Chaos Engineering

<objective>
Chaos engineering is the discipline of experimenting on a system to build confidence in its ability to withstand turbulent conditions. It is not random destruction -- it is hypothesis-driven, controlled experimentation that reveals weaknesses before they cause outages. A retry that "works in the demo" silently double-charges customers when the payment API times out; the only way to know is to inject the timeout and watch.
</objective>

Quick Route

Discovery Questions

Check .agents/qa-project-context.md first. If it exists, use it as context and skip questions already answered there.

Environment and readiness:

• Where will chaos experiments run? (Pre-production only, production with approval, never production)
• What is the team's monitoring maturity? Can you detect problems in real time?
• Has the team practiced incident response? Is there a runbook?
• Is there executive buy-in for chaos engineering? (Important for production experiments)

Architecture:

• What is the architecture? (Monolith, microservices, serverless, hybrid)
• What are the critical dependencies? (Database, cache, message queue, third-party APIs)
• Are there single points of failure? (Single database, single region, no redundancy)
• What redundancy and failover mechanisms exist?

Current resilience practices:

• Do services have health checks? What do they check?
• Are there circuit breakers, retry logic, or timeout configurations?
• What happens when a dependency is unavailable? (Graceful degradation, hard failure, unknown)
• Have you experienced unexpected outages? What failed?

Team and culture:

• Is the team comfortable with controlled failure? (Anxiety is normal and should be addressed)
• Who would be the chaos engineering champion? (Needs someone to own the practice)
• What is the appetite for starting? (Start small or dive in)

Core Principles

1. Hypothesis-driven: define expected behavior before injecting

Every chaos experiment starts with a hypothesis: "We believe that if [failure X occurs], the system will [expected behavior Y]." Without a hypothesis, you are just breaking things. The hypothesis names concrete steady-state metrics (baseline metrics) — error rate, latency, throughput — and the bound each may move to.

Example hypothesis: "We believe that if the primary database becomes unavailable, the application will serve cached data for read requests and queue write requests for up to 5 minutes without user-visible errors. Blast radius: staging, one service. Steady-state baseline: error rate <0.1%, P95 latency <300ms."

2. Start small: one service, controlled blast radius

The first chaos experiment should not be "shut down production." It should be "add 200ms latency to one non-critical service in staging." Increase scope gradually as confidence and tooling mature.

3. Monitoring is a prerequisite

If you cannot detect problems in real time, you cannot safely inject failures. Chaos experiments without monitoring are just outages with extra steps. Verify dashboards, alerts, and on-call processes before running any experiment.

4. Game days build muscle memory

Running chaos experiments in automated pipelines is valuable, but game days -- scheduled sessions where the team runs experiments together and practices response -- build the human skills that matter during real incidents.

Chaos Experiment Workflow

Every chaos experiment follows this five-step process.

Step 1: Define steady state hypothesis

Identify the metrics that define "normal" and predict what should happen during the experiment.

Experiment: Database failover
Steady state:
  - Error rate: < 0.1%
  - P95 latency: < 300ms
  - Successful orders per minute: > 50

Hypothesis: When the primary database fails over to the replica,
  - Error rate will spike to < 2% for < 30 seconds
  - P95 latency will increase to < 1s for < 60 seconds
  - No orders will be permanently lost
  - The application will recover without manual intervention

Step 2: Introduce the variable

Inject the failure in a controlled way with a clear scope and duration.

Injection:
  Target: primary database (PostgreSQL)
  Method: block TCP port 5432 on the primary instance
  Scope: single database instance
  Duration: 60 seconds
  Blast radius: staging environment only (first run)

  Abort conditions:
    - Error rate > 10% for > 2 minutes
    - Any data corruption detected
    - Manual abort by experiment owner

Step 3: Observe

During the experiment, monitor all relevant metrics in real time. Assign observers to specific dashboards.

Observation assignments:
  - Engineer A: application error rate and latency dashboard
  - Engineer B: database metrics (connections, replication lag, failover status)
  - Engineer C: application logs (search for database connection errors)
  - Engineer D: business metrics (order count, payment processing)

Step 4: Analyze recovery and data integrity

After the experiment, analyze what happened versus what was expected.

Analysis checklist:
  - Did the system behave as hypothesized? (Y/N, with details)
  - How long was the impact? (Expected vs. actual duration)
  - Were any errors visible to users?
  - Was any data lost or corrupted?
  - Did monitoring and alerting detect the problem correctly?
  - How long before alerts fired?
  - What was the recovery time?

Step 5: Fix and iterate

Document findings, fix resilience gaps, and schedule a re-run to verify the fix.

Findings document:
  Experiment: Database failover (2026-03-20)
  Hypothesis: Confirmed / Partially confirmed / Disproved
  Recovery time: 45s (expected vs actual: expected <10s, actual 45s)
  Data integrity: no rows lost; 3 writes returned 500 instead of queueing

  Findings:
    - Connection pool did not detect stale connections for 45 seconds (expected: <10s)
    - Retry logic worked correctly for read operations
    - Write operations returned 500 errors for 38 seconds (expected: queued)

  Action items (every one has an owner and a due date — no item is deferred):
    - [ ] Configure connection pool health checks — assigned to @maria, due 2026-03-31
    - [ ] Implement write queue with 5-minute buffer — assigned to @dan, due 2026-04-02
    - [ ] Re-run experiment after fixes deployed (re-run scheduled 2026-04-03);
          specific metrics to check on re-run: stale-connection detection <10s,
          zero write 500s, error rate <2%

Failure Injection Types

Network failures

See references/fault-injection.md for the tc netem latency/packet-loss commands and the toxiproxy latency config.

Service failures

See references/fault-injection.md for the kubectl delete pod command and the LitmusChaos pod-delete ChaosEngine manifest.

