Strangler Fig Planner

Strangler Fig Planner

Plan a route-by-route replacement of a live system: install a routing seam, migrate one vertical capability at a time, and delete the legacy trunk only after each slice has soaked at full traffic.

Workflow

1. Install the routing seam first. Find or insert a place to redirect traffic per slice: a reverse proxy (NGINX, Envoy), an API gateway, a facade service, or a feature-flag check in the legacy entry point. The seam must route at the granularity you intend to migrate - per endpoint, per URL prefix, per message type. Verify you can flip one route to the new system and back without a deploy, in under a minute. If you cannot route incrementally, you cannot strangle incrementally; building the seam is step zero, before any slice.
2. Slice by business capability, not by layer. Cut vertically - a whole capability (checkout, user profile, billing export) end to end, never 'the data layer' or 'all the controllers'. Pick the first slice for low risk and high learning: modest traffic (roughly 1-5% of total volume is a good first target), few upstream dependencies, clear ownership, observable behavior. Do not start with the crown-jewel transaction. Size each slice to ship in days to a few weeks; if an estimate exceeds 6 weeks, cut the slice smaller.
3. Sequence by dependency and data gravity. Map which slices read and write shared data. Migrate leaf capabilities before the tables everyone touches. Where new and old share a database, decide per slice: dual-write with reconciliation, change-data-capture, or new-system-owns-the-table with a read replica for legacy. Order steps so each one leaves the system fully working - never a state that only resolves after three more PRs.
4. De-risk every cutover. Per slice: shadow traffic (send to both, compare outputs, serve old) to validate - hold shadowing until the output mismatch rate is below ~0.1% or every mismatch is explained - then canary a small percent and ramp on a schedule like 1% → 5% → 25% → 100%, watching the comparison metric at each step. Define rollback as flipping the route, not reverting code. Add production metrics that compare old vs new outputs. Keep the kill switch until the slice has run at 100 percent through a full business cycle (month-end, peak) - typically 30 days minimum.
5. Define done and delete the trunk. A slice is done only when the legacy path has served zero traffic for a defined soak period and the dead legacy code is deleted, not left dormant. Track remaining legacy surface explicitly so the project ends rather than stalling at 80 percent forever.