aws-service-chaos-research（AWS service chaos 研究）

AWS Service-Specific Chaos & HA Testing Research

Generate comprehensive chaos engineering and high availability testing scenarios for a specific AWS service. Uses a Scenario-Library-first approach: read the latest FIS Scenario Library documentation for pre-built composite scenarios first, then query individual FIS actions via list-actions, and finally supplement with deep documentation research.

Output Language Rule

Detect the language of the user's conversation and use the same language for all output.

• Chinese input -> Chinese output
• English input -> English output
• Mixed -> follow the dominant language

Prerequisites

Required tools (at least one of each group):

FIS Scenario Library (Group A — documentation-based, always available):

• aws___read_documentation — read FIS Scenario Library pages directly (scenarios are console-only and cannot be queried via CLI, so reading the latest docs is the only way to discover them)

FIS Actions Discovery (Group B — use in order of preference):

1. AWS CLI aws fis list-actions — definitive, real-time list of FIS actions from user's region
2. aws___search_documentation — FIS actions reference page as fallback when CLI is unavailable

Documentation Research (Group C):

• aws___search_documentation — search AWS official docs
• aws___read_documentation — read full doc pages
• aws___recommend — discover related pages

All documentation research uses only the AWS Knowledge MCP tools above. Do NOT use SearXNG or other web search tools for documentation research.

Workflow

CRITICAL — Sequential execution of all AWS Knowledge MCP calls: All calls to aws___search_documentation, aws___read_documentation, and aws___recommend MUST be executed one at a time, sequentially. NEVER send multiple MCP requests in parallel — the aws-knowledge-mcp-server has strict rate limits and will reject concurrent requests with "Too many requests" errors. Wait for each request to return a complete response before sending the next one. This applies to ALL steps below (Step 2, 4b, 4c, 5a, 5b).

Retry on failure: If any MCP call (especially aws___read_documentation) returns a rate limit error ("Too many requests") or any other transient error, retry up to 10 times with a 5-second wait between retries. Only skip the request after all 10 retries have failed.

Multi-service requests: When the user asks about multiple services (e.g., "EKS, RDS, MSK, and ElastiCache"), process them one service at a time. Complete all research steps (Steps 2-5) for one service before starting the next. Do NOT launch parallel research for multiple services — this will trigger rate limiting. The Scenario Library fetch (Step 2) only needs to run once since it covers all services; the per-service steps (3-5) must be repeated sequentially for each service.

Step 1: Identify Target Service

Extract the target AWS service from the user's message and determine the target region.

Region Detection

FIS actions can differ across AWS regions — some actions may be available in us-east-1 but not yet in ap-southeast-1. Always determine the target region first, because service keyword resolution depends on it.

Detection order (use the first one that applies):

1. User explicitly specifies — e.g., "us-west-2", "东京区域", "ap-northeast-1"
2. Infer from context — resource ARNs, previous conversation mentioning a region
3. Check AWS CLI default — run aws configure get region to get the configured default
4. Ask the user — if none of the above yields a region, ask: "Which AWS region are you targeting? FIS actions and scenarios may vary by region."

Store the resolved region as TARGET_REGION for use in subsequent steps.

Service Keyword Resolution

FIS action IDs follow the pattern aws:<service>:<action>. To map the user's input to the correct FIS service keyword, use dynamic discovery from the live FIS action list:

aws fis list-actions --region TARGET_REGION | jq '.actions[].id' | awk -F':' '{print $2}' | sort -u

This returns the definitive list of FIS-supported service keywords in that region (e.g., ebs, ec2, ecs, eks, elasticache, fis, network, rds, s3, ssm...). Match the user's service name against this list. For example, if the user says "Aurora", match it to rds; if "Kubernetes", match to eks.

If the AWS CLI is not available, derive the keyword by lowercasing the AWS service name and removing spaces/hyphens (e.g., "ElastiCache" -> elasticache).

If the service is ambiguous, ask the user to clarify (e.g., "RDS MySQL or Aurora MySQL?").

