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

Situation	Go to
First experiment ever, team is new	Starting Small → First Three Experiments
Designing one experiment	Chaos Experiment Workflow (5 steps)
Picking a tool for your environment	Tools → Choosing a tool decision tree
Running a team session	Game Day Planning
Need runnable injection commands/configs	references/fault-injection.md
Want the abort to fire without a human	references/fault-injection.md → Automated abort

---

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

Failure	Tool	Use Case
Latency injection	tc, toxiproxy, Gremlin	Simulate slow network, distant regions
Packet loss	tc netem, Chaos Mesh	Simulate unreliable network
DNS failure	iptables, CoreDNS manipulation	Simulate DNS outage
Network partition	iptables, Chaos Mesh	Simulate split-brain scenarios
Bandwidth restriction	tc, toxiproxy	Simulate congested network

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

Service failures

Failure	Method	Use Case
Service crash	Kill process, pod delete	Simulate unexpected crash
Service slowdown	CPU stress, thread pool exhaustion	Simulate overloaded service
Error injection	Return 500/503, throw exceptions	Simulate application errors
Memory pressure	stress-ng, Chaos Mesh	Simulate memory leaks

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

Infrastructure failures

Failure	Method	Use Case
Disk full	fallocate, dd	Simulate disk exhaustion
CPU exhaustion	stress-ng	Simulate CPU saturation
Memory exhaustion	stress-ng	Simulate OOM conditions
Clock skew	chrony manipulation, timedatectl	Simulate time drift

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

Dependency failures

Failure	Method	Use Case
API down	toxiproxy, mock server	Simulate third-party outage
Database unavailable	block port, kill process	Simulate database outage
Cache unavailable	block Redis port	Simulate cache miss storm
Message queue full	fill queue, block consumers	Simulate backpressure

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

---

Tools

Tool	Type	Best For
LitmusChaos (3.29.x)	Kubernetes-native, CNCF	K8s environments, CI/CD integration; ChaosCenter UI; Workflows for GameDay-as-code; MCP Server (Oct 2025) drives experiments from an AI assistant
Chaos Mesh (2.8.x)	Kubernetes-native, CNCF	K8s with fine-grained control; eBPF chaos via bpfki runtime for kernel-precision faults
AWS FIS	Managed AWS service	Cloud-chaos for AWS workloads (EC2, ECS, RDS, EKS); CloudWatch-alarm stop-conditions for auto-abort — primary cloud-native option
Gremlin	Managed platform	Teams wanting guided experiments + compliance reporting; Health Checks halt-and-rollback on SLO breach
Steadybit	Managed platform	Reliability hub spanning Kubernetes + cloud + on-prem; direct alternative to Gremlin
kube-monkey	Open source	Lightweight K8s alternative when Litmus/Chaos Mesh feel heavy
Pumba	Open source	Docker-only chaos (containers, networks); pre-K8s and edge
toxiproxy	Network proxy, open source	Network fault injection in integration tests
tc (traffic control)	Linux kernel	Network latency and packet loss
stress-ng	Linux utility	CPU, memory, disk stress testing
k6 (+ xk6-disruptor)	Load testing tool	Combined load + chaos scenarios

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

Chaos without monitoring

Injecting failures without the ability to observe their impact is not chaos engineering -- it is sabotage. You will not know if the experiment revealed a problem until a user complains.

Fix: Before any chaos experiment, verify that you can see error rates, latency, throughput, and dependency health in real time. If you cannot, invest in monitoring first. Go one step further and wire the monitor into the experiment so it auto-aborts on breach — AWS FIS CloudWatch stop-conditions, Gremlin Health Checks, or a Litmus promProbe in mode: Continuous (see references/fault-injection.md).

Starting too big

The first chaos experiment should not be "kill the production database." Starting with high-impact experiments before the team has practiced with low-impact ones creates anxiety and potential real outages.

Fix: Start with staging. Start with non-critical services. Start with reversible injections (latency, not data corruption). Build confidence gradually. Graduate to production only after multiple successful staging experiments.

No rollback plan

"The experiment is only 60 seconds, we don't need a rollback plan." Then the fault injection tool crashes and the failure persists indefinitely (exactly how a 5-minute tc latency injection becomes a 2-hour outage when the command fails midway).

Fix: Every experiment must have a documented rollback procedure that can be executed in under 30 seconds. Test the rollback before running the experiment, and have a second person ready to abort if the experiment owner is unable to. Prefer a tool-enforced stop-condition over a human finger on the kill switch (see Automated abort in references/fault-injection.md) — and pick injections with a built-in timeout (stress-ng --timeout, Litmus TOTAL_CHAOS_DURATION) so the fault self-clears even if nobody is watching.

