aws-cloudwatch-investigation

AWS CloudWatch Investigation Skill

Reusable patterns for investigating production incidents using CloudWatch Logs, Metrics, and Alarms. These patterns are designed to be composed together during incident triage.

Required IAM Permissions

The agent executing this skill requires only the following read-only IAM actions. Construct the execution role with exactly these permissions and no others:

No write permissions are required or permitted. The health:DescribeEvents action requires the IAM principal to be in us-east-1; scope the role's region condition accordingly if your policy enforces regional restrictions.

Security note for programmatic consumers: If these query templates are consumed by an automated tool or agent that performs string substitution, all placeholder values (e.g., <event-data-store-id>, <alarm_time>, <alarm_time_minus_30m>, TARGET) must be validated and sanitized before substitution. Accepted values should be constrained to their expected types (ISO-8601 timestamps, AWS resource-ID formats, alphanumeric names). Do not perform naive string interpolation on untrusted or externally supplied input.

Pattern 1: Logs Insights Query Templates

Error Spike Detection

Find the top errors in a time window, grouped by error type:

fields @timestamp, @message, @logStream
| filter @message like /(?i)(error|exception|fatal|critical)/
| stats count(*) as errorCount by bin(5m), @logStream
| sort errorCount desc
| limit 20

P99 Latency Breakdown by Operation

Identify which operations are driving latency spikes:

fields @timestamp, @duration, operation
| filter ispresent(@duration)
| stats avg(@duration) as avgMs,
        pct(@duration, 50) as p50Ms,
        pct(@duration, 95) as p95Ms,
        pct(@duration, 99) as p99Ms,
        count(*) as invocations
  by operation
| sort p99Ms desc
| limit 15

Lambda Cold Start Detection

Quantify cold start impact during an incident:

fields @timestamp, @duration, @initDuration, @memorySize, @maxMemoryUsed
| filter ispresent(@initDuration)
| stats count(*) as coldStarts,
        avg(@initDuration) as avgInitMs,
        max(@initDuration) as maxInitMs,
        avg(@duration) as avgDurationMs
  by bin(5m)
| sort @timestamp desc

Out-of-Memory (OOM) Detection

Find Lambda functions or containers killed by memory pressure:

fields @timestamp, @message, @logStream, @memorySize, @maxMemoryUsed
| filter @message like /Runtime exited|out of memory|OOMKilled|Cannot allocate memory|MemoryError/
| stats count(*) as oomEvents by @logStream, bin(10m)
| sort oomEvents desc
| limit 10

For memory utilization trending before OOM:

fields @timestamp, @maxMemoryUsed, @memorySize
| filter ispresent(@maxMemoryUsed)
| stats max(@maxMemoryUsed / @memorySize * 100) as peakMemPct,
        avg(@maxMemoryUsed / @memorySize * 100) as avgMemPct
  by bin(5m)
| sort @timestamp desc

Timeout Detection

Find invocations that hit the configured timeout:

fields @timestamp, @duration, @logStream, @requestId
| filter @message like /Task timed out/ or @duration > 28000
| stats count(*) as timeouts by @logStream, bin(5m)
| sort timeouts desc

Pattern 2: Alarm History to Deploy-Event Correlation

Process

1. Get alarm transition time — note the exact timestamp when the alarm entered ALARM state.
2. Query CloudTrail for deployment-related events in a window of [alarm_time - 30min, alarm_time]:

# CloudTrail Lake query for deployment events
SELECT eventTime, eventName, userIdentity.arn, requestParameters
FROM <event-data-store-id>
WHERE eventTime > '<alarm_time_minus_30m>'
  AND eventTime < '<alarm_time>'
  AND eventName IN (
    'UpdateFunctionCode', 'UpdateFunctionConfiguration',
    'UpdateService', 'CreateDeployment', 'RegisterTaskDefinition',
    'CreateChangeSet', 'ExecuteChangeSet',
    'StartPipelineExecution', 'PutImage'
  )
ORDER BY eventTime DESC

1. Correlation criteria — a deploy is "correlated" if:

   • It targets the same service/resource as the alarm
   • It completed within 15 minutes before the alarm transition
   • The deployer identity matches a CI/CD role (not a human applying a hotfix)

2. Strengthening the correlation:

   • Check if the same alarm was healthy in the previous deployment cycle
   • Verify no other environmental changes (scaling events, config changes) in the same window
   • Look for canary/synthetic monitor failures that started at the same time

Output Format

Deploy Correlation:
  Event: UpdateFunctionCode
  Time: 2024-03-15T14:23:07Z (12 min before alarm)
  Actor: arn:aws:sts::123456789012:assumed-role/github-actions-deploy/session
  Resource: arn:aws:lambda:us-east-1:123456789012:function:payment-processor
  Correlation: STRONG — same resource, CI/CD actor, alarm was OK prior cycle

Pattern 3: Narrow the Blast Radius Decision Tree

Use this tree to systematically scope an incident from broadest to most specific:

START
  |
  v
[1] ACCOUNT — Which account(s) show the alarm?
  |  - Check: Are alarms firing in multiple accounts?
  |  - If yes → suspect shared service (SSO, networking, shared deployment pipeline)
  |  - If no → proceed to Region
  v
[2] REGION — Which region(s) are affected?
  |  - Check: Same alarm in other regions?
  |  - If multi-region → suspect global service (IAM, Route53, S3 global)
  |  - If single-region → proceed to Service
  v
[3] SERVICE — Which service namespace shows degradation?
  |  - Check CloudWatch namespace: AWS/Lambda, AWS/ECS, AWS/ApiGateway, etc.
  |  - If multiple services → suspect shared dependency (VPC, NAT, DNS, IAM)
  |  - If single service → proceed to Operation
  v
[4] OPERATION — Which API action or function is failing?
  |  - For Lambda: which function name?
  |  - For ECS: which service/task definition?
  |  - For API GW: which stage/resource/method?
  |  - If all operations → suspect service-level issue (throttling, quota)
  |  - If specific operation → proceed to Resource
  v
[5] RESOURCE — Which specific resource instance?
     - Function ARN, Task ID, DB instance identifier
     - This is your investigation target
     - Proceed to log and trace analysis scoped to this resource