Also determine the deployment architecture if the user mentions it:

• Multi-AZ, Multi-Region, Single-AZ
• Read replicas, Global Tables, Cross-region replication
• This affects which scenarios are relevant

Step 2: Fetch FIS Scenario Library (Scenario-Library-First)

This step has the highest priority. The FIS Scenario Library provides AWS-curated composite scenarios that orchestrate multiple fault injection actions into realistic failure simulations. These are the most valuable starting point because they represent AWS's own recommendations for how to test resilience.

Scenario Library scenarios are console-only — they cannot be listed or queried via AWS CLI or API. The only way to discover them is by reading the latest documentation.

Fetch the Scenario Library pages listed in references/search-queries.md under "FIS Scenario Library Pages (Always Fetch)". Read both the overview and detailed scenario pages relevant to the target service. Read pages one at a time, sequentially — wait for each aws___read_documentation call to complete before starting the next one.

From the scenario documentation, extract for each relevant scenario:

• Scenario name and description
• Which sub-actions the scenario orchestrates
• Which sub-actions are relevant to the target service
• What resource tags are required to target specific resources
• The default durations (interruption + recovery phases)
• Any prerequisites or limitations
• Stop condition recommendations

Decision: Which scenarios apply?

After reading the documentation, classify each scenario's relevance:

Include both directly and indirectly relevant scenarios in the output.

Step 3: Query FIS Actions

After the Scenario Library research, query individual FIS actions to discover service-specific fault injection capabilities that may not be covered by composite scenarios.

Path A: AWS CLI Available (Preferred)

Step 3a: Fetch ALL FIS actions in the target region:

aws fis list-actions --region TARGET_REGION --query 'actions[].{id:id, description:description}' --output json

Replace TARGET_REGION with the region resolved in Step 1 (e.g., us-east-1). If no region was determined, omit --region to use the CLI default, but warn the user that results reflect their default region and may differ in other regions.

Step 3b: Filter for target service — from the full list, find actions whose id contains the search keyword(s) from Step 1:

aws fis list-actions --region TARGET_REGION --query 'actions[?starts_with(id, `aws:KEYWORD:`)].{id:id, description:description}' --output json

Also scan the description field for the service name, because some actions may reference a service in their description even if the action prefix is different.

Step 3c (Optional): Collect cross-cutting actions — these affect services indirectly. Include them if the user's service would benefit from network, API, or infrastructure-level fault injection testing:

aws fis list-actions --region TARGET_REGION --query 'actions[?starts_with(id, `aws:network:`) || starts_with(id, `aws:fis:inject`) || starts_with(id, `aws:ssm:`) || starts_with(id, `aws:ec2:stop`) || starts_with(id, `aws:ec2:terminate`)].{id:id, description:description}' --output json

Cross-cutting actions and when they're useful:

• aws:network:disrupt-connectivity — useful for any VPC-based service
• aws:network:disrupt-vpc-endpoint — useful for services accessed via PrivateLink
• aws:fis:inject-api-internal-error — useful to test app handling of AWS API failures
• aws:fis:inject-api-throttle-error — useful to test backoff/retry logic
• aws:fis:inject-api-unavailable-error — useful to test graceful degradation
• aws:ec2:stop-instances / terminate-instances — useful for services running on EC2
• aws:ssm:send-command / start-automation-execution — useful for custom fault scripts

Whether to include cross-cutting actions depends on context:

• Include when the service runs on EC2, uses VPC networking, or the user is interested in infrastructure-level failure testing
• Skip when the user is focused only on service-native failures, or the service is fully managed with no user-accessible infrastructure layer

Path B: AWS CLI Not Available

Search the FIS actions reference documentation:

aws___search_documentation(
  search_phrase="AWS FIS actions [SERVICE_NAME] fault injection",
  topics=["reference_documentation"],
  limit=10
)

Then read the FIS actions reference page:

aws___read_documentation(
  url="https://docs.aws.amazon.com/fis/latest/userguide/fis-actions-reference.html",
  max_length=10000
)

Decision Point: FIS Actions Found?