Chaos in production without approval

Running chaos experiments in production without explicit approval from engineering leadership and affected teams destroys trust and careers.

Fix: Before any production run: get explicit, documented engineering leadership approval; send affected teams notification and brief the on-call team and support team; communicate scope and duration and share the abort criteria with named abort authority; and confirm the blast radius is bounded to the smallest viable target. Start with staging first. Anything skipped here is what turns a controlled experiment into an incident.

Running chaos experiments during incidents

Adding controlled failures to a system that is already experiencing problems makes diagnosis harder and extends the outage.

Fix: Cancel or postpone chaos experiments if the system is not in steady state. Check for active incidents before starting. If an unrelated incident starts during an experiment, abort the experiment immediately.

No follow-through on findings

The experiment revealed that the circuit breaker does not work correctly. The team says "interesting" and moves on. The finding is never fixed. The next real outage triggers the same failure.

Fix: Every chaos experiment finding gets a ticket with an owner and a due date. Re-run the experiment after the fix to verify. Track the backlog of chaos findings alongside production incident action items.

---

Verification

Prove the safety machinery works before trusting a single live experiment — an abort path you have never triggered is a hypothesis, not a control. Smallest check first:

1. Inject and reverse in staging. Run the lowest-blast-radius injection (e.g. tc qdisc add dev eth0 root netem delay 200ms) against one staging service, confirm the steady-state dashboard moves, then run the documented rollback (tc qdisc del dev eth0 root) and confirm metrics return to baseline. Time the rollback — it must complete inside its stated window (under 30s).
2. Trip the automated abort. Force the monitored metric past its threshold (push error rate over the stop-condition) on an experiment wired to AWS FIS CloudWatch stop-conditions, a Gremlin Health Check, or a Litmus promProbe in mode: Continuous. Confirm the experiment halts on its own with no human action. If it does not fire, the abort is unverified (see references/fault-injection.md).
3. Confirm the self-clearing timeout. Start a bounded injection (stress-ng --cpu 4 --timeout 60s, or Litmus TOTAL_CHAOS_DURATION: '60'), then walk away. Confirm the fault clears itself at the deadline even if nobody aborts.
4. Confirm the findings loop closes. After a staging run, confirm a findings doc was produced with baseline-vs-actual numbers and that every gap became a tracked ticket with an owner — the experiment record is the artifact, not the injection.

If steps 1–3 cannot be demonstrated in staging, the experiment is not safe to promote toward production.

---

Done When

• Every experiment has a written hypothesis ("We believe that if [failure X], the system will [expected behavior Y]") recorded before any fault is injected
• Blast radius is explicitly bounded — target scope, duration, and abort conditions are written in the experiment record, and no experiment runs in production before at least one passing run of the same experiment in staging
• A steady-state snapshot (dashboard link or the actual baseline metric values) is attached to each experiment record, captured immediately before injection
• A findings doc exists per experiment with baseline-vs-actual numbers, recovery time, and an explicit data-integrity check (lost/corrupted: yes/no)
• Each weakness found has a tracked ticket with a named owner, a due date, and a scheduled re-run of the same experiment to verify the fix

Reference Files (in references/)

• fault-injection.md — Runnable commands, configs, and test code for injecting each failure class: tc netem and toxiproxy network faults, kubectl/LitmusChaos service faults (ChaosEngine with engineState: active), fallocate/stress-ng infrastructure faults, the programmatic toxiproxy dependency-failure test, plus automated abort / stop-condition examples (AWS FIS, Gremlin, Litmus probes, Litmus MCP) and a cron-gated continuous-chaos CI job.

Related Skills

• testing-in-production — for safe-rollout mechanics during a release (feature flags, canary, dark launch). Chaos engineering deliberately breaks things to find weaknesses; testing-in-production controls exposure so a breakage stays contained. Go there for the rollout, here for the fault injection.
• observability-driven-testing — when production telemetry (traces, logs, error patterns) is the input that tells you which tests or experiments to design. It feeds the hypothesis; chaos engineering then proves or disproves it. Observability is also a hard prerequisite for safe chaos.
• performance-testing — load testing complements chaos; combine load + fault for realistic failure scenarios (see k6 + xk6-disruptor).
• release-readiness — chaos experiment results feed into go/no-go release confidence assessments.
• test-environments — pre-production environments are the safe starting point for chaos experiments.
• qa-metrics — chaos experiment results (recovery time, error impact) are quality metrics worth tracking.