Infrastructure failures

See references/fault-injection.md for the fallocate disk-fill and stress-ng CPU/memory commands.

Dependency failures

See references/fault-injection.md for the programmatic toxiproxy integration test that disables Redis and asserts graceful degradation.

Tools

Avoid: Chaos Monkey (Netflix) for new projects — low activity, Spinnaker-only path (as of mid-2026). It still works and the repo is not archived, but it only injects instance termination and requires a Spinnaker deployment pipeline. Greenfield work should pick Chaos Mesh, LitmusChaos, or AWS FIS. (The older SimianArmy repo was archived in 2021; don't confuse the two.)

Choosing a tool

Decision tree:
  Running on Kubernetes?
    → Cloud-managed AWS workloads: AWS FIS (cloud-native, IAM-integrated)
    → On K8s with sidecar tolerance: Chaos Mesh (eBPF, fine-grained)
    → On K8s wanting workflows + UI: LitmusChaos (ChaosCenter, Workflows)
    → On K8s lightweight: kube-monkey

  Running on plain VMs / Docker?
    → Docker only: Pumba
    → Linux: tc + stress-ng (manual)

  Need network fault injection in integration tests?
    → toxiproxy (lightweight, programmatic API)

  Need to combine load testing with chaos?
    → k6 with xk6-disruptor extension

  Need managed platform with UI and compliance?
    → Gremlin or Steadybit (both commercial)

GameDay-as-code

The 2026 trend is treating chaos as scheduled CI jobs rather than ad-hoc events: Litmus Workflows, Steadybit reliability hub, and Gremlin Scenarios all let you define a chaos run as YAML and trigger it from CI on a cron. Pair with the Game Day Planning section below — the human practice still matters; the automation just removes the bottleneck of "we never had time to schedule one." For a concrete cron-gated pipeline (nightly pod-delete on an off-peak window) plus automated abort/stop-condition examples, see references/fault-injection.md.

LitmusChaos shipped an MCP Server in October 2025 that connects an AI assistant such as Claude directly to ChaosCenter: you can list, run, and stop experiments in natural language ("run pod-delete on the frontend pods," "stop the network latency experiment") instead of hand-writing YAML. Relevant if your team already drives ops through an AI agent.

Game Day Planning

A game day is a scheduled session where the team runs chaos experiments together, practices incident response, and builds confidence in the system's resilience.

Preparation checklist

2 weeks before:
  - [ ] Define 2-3 experiments to run (don't overload the schedule)
  - [ ] Write hypotheses for each experiment
  - [ ] Get approval from engineering leadership and affected teams
  - [ ] Notify support team and stakeholders
  - [ ] Verify monitoring and alerting are working
  - [ ] Identify rollback procedures for each experiment
  - [ ] Schedule 3-4 hour block (experiments + analysis + retro)

1 day before:
  - [ ] Confirm all participants and their roles
  - [ ] Test that fault injection tools work in the target environment
  - [ ] Verify rollback procedures work (dry run)
  - [ ] Prepare dashboards and observation assignments
  - [ ] Brief the on-call team
  - [ ] Confirm abort criteria for each experiment

Communication and roles

Communicate before (schedule, scope, abort authority), during (live updates every 15 minutes in a dedicated channel), and after (summary within 24 hours with findings and action items).

Assign roles per experiment: experiment owner (runs it, makes abort decisions), observers (application metrics, infrastructure metrics, logs, user experience), and a scribe (records timeline and decisions).

Post-game retrospective

For each experiment: was the hypothesis confirmed? What surprised us? What action items do we have? For the process: did monitoring detect problems? Did alerts fire? Were we comfortable with the blast radius? Close with action items (with owners and due dates) and schedule the next game day.

Starting Small: First Three Experiments

For teams new to chaos engineering, start with these three experiments in a pre-production environment.

Experiment 1: Slow database

Why first: Database latency is the most common cause of user-facing slowness, and the experiment is easy to set up and reverse.

Hypothesis: When database latency increases by 500ms, the application
will remain functional with response times under 3 seconds.

Injection: Add 500ms latency to the database connection using toxiproxy.
Duration: 5 minutes.
Environment: staging.

What to observe:
  - Application response times (should increase by ~500ms, not 10x)
  - Connection pool behavior (should not exhaust connections)
  - Timeout handling (requests should not hang indefinitely)
  - Circuit breaker activation (if implemented)
  - Cache effectiveness (cached reads should be unaffected)

Experiment 2: Third-party API returns 500s

Why second: Third-party dependencies fail regularly, and the application's handling of those failures is often untested.

Hypothesis: When the payment provider returns 500 errors, the
application will show a user-friendly error message and allow
retry without duplicate charges.

Injection: Configure mock/proxy to return 500 for payment API calls.
Duration: 10 minutes.
Environment: staging.

What to observe:
  - Error message quality (user-friendly, not stack traces)
  - Retry behavior (does the application retry? How many times?)
  - Idempotency (retries don't create duplicate transactions)
  - Fallback (is there an alternative payment path?)
  - Monitoring (does the payment failure show up in alerts?)

Experiment 3: Cache unavailable

Why third: Cache failures cause "thundering herd" problems where all traffic suddenly hits the database, often causing cascading failures.

Hypothesis: When Redis becomes unavailable, the application will fall
back to direct database queries with degraded but functional performance.

Injection: Block Redis port using toxiproxy or iptables.
Duration: 5 minutes.
Environment: staging.

What to observe:
  - Database query volume (should increase but not overwhelm)
  - Response times (should increase but remain under 5 seconds)
  - Error rate (cache miss should not cause errors)
  - Connection pool (database connections should not exhaust)
  - Recovery (when cache returns, does the application resume normal behavior?)

Anti-Patterns