Lambda Recursive Loop Detection: Break Accidental Invoke Storms Early

One miswired SNS → Lambda → SNS, or a “helpful” handler that re-invokes itself on every error, can exhaust regional concurrency and print a five-figure bill before anyone wakes up. AWS recursive loop detection is the seatbelt; your architecture and billing alarms are the brakes.

⚡ TL;DR: Prefer event designs that cannot fan into themselves; enable recursive loop detection; stamp idempotency + hop counters on every async invoke; alarm on RecursiveInvocationsDropped and unexpected spend. Pair with Lambda Destinations and Lambda Timeouts & DLQs.

How loops form

Classic patterns:

1. SQS → Lambda → same SQS (publish on failure without poison handling)
2. SNS topic A → Lambda → publish to topic A
3. Async self-invoke “retry” without a counter
4. S3 event on bucket B → Lambda writes to bucket B under a watched prefix

// ❌ Accidental loop: error path republishes same message shape to same bus
await eventBridge.putEvents({
  Entries: [{
    EventBusName: "app",
    Source: "orders",
    DetailType: "OrderFailed",
    Detail: JSON.stringify(event.detail), // re-enters the same rule
  }],
});

// ✅ Hop counter + different detail type for remediation
const hops = Number(event.detail?.hops ?? 0);
if (hops >= 3) {
  await sendToDestinations(event); // on-failure queue, not the bus
  return;
}
await eventBridge.putEvents({
  Entries: [{
    EventBusName: "app",
    Source: "orders.remediator",
    DetailType: "OrderRemediationRequested",
    Detail: JSON.stringify({ ...event.detail, hops: hops + 1 }),
  }],
});

Turn on detection and watch the metric

Recursive loop detection drops invokes that look like cycles among supported AWS services and emits RecursiveInvocationsDropped. Treat that metric as a sev page, not a curiosity.

# Alarm sketch
aws cloudwatch put-metric-alarm \
  --alarm-name lambda-recursive-dropped \
  --metric-name RecursiveInvocationsDropped \
  --namespace AWS/Lambda \
  --statistic Sum \
  --period 60 \
  --evaluation-periods 1 \
  --threshold 1 \
  --comparison-operator GreaterThanOrEqualToThreshold \
  --dimensions Name=FunctionName,Value=orders-consumer \
  --alarm-actions arn:aws:sns:region:acct:oncall

Also alarm on:

• UnreservedConcurrentExecution approaching account limit
• Cost anomaly / daily Lambda spend budget
• SQS NumberOfMessagesSent self-correlation with the consumer function

Defense in depth beyond AWS detection

# assert_no_self_invoke.py (CI check on IaC)
import json, sys
template = json.load(open(sys.argv[1]))
# naive: flag Lambda whose EventBridge rule target equals its own publish bus+source
# fail build if detected without explicit allow annotation

Closing checklist

✅ Dos
– ✅ Design acyclic event graphs; document allowed remediations
– ✅ Enable recursive loop detection on async workflows
– ✅ Page on RecursiveInvocationsDropped
– ✅ Cap reserved concurrency on risky consumers
– ✅ Use destinations/DLQ instead of republishing the same event

❌ Don’ts
– ❌ Don’t async self-invoke without a hop budget
– ❌ Don’t write S3 notifications back to the watched prefix
– ❌ Don’t disable detection to “fix retries”
– ❌ Don’t rely on humans noticing the bill first

Related reading

• Lambda Destinations: Async Failure Routing
• Lambda Timeouts, Retries, and DLQs
• AWS Lambda Best Practices
• Lambda Event Source Mapping Parallelization

Last updated on September 11, 2026