◆ Case 04 / 06— Legacy Upgrade · 2022 · 22mo engagement
ClientManufacturing / mid-market
RoleMigration Lead
StackJava · Spring · Postgres · Debezium
OutcomeZero-downtime cutover
Fig.01 — Inference topology (v2)
● LIVE
Go/TypeScript/Python/Rust/Kubernetes/Kafka/Postgres/Ray/Redis/gRPC/Terraform/AWS/GCP/DDD/Event-Sourcing/Temporal/OpenTelemetry/ClickHouse/pgvector/Go/TypeScript/Python/Rust/Kubernetes/Kafka/Postgres/Ray/Redis/gRPC/Terraform/AWS/GCP/DDD/Event-Sourcing/Temporal/OpenTelemetry/ClickHouse/pgvector/
The Challenge —
Sixteen years of accumulated business truth lived inside stored procedures no one was willing to touch. The vendor had quietly stopped shipping patches.
The requirement wasn't parity. It was a graceful exit.
The Approach —
Strangler-fig migration: CDC via Debezium mirrored live writes into the new schema, feature-by-feature. The old system kept shipping invoices while the new one learned to.
Cutover was a single DNS flip, performed on a Tuesday afternoon.
0min
Cutover Downtime
0%
Data Parity
0 modules
Ported In Flight
Outcome —
No downtime. No data loss. The legacy stack ran in warm standby for another six months, purely for auditor comfort.
Migrations like this are boring when they go right. Make them boring.
Next Case — 05/06
RAG Knowledge Search↗