TLS is terminated at the relay so path and method rules apply. Workloads must trust the relay CA, which the relay publishes as a ConfigMap into every namespace an inspect-mode policy selects.

The origin leg of an inspected connection is always verified normally, against system roots with hostname checking. Interception never becomes a downgrade.

Plain http:// requests carry their path in the clear, so path rules apply there with no interception at all.

⁠Quick start

# 1. CRDs and the relay
helm install relay deploy/helm -n relay-system --create-namespace \
  --set denyCIDRs='{10.244.0.0/16,10.96.0.0/12}'   # your pod and service CIDRs

# 2. The corporate proxy account for one tenant
kubectl -n relay-system create secret generic corp-team-a \
  --from-literal=username=svc-team-a --from-literal=password='...'

# 3. The profile and the grant
kubectl apply -f deploy/examples/

Then, in a tenant pod:

export http_proxy=http://relay.relay-system:3128
export https_proxy=$http_proxy
curl https://api.github.com/zen

⁠Before you trust pod-IP identity

Identity rests on the relay seeing the client's real pod IP. Check it:

kubectl run preflight --rm -it --image=cropalato/proxy-relay-control:0.1.0 \
  --env POD_IP=... -- preflight --url http://relay.relay-system:9090

A mismatch means something is rewriting the source address — kube-proxy --masquerade-all, a CNI masquerade rule, or a hostNetwork client. See docs/operations.md⁠.

⁠Documentation

• docs/policy.md⁠ — writing EgressPolicy and UpstreamProxy
• docs/onboarding.md⁠ — what a tenant has to do
• docs/operations.md⁠ — CA rotation, preflight, troubleshooting
• docs/testing.md⁠ — the end-to-end suite

⁠Status

v0.1.0. Identity is pod-IP based; ServiceAccount-token and mTLS/SPIFFE providers fit the same interface and are the intended hardening path.

Images are published to Docker Hub⁠ as cropalato/proxy-relay-control.

⁠License

MIT⁠.