Count the number of service-specific actions found (exclude cross-cutting actions).

• YES (1+ service-specific actions found) -> Continue to Step 4 (FIS-Enriched Path)
• NO (zero service-specific actions) -> Jump to Step 5 (Documentation-Only Path)

Step 4: FIS-Enriched Path

When FIS has native actions for the target service, combine Scenario Library findings with FIS-action-specific details.

4a: Organize FIS Actions into Testing Scenarios

Map each FIS action to a testing scenario. Use the "FIS Native Fault Injection Scenarios" table format from references/output-template.md.

IMPORTANT — Scenario Library deduplication (must apply before building the table): Before listing any FIS action in the per-service table, check whether that exact action ID appeared as a sub-action in any Scenario Library composite scenario discovered in Step 2. Common examples of overlap:

• aws:rds:failover-db-cluster — sub-action of AZ Power Interruption
• aws:elasticache:replicationgroup-interrupt-az-power — sub-action of AZ Power Interruption
• aws:eks:pod-network-latency — sub-action of AZ Application Slowdown
• aws:eks:pod-network-packet-loss — sub-action of Cross-AZ Traffic Slowdown
• aws:ec2:stop-instances — sub-action of AZ Power Interruption

Rules:

1. If an action is a Scenario Library sub-action, still list it in the per-service table but append to the "HA Verification Purpose" column: "(Also sub-action of {Scenario Name} — see Scenario Library section)".
2. If all service-specific FIS actions are Scenario Library sub-actions (e.g., ElastiCache has only replicationgroup-interrupt-az-power which is covered by AZ Power Interruption), omit the "FIS Native Fault Injection Scenarios" sub-section entirely and replace with: All FIS native actions for {SERVICE} are covered by Scenario Library composite scenarios. See the Scenario Library and Cross-Cutting section for details.

Group scenarios by failure domain:

1. Instance/Task Level — individual resource failure
2. Storage Level — disk/volume failure or degradation
3. Network Level — connectivity disruption
4. AZ Level — availability zone failure simulation
5. Region Level — cross-region failover
6. API/Control Plane — AWS API errors

Scenario Library cross-reference: For each FIS action, check whether it also appears as a sub-action in any Scenario Library composite scenario discovered in Step 2. If it does, append a note in the "HA Verification Purpose" column (e.g., "Also a sub-action of AZ Power Interruption — see Scenario Library section"). If all service-specific FIS actions are sub-actions of Scenario Library scenarios, omit the "FIS Native Fault Injection Scenarios" sub-section entirely and replace it with a note: "All FIS native actions for this service are covered by Scenario Library composite scenarios — see the Scenario Library and Cross-Cutting section."

4b: Enrich with Service-Specific Capabilities

Some services have built-in fault injection beyond FIS. Search for these (sequentially — wait for the search to complete before reading any result pages):

aws___search_documentation(
  search_phrase="[SERVICE_NAME] fault injection testing failover simulation",
  topics=["general", "reference_documentation"],
  limit=10
)

If found, add a "Service Built-in Fault Injection" section using the table format from references/output-template.md.

4c: Deep Documentation Research

Use the search queries from references/search-queries.md under "FIS-Enriched Path". Run all 5 queries sequentially (one at a time). After searches, read the top 3-5 most relevant pages one at a time and use aws___recommend on the most relevant page for discovery. Never send multiple read or recommend requests in parallel.

Step 5: Documentation-Only Path (No FIS Actions)

When FIS has no native actions for the target service, fall back to comprehensive documentation research. Note that Scenario Library findings from Step 2 still apply.

5a: Deep Documentation Search

Use the search queries from references/search-queries.md under "Documentation-Only Path". Run all 6 queries sequentially (one at a time, wait for each to complete).

5b: Read Key Pages and Discover Related Content

From the combined search results, read the top 5 most relevant pages following the priority order in references/search-queries.md. Read pages one at a time — wait for each aws___read_documentation call to complete before the next. Then use aws___recommend on the service's main documentation page to discover related content.

