Chaos Engineering Basics

Overview

Use this skill to design safe, evidence-driven fault injection experiments that verify system resilience under realistic failure conditions.

Scope Boundaries

• Use this skill when the task matches the trigger condition described in description.
• Do not use this skill when the primary task falls outside this skill's domain.

Inputs To Gather

• Critical user journeys and service dependency map.
• Current SLI/SLO and alert signal quality.
• Failure budget, allowed blast radius, and rollback authority.
• Existing runbooks and on-call escalation paths.

Deliverables

• Experiment charter (hypothesis, steady state, blast radius, abort criteria).
• Fault-injection plan (what fails, where, for how long, at what traffic share).
• Observation plan (metrics, logs, traces, and decision thresholds).
• Findings with remediation owners and re-test schedule.

Quick Start Example

Example experiment charter

• Hypothesis: "API p95 remains < 400ms when one cache node fails."
• Steady-state metrics: p95 latency, error rate, queue depth.
• Blast radius: 5% traffic, one AZ only, 10 minutes max.
• Abort immediately if:
  • error rate > 2x baseline for 3 minutes,
  • user checkout success drops below threshold,
  • paging alerts fire in unrelated services.

Example decision rule

• pass: steady-state metrics remain inside pre-registered limits.
• fail: any hard guardrail breaches abort threshold.
• inconclusive: observability gaps prevent causal interpretation.

Quality Standard

• Steady state is measurable and agreed before injection.
• Abort/rollback conditions are explicit and executable.
• Blast radius is bounded by environment, traffic, and time.
• Experiment outcomes produce owned remediation actions.
• Re-test conditions are defined for failed assumptions.

Workflow

1. Select one reliability assumption tied to a business-critical flow.
2. Define steady-state metrics and hard guardrails.
3. Design smallest useful fault experiment with bounded blast radius.
4. Run experiment under live observation with explicit abort authority.
5. Classify result (pass/fail/inconclusive) and capture learning.
6. Assign remediation and schedule follow-up verification.

Failure Conditions

• Stop when steady-state metric or abort threshold is undefined.
• Stop when observability cannot detect degradation quickly.
• Escalate when proposed blast radius exceeds approved risk budget.