Serverless Application Modernization Playbook (2026): Strangler, Lambda, LMI, and Event Seams

Lambda Managed Instances (LMI) lets functions run on selected EC2 instance types with Lambda’s programming model — useful for steady, high-performance workloads that need specialized compute or EC2 pricing models (LMI launch coverage, ICYMI). Most modernization programs still start with standard Lambda and event seams, not a fleet of capacity providers.

This is the serverless modernization playbook. It is not general refactor/replatform/rearchitect, not scale-in-place before decomposition, not ECS Fargate zero-downtime migration, and not the LMI operations guide.

Artifacts: compute-path matrix, strangler checklist, architecture diagram (draw.io).

Benchmark silhouette (not a cited client) — .NET Framework monolith, ~220 endpoints, ~$28k/mo EC2+ALB, change lead time ~12 days. Strangled 14 high-churn endpoints to API Gateway + Lambda: those routes p95 1.4s → 210ms, deploy lead time same day. Premature LMI on a spiky reporting endpoint left capacity providers warm: ~$3.2k idle in a month — moved back to standard Lambda. Lesson: traffic shape picks compute, not fashion.

Strangler sequence

1. Map top endpoints by RPS × change frequency
2. Extract one capability behind API Gateway
3. Anti-corruption layer to shared data — no casual dual-write
4. Publish domain events to EventBridge
5. Retire monolith path within 30 days of 100% traffic

Opinionated take: Prefer standard Lambda for the first three extracts. Introduce LMI only when metrics show steady concurrency and EC2 economics win — then read the LMI ops guide.

Compute path

Use compute-path-decision-matrix.md:

Reference flow

Monolith (remaining)
   │
   ├── unchanged routes
   │
   └── strangled routes ──► API Gateway ──► Lambda or LMI
                                              │
                                              ├── Aurora / DynamoDB (owned tables)
                                              └── EventBridge ──► SQS consumers

What broke — LMI on spiky traffic

What broke — Reporting extract moved to LMI in week 2 because “we might need Graviton4.” Traffic was weekday 9–11 only. Capacity provider hours stayed warm. Detection: Cost Explorer + capacity provider metrics. Fix: return to standard Lambda; revisit LMI when p95 concurrency stays high >12 hours/day for two weeks.

What to Do This Week

1. Complete the strangler checklist Week 0 section.
2. Pick one endpoint; ship behind a flag.
3. Score Lambda vs LMI vs Fargate with the matrix — write the choice in the PR.
4. If the monolith is still fine on SLOs, read scale-in-place before a rewrite program.

What This Post Doesn’t Cover

• Full ECS/EKS platform builds
• LMI capacity provider operations (dedicated post)
• Mainframe or packaged ERP exits
• GenAI/agent rewrites of business logic