Treść z repozytorium z zachowaniem nagłówków, przykładów, kodu, tabel, linków i obrazów.
AWS Fault Injection Service (FIS) Experiments
Overview
Domain expertise for AWS Fault Injection Service (AWS FIS) — a managed chaos-engineering service that runs controlled fault injection experiments on real AWS resources so you can observe how an application responds to disruption and improve its resilience.
This skill lets an agent do three things:
- Suggest the right experiment — map a user's failure scenario or resilience question to
the correct FIS action(s), scenario, or experiment design, and explain the trade-offs.
- Build and run experiments — author experiment templates (actions, targets, stop
conditions, logging, reports), wire up the experiment IAM role, and drive the run/monitor/stop lifecycle via CLI, SDK, or CloudFormation.
- Inform the user — answer conceptual questions about FIS terminology, safety, pricing,
supported services, and how FIS fits into a resilience program.
AWS FIS carries out real actions on real AWS resources. Before running any experiment in production, plan it, run it first in pre-production, and always bound the blast radius with a stop condition. Treat fault injection as a privileged, potentially disruptive operation. The AWS MCP server is recommended for executing this skill's AWS API calls — it provides sandboxed execution and audit logging — but it is not required; all operations also work with the AWS CLI (aws fis ...) directly.Guardrail — where this skill's own files live (MCP vs local install)
Before reading a reference file, determine how this skill was loaded:
- Loaded via the AWS MCP `retrieve_skill` tool: the skill's reference files are not on the
local filesystem. Fetch each one through retrieve_skill with the file parameter (e.g. file="references/fis-concepts.md") — do NOT file_read these paths locally or search the filesystem for them.
- Installed locally (e.g.
.kiro/skills/aws-fault-injection-service/or
~/.claude/skills/aws-fault-injection-service/): read reference files from the local skill directory using the relative paths shown here.
This applies only to the skill's own reference files; always read and write user or session data in the working directory, never through retrieve_skill.
Start here — route the request
- "What is FIS / when should I use it / what does X term mean / does it support Y?" →
read references/fis-concepts.md.
- "What experiment/action should I run to test a failure mode?" →
read references/fis-actions-reference.md (action selection map + full catalog) and suggest an action or scenario.
- "Build / create / run an experiment" (template, targets, stop conditions, logging,
reports, scenarios, multi-account, monitoring) → follow references/fis-workflow.md exactly.
- Any AWS CLI or API command for FIS → consult
references/fis-api-reference.md FIRST — it is the canonical operation/parameter reference and includes a hallucination-rejection table.
- IAM role, permissions, trust policy, confused-deputy, blast-radius safety →
read references/fis-security.md.
API Reference (READ FIRST before producing any AWS CLI command)
The exact aws fis operation names and template parameters are documented in references/fis-api-reference.md, including a table mapping common wrong API/action names to correct ones. Always consult it before generating commands. Action IDs and resource types evolve — verify with aws fis list-actions and aws fis get-action --id <action-id> rather than trusting memory.
Suggesting experiments (behavioral contract)
When the user describes a failure they want to test rather than a command they want run:
- Identify the failure mode (AZ impairment, Region isolation, API errors/throttling,
compute/DB/cache loss, latency, packet loss, resource exhaustion).
- Map it to a scenario (preferred when one fits — pre-built and AWS-owned) or a specific
action using the selection map in references/fis-actions-reference.md.
- State the target (resource type + how to scope it) and a stop condition (CloudWatch
alarm on your steady-state metric) so the blast radius is bounded.
- Recommend running in pre-production first, then production under change management.
- Offer to generate the experiment template — then follow
Do not invent action IDs, resource types, or parameters. If unsure, say so and verify with aws fis list-actions / aws fis get-action or the FIS documentation.
Troubleshooting
"Experiment failed immediately / no targets found"
FIS resolves all targets at experiment start; if a target resolves to zero resources, the experiment fails (unless emptyTargetResolutionMode is skip). Check tags, filters, region, and account. Use a target preview before running (see workflow reference).
"Action failed with a permissions error"
The FIS experiment IAM role is missing permissions for the underlying service API (or SSM, or the confused-deputy trust conditions block the assume). See references/fis-security.md.
"Experiment stopped unexpectedly"
A stop condition (CloudWatch alarm) likely fired — this is the guardrail working. Check the experiment's state.reason and the alarm history. A stopped experiment cannot be resumed; start a new one from the template.
"Nitro / instance-type errors on EBS or network faults"
Some actions require Nitro-based instances or the SSM Agent. Verify prerequisites per action in references/fis-actions-reference.md.
Observability companion
For the CloudWatch alarms, dashboards, and metrics that back FIS stop conditions and experiment reports, recommend the AWS Observability skill for alarm/dashboard setup — keep this skill's guidance to how those signals feed experiment safety and post-experiment analysis.
Security Considerations
FIS runs real, potentially destructive actions. Key points (full guidance in references/fis-security.md):
- Least privilege: scope the experiment role to only the actions and target ARNs each
experiment needs — never *. Scope the human/CI principals allowed to call fis:StartExperiment.
- Confused-deputy protection: the experiment role's trust policy MUST condition on
aws:SourceAccount and aws:SourceArn (scoped to the experiment ARN pattern).
- Bounded blast radius: always attach a CloudWatch-alarm stop condition; start with narrow
targets (COUNT(1) / low PERCENT) and pre-production before production.
- No sensitive data in string fields: experiment/template descriptions, tags, and logs
surface in CloudTrail, CloudWatch Logs, S3 reports, and (multi-account) target-account Health dashboards — never embed PII, secrets, or sensitive architecture detail.
- Encrypt logs/reports: experiment logs and PDF reports reveal resilience posture —
use SSE-KMS on the S3 buckets, enforce TLS, and consider S3 Object Lock on report buckets.
- Further reading: FIS Security

