WebFlux Reactive Patterns

Use WebFlux when the complete request path benefits from non-blocking I/O.

Dependencies and pipeline rules

• Use the dedicated Boot 4 WebFlux and WebFlux test starters.
• Return Mono or Flux; never call subscribe() in request-handling code.
• Choose flatMap, concatMap, or sequential composition according to ordering needs.
• Keep JDBC, filesystem, and blocking SDK calls off Reactor event loops.
• Isolate unavoidable blocking work on boundedElastic at the adapter boundary.
• Apply timeouts to external calls and preserve cancellation.

Context and errors

• Put request-scoped metadata in Reactor Context, not ThreadLocal.
• Configure observability context propagation deliberately.
• Translate domain failures centrally without swallowing cancellation.
• Use onErrorResume only for a defined fallback and avoid onErrorContinue.

Persistence and streaming

• Use R2DBC for reactive database access.
• Bound fan-out concurrency according to downstream capacity.
• Define backpressure, buffering, and maximum in-memory sizes.
• Never collect an unbounded stream solely to simplify downstream code.

Testing

• Use StepVerifier for publisher behavior and WebTestClient for contracts.
• Test cancellation, timeout, empty results, errors, ordering, and backpressure.
• Detect blocking calls in tests when practical.

Examples

• See examples/good-reactive-service.java and examples/bad-reactive-service.java.

Official sources

• Spring WebFlux: https://docs.spring.io/spring-framework/reference/web/webflux.html
• Reactor reference: https://projectreactor.io/docs/core/release/reference/
• Spring Data R2DBC: https://docs.spring.io/spring-data/relational/reference/r2dbc.html

Gotchas

• Agent adds MVC and WebFlux starters accidentally - choose the intended web stack.
• Agent calls block() or subscribe() in application flow - keep subscription with the runtime.
• Agent uses ThreadLocal for tenant or trace data - use Reactor Context.
• Agent uses unbounded flatMap - set concurrency from downstream limits.
• Agent wraps JDBC with Mono.just - this still blocks the event loop.