[RFC] AI-guided fault injection + cross-region latency & Redis cluster failure NetworkChaos actions

Feature request

Background

Chaos engineering today requires practitioners to translate informal resilience goals ("our checkout service should survive a database latency spike") into precise Chaos Mesh YAML. This translation is manual, error-prone, and requires deep knowledge of both the application topology and the Chaos Mesh API surface.

Large language models (LLMs) have demonstrated strong capability in understanding natural-language system descriptions and generating structured configuration from them. Combining Chaos Mesh's rich fault vocabulary with an LLM-backed planning layer could dramatically lower the barrier to running principled chaos experiments.

Feature Request

Add an AI-guided fault injection workflow to Chaos Mesh that allows an operator to describe a resilience hypothesis in natural language and receive a complete, ready-to-apply Workflow YAML.

Proposed user experience

$ chaosctl ai suggest \
    --describe "Simulate a 200ms latency spike between the order service and the payments database, \
                then verify that the circuit breaker opens within 5 seconds"

The AI layer is additive and offline-capable:

1. Intent parsing — the LLM extracts: target services, fault type, magnitude, duration, and verification steps
2. Topology enrichment — chaosctl queries the Kubernetes API to resolve service labels and pod selectors
3. YAML generation — the LLM produces a Workflow YAML with the CRD JSON Schema injected as a system prompt to constrain output
4. Dry-run validation — generated YAML is validated client-side against the CRD schema
5. Apply with confirmation — user reviews then applies

The LLM is called only at step 3. Steps 1, 2, 4, and 5 are deterministic. Architecture supports any OpenAI-compatible API endpoint (including self-hosted via Ollama/vLLM).

New fault patterns (shipped with this RFC as a companion PR)

To demonstrate AI suggestion capability and address long-requested community features, this RFC accompanies a PR adding two new NetworkChaosAction values:

cross-region-latency

Simulates variable WAN latency toward external CIDR ranges representing cloud-provider regional endpoints. Each region gets independent latency, jitter, and bandwidth cap applied simultaneously via separate filtered tc qdiscs.

action: cross-region-latency
crossRegionLatency:
  profiles:
    - regionName: us-east-1
      cidrs: ["52.0.0.0/8", "54.80.0.0/13"]
      latency: 5ms
      jitter: 1ms
    - regionName: eu-west-1
      cidrs: ["54.72.0.0/13", "176.34.0.0/16"]
      latency: 80ms
      jitter: 8ms
      bandwidthRate: "50mbit"

redis-cluster-failure

Injects targeted network faults scoped to Redis protocol ports (6379 + cluster bus 16379) without disturbing unrelated pod traffic. Three modes: partition (drops 100% of Redis traffic), latency (netem delay on Redis port), bandwidth (TBF throttle on Redis port).

action: redis-cluster-failure
redisClusterFailure:
  mode: partition
  redisPort: 6379
  clusterBusPort: 16379

Implementation plan

Phase 1 (companion PR): New fault types

• api/v1alpha1/networkchaos_types_additions.go — type definitions
• controllers/chaosimpl/networkchaos/trafficcontrol/impl_additions.go — ApplyTc case handlers

Phase 2 (follow-on PR): chaosctl ai suggest

• CLI command, OpenAI-compatible HTTP client, CRD-schema-based prompt builder, dry-run validator
• Config: CHAOS_MESH_AI_ENDPOINT + CHAOS_MESH_AI_API_KEY env vars

Phase 3 (optional): UI panel in Chaos Dashboard

Related issues

• #2201 — NetworkChaos: support for multi-region latency simulation
• #1893 — Redis-aware chaos actions

Proposed solution

See implementation plan in the feature request body above. The companion PR (to be linked once opened) contains:

• api/v1alpha1/networkchaos_types_additions.go — CrossRegionLatencySpec, RegionLatencyProfile, RedisClusterFailureSpec, RedisFailureMode types
• controllers/chaosimpl/networkchaos/trafficcontrol/impl_additions.go — applyCrossRegionLatency() and applyRedisClusterFailure() helpers that extend the ApplyTc switch

The AI suggestion layer (Phase 2) will live in a new pkg/ai/ package and chaosctl ai suggest CLI subcommand, supporting any OpenAI-compatible endpoint.

Additional context

No response