# Serverless modernization compute-path decision matrix (2026)

| Workload shape | Standard Lambda | Lambda Managed Instances (LMI) | ECS Fargate / Express Mode | Stay on monolith VM |
|----------------|-----------------|--------------------------------|----------------------------|---------------------|
| Spiky, event-driven, <15 min | **2** | 1 | 1 | 0 |
| Steady high RPS, needs EC2 family / Graviton4 / multi-concurrent | 1 | **2** | **2** | 0 |
| Long-running workers / sidecars | 0 | 1 | **2** | 1 |
| Needs fleets of containers / service mesh | 0 | 0 | **2** | 1 |
| Unknown traffic, early strangler extract | **2** | 0 | 1 | 1 |

## Opinionated default

1. **Strangle** at HTTP or queue boundaries first — extract one capability behind API Gateway.
2. Default extract to **standard Lambda** + EventBridge/SQS.
3. Move to **LMI** when concurrency is steady and you need EC2 pricing (Savings Plans/RI) or specialized instance types — see LMI ops guide.
4. Use **ECS Express Mode / Fargate** when the unit of deploy is a container service, not a function.

## When NOT to

- Do not rewrite the whole monolith to Lambda in one program — extract seams.
- Do not put multi-minute CPU jobs on standard Lambda without chunking or Step Functions.
- Do not adopt LMI for cold, spiky workloads — you will overpay for capacity providers.