Lambda Reserved Concurrency: Bulkheads That Protect Tenant Workloads

Account concurrency is a shared fate pool. One tenant’s report export or AI agent fan-out will starve checkout if you leave every function unreserved. Bulkheads via reserved concurrency are the unfair advantage: explicit partitions for critical paths, unreserved headroom for bursty best-effort work, and alarms before the pool hits zero.

⚡ TL;DR: Reserve concurrency for auth, checkout, and payment webhooks first. Cap noisy neighbors (exports, backfills, agent tools) with low reserved ceilings or separate accounts. Leave unreserved headroom for spikes — reserved is subtracted from the account pool. Simulate noisy-neighbor drills. Tie agent backends to Lambda Warm Pools and sandboxes in Secure AI Sandboxes.

Mental model: reserved subtracts from the pool

Account limit L (e.g. 1000)
Reserved sum R across functions
Unreserved capacity = L - R

Function with reserved=50 can never exceed 50
Function with no reservation competes for L - R

// GOOD: critical path reserved; batch capped
await lambda.putFunctionConcurrency({
  FunctionName: "checkout-api",
  ReservedConcurrentExecutions: 200,
});
await lambda.putFunctionConcurrency({
  FunctionName: "tenant-export",
  ReservedConcurrentExecutions: 20, // hard bulkhead
});

// BAD: reserve 900 of 1000 for "important" batch jobs
// leaves auth/checkout fighting over 100 unreserved

✅ Reserve the money path; cap the analytics path.
❌ Reserving almost everything “just in case” (destroys burst capacity).

Per-tenant isolation patterns

// Event source mapping max concurrency as a second bulkhead
await lambda.updateEventSourceMapping({
  UUID: esmId,
  ScalingConfig: { MaximumConcurrency: 10 },
});

For multi-tenant SaaS, do not invent “dynamic reserved concurrency per tenant” on every request — use queue depth, per-tenant rate limits, and tiered function aliases instead.

Drill the failure mode

# Synthetic load on export while measuring checkout Availability + Throttles
aws cloudwatch put-metric-alarm --alarm-name checkout-throttles \
  --metric-name Throttles --namespace AWS/Lambda \
  --dimensions Name=FunctionName,Value=checkout-api \
  --threshold 1 --comparison-operator GreaterThanOrEqualToThreshold \
  --evaluation-periods 1 --statistic Sum --period 60

If checkout throttles while export is healthy, your bulkhead math is inverted.

Closing checklist

• [ ] Money-path functions have reserved concurrency documented in an ADR
• [ ] Noisy jobs capped (reserved low or ESM max concurrency)
• [ ] Sum(reserved) leaves intentional unreserved headroom
• [ ] Alarms on Throttles for critical functions
• [ ] Quarterly noisy-neighbor game day
• [ ] Agent/tool Lambdas isolated from interactive APIs

Related reading

• Lambda Warm Pools: Low-Latency Backends for Coding Agent Tools
• Secure AI Sandboxes: Ephemeral ECS Tasks for Agent Tool Execution
• Agent Tool Allowlists: Least Privilege for Filesystem and Shell Access
• Cross-Account Bedrock Access: Platform Teams Without Shared Keys

Last updated on September 11, 2026