First published in REST 4.8.0. Generated dispatch changes how eligible annotated handlers are called, not their Java API, request contract or native ABI.
| Processor option | Behavior |
|---|---|
Omitted or -Areactor.codegen.dispatch=auto |
Default. Use switch dispatch where supported; otherwise warn and keep the existing generated direct-call adapter. |
-Areactor.codegen.dispatch=switch |
Strict. Unsupported methods or table overflow fail compilation. |
-Areactor.codegen.dispatch=legacy |
Keep the previous generated route adapters and exception-handler lambdas. |
Other values fail compilation. Selection happens during annotation processing, not on each request. Changing a JVM property at startup does not change already compiled classes. A clean rebuild is required when changing modes.
Switch dispatch requires a registered component or @Bean owner: a public,
concrete, non-generic class with public, non-static, non-generic methods and
publicly accessible parameter types, including enclosing types and generic
arguments. Ordinary parameters need exactly one supported request binding.
In auto, package-private owners, methods, parameter types and exception handlers
keep their existing package-local generated factories. This includes
package-private @Bean return types. Registration, conditions and lifecycle stay
where they are; eligible public methods can still use switch in the same app.
Unregistered owners and previously invalid signatures do not gain new support.
If warnings are treated as errors, choose legacy explicitly where needed.
Routes support arities 0..8, with at most 256 switch entries per arity and a
separate limit of 256 exception-handler methods. IDs and ordering are deterministic.
auto keeps overflow entries on generated legacy calls; strict switch rejects
them. legacy does not remove the existing eight-parameter route limit. These
are generated-table limits, not connection or concurrency limits.
Each generated RouteN adapter contains its switch and direct handler call in
invokeN itself, not in a forwarding dispatcher. Specialized primitive
invokeInt, invokeLong, invokeBoolean, invokeDouble and invokeShort methods
also contain their own switches. Existing binding and direct-buffer/scalar paths
are unchanged, including @DirectQueryInt / @DirectPathInt runtime-direct
MethodHandle invocation. Generating an adapter does not reroute those calls.
Async handlers borrowing framework buffers remain unsupported.
The platform parent defaults reactor.codegen.dispatch to auto and forwards it
to the annotation processor. From your application's directory, choose a mode:
mvn -B -ntp -DskipTests -Dreactor.codegen.dispatch=auto clean package
mvn -B -ntp -DskipTests -Dreactor.codegen.dispatch=switch clean package
mvn -B -ntp -DskipTests -Dreactor.codegen.dispatch=legacy clean packageIf the application or a profile adds compiler arguments, append instead of
replacing the parent's options. For example, inside maven-compiler-plugin:
<configuration>
<compilerArgs combine.children="append">
<arg>-parameters</arg>
</compilerArgs>
</configuration>Apply the same merge rule to each profile that defines compilerArgs. Without
the platform parent, pass -Areactor.codegen.dispatch=... through your existing
compiler configuration; a Maven property alone is not an annotation-processor
option. The compiler's Generated dispatch mode=... note reports switch/legacy
route and exception-handler counts. Generated sources show the actual calls.
OpenJ9 runs the same precompiled Java adapters. The generated request path adds no reflective invocation, dynamic proxies or runtime recompilation. Existing startup inspection and runtime-direct MethodHandle paths are not removed. Guards, async completion, status/headers, response ownership and lifecycle keep their existing behavior. Java APIs and REST29 / Dubbo7 / SpringDubbo3 / Redis6 / Glowroot6 remain unchanged.
On 2026-10-27, source phase79-dc6e245883e69cefe9e6 passed the declared
EmaTech/OpenJ9 reader matrix: five fresh process pairs, six small/raw/heavy cells
at c64/c256, 19,816,245 exact-body validated warm/measured responses and 100
separate HTTP contract checks, with zero errors. The real cache-reader detail,
segment and generated health routes were measured; customer-number binding and
the expected missing-customer 404 were also checked. HTTP used pod loopback,
not public ingress. All five pairs, including the original failing cells, were
retained after adding the missing relative-order coverage.
Both arms used 1 CPU / 128 MiB, affinity 0,3, L1 disabled, seeded Redis without
a database, and 15-second warm / 8-second measurement windows per arm/cell.
Benchmark-only recipe: admission queue timeout was 125 ms for both
cache-read and cache-read-point, with unchanged concurrency caps 64/128.
Production defaults were not changed; this is not acceptance of the default
25 ms recipe. Acceptance used per-cell medians across all five paired results:
RPS >= 0.98, p99 <= 1.10, RSS/container delta <= +3 MiB and zero errors.
| Cell | RPS ratio | p99 ratio | RSS delta MiB | Container delta MiB |
|---|---|---|---|---|
| small c64 | 1.026040 | 1.027248 | -0.795 | -0.369 |
| small c256 | 1.036327 | 1.029094 | -0.801 | -0.428 |
| raw c64 | 0.998793 | 0.992910 | -0.598 | -0.223 |
| raw c256 | 0.991837 | 1.003699 | -0.703 | -0.277 |
| heavy c64 | 1.025231 | 0.991344 | -2.230 | -2.086 |
| heavy c256 | 0.989412 | 1.021405 | -0.488 | -0.910 |
A separate same-JVM HotSpot primitive-invoker allocation check measured 0 bytes for each mode over 100,000 calls after a positive 532,480-byte counter calibration. The OpenJ9 allocation-method abort is not counted as a pass; this does not prove that complete HTTP requests allocate nothing. Earlier FAIL receipts remain preserved. These bounded results do not establish universal speedup, consumer performance, public-network behavior, long-soak stability, other platforms or full production/release qualification. This measurement predates publication; shipping the feature in 4.8.0 does not turn it into full production or user acceptance.