message-bus

v1

Benchmarks are written with PHPBench and live under benchmarks/. They are not part of mago's checked [source] paths, this is harness/tooling code, not library source, so it isn't held to the same lint/analyze conventions as src/.

Running the benchmarks

composer benchmark

This runs phpbench run --report=aggregate using phpbench.json.dist.

What's measured

  • MiddlewarePipeBench: dispatches a command through MessageBus -> MiddlewarePipe with a single pass-through middleware ahead of the terminal MessageHandlerMiddleware. Handler resolution is stubbed with a static resolver so the number isolates the cost of pipeline traversal itself (queue cloning/dequeue per Next invocation).

  • MessageBusDispatchBench: the total cost of handling a command end-to-end through a container-backed setup (Laminas ServiceManager), including real MessageHandlerResolver container lookups, matching how the bus is actually wired in an application, at 1 and 5 pass-through middleware.

Latest results

Recorded on 2026-08-06, PHP 8.4.24 (Linux x86_64, WSL2), 1000 revolutions / 5 iterations per subject.

Benchmark Subject Set Mode (μs) RStDev

MiddlewarePipeBench

benchHandle

1 middleware

3.363319

±3.81%

MessageBusDispatchBench

benchHandle

1 middleware

4.650117

±1.62%

MessageBusDispatchBench

benchHandle

5 middleware

6.909352

±1.93%

MessageBusDispatchBench rose by roughly 0.35-0.56μs from the previous baseline (4.300098μs / 6.345840μs) after MessageHandlerResolver::resolve() gained a command/query handler-type check (rejecting a resolved handler that doesn't implement the expected CommandHandlerInterface/ QueryHandlerInterface). MiddlewarePipeBench, which never touches the resolver, stayed flat, confirming the increase is attributable to that added validation rather than noise.

At 1 middleware, MessageBusDispatchBench (real PSR-11 container + config-driven wiring) costs about 1.3μs more than the equivalent MiddlewarePipeBench case, that's the overhead MessageHandlerResolver's container lookup and handler-type check add on top of pure pipeline traversal. Going from 1 to 5 middleware in MessageBusDispatchBench adds roughly 0.6μs per middleware, consistent with MiddlewarePipe cloning the internal SplQueue once per Next::handle() call.

Re-run composer benchmark and update this table after any change likely to affect pipeline or resolver performance.