Skip to content

NetworkPolicy evaluation

The chart can optionally render a NetworkPolicy, disabled by default. The runtime's network destinations are deployment-specific, so enabling the policy requires explicit ingress peers and egress destinations.

Traffic contract

Direction Traffic Ownership
Ingress Client requests to the runtime Service on TCP 9000 Operator selects clients and any ingress controller or gateway
Egress Provider API requests and model metadata checks Operator configures provider endpoints and allowed destinations
Egress DNS lookups to the cluster DNS service Cluster platform provides DNS and its service IP
Egress OTLP HTTP traces when tracing is enabled Operator supplies the collector endpoint and network path
Ingress Kubernetes probes to /health/* Chart configures probes; cluster networking must permit them
Ingress Metrics scraping at /metrics when enabled Operator-owned scraper and selection policy

The chart cannot infer provider CIDRs, collector addresses, ingress identities, DNS service IPs, or scraper namespaces. Provider credentials and endpoints also remain outside the chart's ownership boundary.

Minimum contract for a future opt-in policy

A policy is disabled by default and lets operators explicitly provide:

  • allowed ingress namespace and pod selectors;
  • allowed egress CIDRs or named destinations for providers and collectors;
  • the cluster DNS namespace/pod selector and port 53 UDP/TCP;
  • an explicit metrics-scraper selector, if metrics ingress is restricted.

It renders only networking.k8s.io/v1, preserves the runtime Service and probe paths, and include contract tests for disabled, ingress-only, and fully-configured modes. It must not silently add 0.0.0.0/0 egress as a security policy disguised as a default.

Current recommendation

Enable the chart policy only when those selectors and destinations are known. Teams with more complex requirements can still apply organization-owned policies in their platform repository, using the traffic table above and validating them against their provider and observability topology.