Extract from all pages:

• Failure modes the service can experience
• Built-in HA mechanisms (automatic failover, replication, etc.)
• Testing approaches documented in official guides
• Monitoring/metrics to watch during tests

5c: Compile Alternative Testing Approaches

Use the "Testing Methods (No Native FIS Actions)" section format from references/output-template.md, including both indirect FIS actions and AWS API/Console methods.

Step 6: Compile Output and Save to Local File

Write the report directly to a local markdown file instead of outputting the full content to the terminal. Use the following file naming convention:

TIMESTAMP=$(TZ=Asia/Shanghai date +%Y-%m-%d-%H-%M-%S)
SERVICE_SLUG=$(echo "{SERVICE_NAME}" | tr '[:upper:]' '[:lower:]' | tr ' :/' '-')
# File name: ${TIMESTAMP}-${SERVICE_SLUG}-chaos-research.md

For multi-service requests, generate one file per service:

• ${TIMESTAMP}-rds-chaos-research.md
• ${TIMESTAMP}-eks-chaos-research.md
• etc.

Compile the report content using the exact format defined in references/output-template.md and save it to the file. The report must include all sections in this order:

1. Executive Summary — overview with region, FIS support status, key recommendation
2. Scenario Library and Cross-Cutting — Scenario Library composite scenarios (highest priority), cross-cutting actions as optional supplement. This section comes BEFORE per-service sections.
3. Per-service sections — each with: FIS scenarios (using {SVC}-# test IDs, e.g., EKS-1, Redis-1), built-in methods, recommended testing scenario matrix, environment observations, and stop conditions
4. Recommended Test Priority (Consolidated) — references test IDs from per-service sections; do NOT duplicate full descriptions; do NOT list a FIS action separately if already covered by a Scenario Library scenario in the same table
5. Implementation Best Practices — steady state, DNS/connection, blast radius
6. Reference Materials — only URLs from actual search results or pages read
7. Next Steps — 3-4 actionable next steps

After saving, print a brief summary to the terminal listing only:

• The file path(s) of the generated report(s)
• Target service(s) and region
• Number of FIS actions found (service-specific + cross-cutting)
• Number of Scenario Library scenarios identified
• Top 3 recommended test priorities

Important Guidelines

• Scenario Library first, always. The FIS Scenario Library represents AWS's own curated resilience testing scenarios. Always read the latest Scenario Library documentation before anything else. These are documentation-based (console-only), not CLI-queryable.
• Region matters. Always resolve the target region before querying FIS actions. FIS action availability varies by region. Always pass --region to the AWS CLI and clearly state the region in the output.
• Don't fabricate FIS actions. If an action doesn't exist, say so clearly. The fallback path exists precisely for services FIS doesn't cover.
• Don't fabricate links. Only include URLs from actual search results or known documentation pages you've read.
• Be specific about the service. Every recommendation should reference the specific service, its HA mechanisms, and its specific metrics.
• Cross-cutting actions are optional context. Include them when they add value, but focus on service-specific actions and Scenario Library scenarios first.
• AWS Knowledge MCP only for docs research. Do NOT use SearXNG or other web search. Use aws___search_documentation, aws___read_documentation, and aws___recommend.
• Search across multiple topics. Use different topics values (general, reference_documentation, troubleshooting) sequentially.
• Use aws___recommend for discovery. After reading a key page, call aws___recommend to find related content that keyword search may miss.
• NEVER send MCP requests in parallel. All calls to aws___search_documentation, aws___read_documentation, and aws___recommend MUST be executed one at a time. Wait for each response before sending the next request. Parallel calls will trigger "Too many requests" errors from the aws-knowledge-mcp-server. This is the single most common cause of failures — enforce strictly in every step.
• Retry on failure — up to 10 times. If any MCP call fails with a rate limit or transient error, wait 5 seconds and retry. Repeat up to 10 times before skipping.
• Respect language. Output in the same language as the user's conversation.