Migration Strategies: Moving Away from Legacy Without Burning Down the House

Every mature organization has legacy systems. The question is never whether to modernize but how to do it without halting business value delivery. The graveyard of failed "big bang" rewrites is large. Incremental strategies win.

Migration Pattern Comparison

Pattern	Approach	Risk	Duration	Team Size	Rollback
Strangler Fig	Replace piece by piece behind a facade	Low	Long (months-years)	Small-Medium	Per component
Branch by Abstraction	Abstract interfaces, swap implementations	Low-Medium	Medium	Medium	Per abstraction
Parallel Run	Run old and new simultaneously, compare	Medium	Medium	Large	Instant (keep old)
Big Bang Rewrite	Replace everything at once	Very High	Long	Large	All or nothing
Lift and Shift	Move as-is to new infra	Low	Short	Small	Easy
Feature Parity + Cutover	Build new until feature parity, then switch	High	Very Long	Large	Complex

Strangler Fig Pattern

Phase 1: Facade                  Phase 2: Migrate                Phase 3: Complete
┌──────────────┐                 ┌──────────────┐                ┌──────────────┐
│   Facade /   │                 │   Facade /   │                │   Facade /   │
│   API GW     │                 │   API GW     │                │   API GW     │
└──┬────────┬──┘                 └──┬────────┬──┘                └──────┬───────┘
   │        │                       │        │                          │
   │ 100%   │ 0%                    │ 40%    │ 60%                     │ 100%
   │        │                       │        │                          │
┌──▼──┐  ┌──▼──┐                 ┌──▼──┐  ┌──▼──┐                   ┌──▼──┐
│OLD  │  │NEW  │                 │OLD  │  │NEW  │                   │NEW  │
│     │  │     │                 │     │  │     │                   │     │
└─────┘  └─────┘                 └─────┘  └─────┘                   └─────┘

Risk Assessment Matrix

Risk Factor	Low (1)	Medium (2)	High (3)
Data complexity	Simple schemas, few relations	Moderate schemas, some coupling	Complex schemas, heavy cross-refs
Integration count	0-3 downstream systems	4-10 downstream systems	10+ downstream systems
Domain knowledge	Well-documented, team knows it	Partial docs, some tribal knowledge	No docs, original team gone
Business criticality	Internal tool	Revenue-supporting	Revenue-critical, SLA-bound
Test coverage	> 70%	30-70%	< 30% or unknown
Data volume	< 10GB	10GB-1TB	> 1TB
Regulatory constraints	None	Audit trail needed	Compliance certification required

Scoring: 7-10 = Strangler Fig or Lift-and-Shift, 11-16 = Branch by Abstraction or Parallel Run, 17-21 = start with assessment phase, consider external help.

Timeline Planning Framework

Phase	Duration	Activities	Gate Criteria
Discovery	2-4 weeks	Map dependencies, identify seams, assess risk	Risk matrix completed
Foundation	2-6 weeks	Build facade/proxy, set up dual-write, CI/CD	Routing works, zero traffic to new system
Migration Wave 1	4-8 weeks	Migrate lowest-risk component	Production traffic on new component
Migration Wave N	4-8 weeks each	Migrate next component by priority	Each wave validated in production
Decommission	2-4 weeks	Remove legacy code, clean up data	Old system fully offline
Stabilization	2-4 weeks	Monitor, fix edge cases, update docs	SLOs met for 2 consecutive weeks

The Feature Parity Trap

Effort ▲
       │                          ╱ "Just one more feature"
       │                        ╱   before we can switch
       │                      ╱
       │                    ╱        ← Trap zone: legacy keeps
       │                  ╱            evolving, target moves
       │                ╱
       │         ●────╱── Planned cutover (missed)
       │       ╱
       │     ╱
       │   ╱
       │ ╱
       └──────────────────────────────────────────► Time
         Start    6mo      12mo     18mo     24mo

  Solution: Define a FROZEN feature set. New features go
  to the new system only. Accept temporary gaps.

Key Principles

Migrate data last. Move routing and compute first. Data migration is the riskiest and most irreversible step. Delay it until everything else is proven.

Measure parity, do not feel it. Define quantitative criteria for when a component is ready: latency P99, error rate, throughput. "It seems to work" is not a migration gate.

Budget for the long tail. The last 10% of migration takes 50% of the effort. Plan for it. Edge cases, data inconsistencies, and undocumented behaviors live in that tail.