Migrating off a legacy ESB without losing what it did

Enterprise service buses are unfashionable, and largely for good reasons. Centralised, heavily customised buses became bottlenecks: a change to one consumer required a bus release, which required a regression pack covering every other consumer.

What tends to get lost is that the bus was doing real work. Mediation, transformation, protocol bridging, guaranteed delivery, retry and dead-letter handling are all still required. Teams that decommission the bus without replacing those capabilities rediscover each one individually, usually during an incident.

A migration sequence that works

Step one: inventory by observation

Instrument the bus and record every flow that actually executes over a full business cycle. The deployed flow count is invariably higher than the executing count — dormant flows are the cheapest ones to migrate, because they simply get deleted.

Step two: classify by interaction pattern

• Request-response where the caller genuinely needs an answer now: these become synchronous APIs behind a gateway.
• Fire-and-forget notifications: these become events on the streaming backbone.
• Batch file transfers: these usually stay batch, but move to object storage with event notification.
• Orchestrated multi-step processes: these need an explicit workflow engine, not a bus flow with hidden state.

Step three: build the backbone before migrating flows

Streaming platform, schema registry, gateway, contract testing in the pipeline and observability wired through. Migrating flows onto an incomplete platform means migrating them twice.

Step four: migrate by domain, retire as you go

Take one business domain at a time, migrate every flow in it, and switch off the corresponding bus flows in the same wave. Domains give clean boundaries and a demonstrable reduction in the legacy footprint after each increment.

What genuinely improves

The measurable change is not raw throughput, which the bus usually handled adequately. It is change independence: a consumer can now be modified without a coordinated release across every other consumer. In the programmes we have run, that alone typically halves integration lead time.