Migration Halfway House
Teams plan incremental migrations — move from monolith to microservices, from REST to GraphQL, from one database to another — one piece at a time. The plan sounds responsible: reduce risk by migrating gradually. But incremental migrations create a prolonged state where both old and new systems coexist. This halfway house is the worst of both worlds. Engineers must understand and maintain two systems. Data flows through compatibility layers that add latency and bugs. New features must work in both paradigms. The migration stalls at 60-70% completion because the remaining 30% is the hardest, least-rewarding work. The organization lives in the halfway house for years, paying the tax of dual systems while getting the benefits of neither.
What people believe
“Incremental migration reduces risk compared to big-bang migration.”
| Metric | Before | After | Delta |
|---|---|---|---|
| Migration completion rate | 100% planned | 60-70% actual | -35% |
| Migration timeline | 6-12 months estimated | 2-4 years actual | +300% |
| Feature velocity during migration | Baseline | -40-60% | -50% |
| Operational complexity | One system | Two systems + compatibility layer | +150% |
Don't If
- •You don't have a hard deadline and dedicated team for completing the migration
- •The migration doesn't have executive sponsorship that survives priority changes
If You Must
- 1.Set a hard completion deadline and cut scope rather than extending timeline
- 2.Dedicate a team to migration — don't split engineers between migration and features
- 3.Migrate the hardest components first, not last
- 4.Define 'done' clearly — what gets turned off and when
Alternatives
- Strangler fig pattern — Route new traffic to new system, let old system die naturally
- Big-bang with rollback — Migrate everything at once with a tested rollback plan
- Don't migrate — Sometimes the cost of migration exceeds the benefit — invest in the existing system
This analysis is wrong if:
- Incremental migrations consistently complete within 150% of their estimated timeline
- Dual-system operation during migration doesn't measurably reduce feature velocity
- Migration completion rates exceed 90% without dedicated migration teams
- 1.Martin Fowler: Strangler Fig Application
Pattern for incremental migration that avoids the halfway house problem
- 2.Charity Majors: The Migration Tax
Analysis of how migrations consume engineering capacity for years beyond estimates
- 3.Shopify: Modular Monolith Migration
Shopify's experience with a multi-year migration that required dedicated teams
- 4.Uber Engineering: Microservice Migration Lessons
Uber's documentation of the prolonged dual-system state during their migration
This is a mirror — it shows what's already true.
Want to surface the hidden consequences of your engineering decisions?