Shared Dependency Investigation

When blast radius spans multiple services, investigate in this order:

1. VPC/Networking — NAT Gateway ErrorPortAllocation, packet drops, DNS resolution failures
2. IAM/STS — ThrottlingException on AssumeRole, token vending latency
3. Downstream dependency — shared database, cache, or external API
4. Deployment pipeline — simultaneous deploys across services from same pipeline run
5. AWS service event — check AWS Health Dashboard and Service Health for the region

Pattern 4: PromQL-Style Metric Query Patterns

These patterns use CloudWatch metric math and GetMetricData to build composite signals. Express them as metric queries for dashboards or programmatic retrieval.

Error Rate as Percentage

MetricDataQueries:
  - Id: errors
    MetricStat:
      Metric:
        Namespace: AWS/Lambda
        MetricName: Errors
        Dimensions: [{Name: FunctionName, Value: TARGET}]
      Period: 60
      Stat: Sum
  - Id: invocations
    MetricStat:
      Metric:
        Namespace: AWS/Lambda
        MetricName: Invocations
        Dimensions: [{Name: FunctionName, Value: TARGET}]
      Period: 60
      Stat: Sum
  - Id: error_rate
    Expression: "errors / invocations * 100"
    Label: "Error Rate %"

Latency Anomaly Detection (Compare to Baseline)

MetricDataQueries:
  - Id: current_p99
    MetricStat:
      Metric:
        Namespace: AWS/Lambda
        MetricName: Duration
        Dimensions: [{Name: FunctionName, Value: TARGET}]
      Period: 300
      Stat: p99
  - Id: baseline_p99
    MetricStat:
      Metric:
        Namespace: AWS/Lambda
        MetricName: Duration
        Dimensions: [{Name: FunctionName, Value: TARGET}]
      Period: 300
      Stat: p99
    # Use StartTime/EndTime set to same window last week
  - Id: anomaly_ratio
    Expression: "current_p99 / baseline_p99"
    Label: "Latency vs Baseline (ratio > 2 = anomaly)"

Throttling Pressure Score

Combine multiple throttling signals into a single pressure metric:

MetricDataQueries:
  - Id: lambda_throttles
    MetricStat:
      Metric: {Namespace: AWS/Lambda, MetricName: Throttles}
      Period: 60
      Stat: Sum
  - Id: api_gw_429s
    MetricStat:
      Metric: {Namespace: AWS/ApiGateway, MetricName: 4XXError, Dimensions: [{Name: ApiName, Value: TARGET}]}
      Period: 60
      Stat: Sum
  - Id: dynamo_throttles
    MetricStat:
      Metric: {Namespace: AWS/DynamoDB, MetricName: ThrottledRequests, Dimensions: [{Name: TableName, Value: TARGET}]}
      Period: 60
      Stat: Sum
  - Id: throttle_pressure
    Expression: "lambda_throttles + api_gw_429s + dynamo_throttles"
    Label: "Combined Throttle Pressure"

Concurrent Execution Headroom

MetricDataQueries:
  - Id: concurrent
    MetricStat:
      Metric: {Namespace: AWS/Lambda, MetricName: ConcurrentExecutions}
      Period: 60
      Stat: Maximum
  - Id: headroom
    Expression: "<ACCOUNT_CONCURRENCY_LIMIT> - concurrent"
    Label: "Remaining Concurrency"

Before using this query: substitute <ACCOUNT_CONCURRENCY_LIMIT> with the actual limit for the target account:

aws lambda get-account-settings \
  --query AccountLimit.TotalConcurrentExecutions \
  --output text

Pattern 5: Incident Timeline Reconstruction

Process

Reconstruct a precise timeline by merging data from multiple sources:

1. Collect timestamps:

1. Build the timeline:

fields @timestamp, @message
| filter @message like /ERROR|WARN|timeout|refused|denied/
| stats earliest(@timestamp) as firstSeen, latest(@timestamp) as lastSeen, count(*) as occurrences
  by @message
| sort firstSeen asc
| limit 20

1. Identify the sequence:

Timeline:
  T-15m: CloudTrail — UpdateFunctionCode by CI/CD role
  T-12m: Logs — first error "Connection refused to payments-api.internal"
  T-10m: Metrics — Error count crosses 5/min threshold
  T-8m:  Alarm — PaymentProcessorErrors enters ALARM
  T-5m:  Metrics — p99 latency spikes to 28s (timeout)
  T-0:   Current — error rate at 45%, alarm still firing

1. Determine root event — the earliest change that preceded all symptoms. Walk backward from the first symptom to the most recent mutation (deploy, config change, scaling event, or external dependency shift).

Gotchas

• CloudWatch metric timestamps are end-of-period. A 1-minute datapoint at 14:05 covers 14:04-14:05.
• CloudTrail events can have up to 15-minute delivery delay. Use eventTime, not ingestion time.
• Log group timestamps depend on the agent/SDK flush interval. Allow for 30-60s of clock skew.
• Alarm state changes have a built-in evaluation delay (periods x evaluation periods). The actual anomaly started earlier.