Benchmarking

ERPBridge includes deterministic Go microbenchmarks for cache and request-preparation paths. They run in process and do not require Redis, MockERP, Docker, credentials, or a live MCP connection.

Local performance baseline

Measure the code paths that run for every cached tool call, then compare the same workload on the same machine.

Measures

Latency, memory, and allocations

Excludes

ERP, Redis, plugins, and HTTP

Use

Local regression comparison

Run microbenchmarks

Core suite

Run all microbenchmarks

Runs only benchmark functions and reports latency and allocation metrics.

make bench

Compare locally

Collect five samples

Keep the machine otherwise idle before comparing a change.

go test -run '^$' -bench . -benchmem -count=5 ./internal/cache ./internal/mcp

The commands use -run '^$', so they do not run ordinary tests. They report ns/op, B/op, and allocs/op.

Local baseline

The following five-sample medians were measured on 2026-08-31 with Go 1.26.2 on Linux/amd64, using an 11th Gen Intel Core i5-1135G7. They are a local baseline only.

Empty argument hash2.5 ns0 allocations

Memory cache hit8.5 us22 allocations

Full cache middleware hit17.3 us58 allocations

On small screens, scroll this table horizontally to view all result columns.

Benchmark	Median	Allocation
ArgsHash/empty	2.5 ns/op	0 B/op, 0 allocs/op
ArgsHash/small	2348 ns/op	384 B/op, 7 allocs/op
ArgsHash/nested	4352 ns/op	656 B/op, 12 allocs/op
ExactKey	2910 ns/op	528 B/op, 11 allocs/op
MemoryManager/hit	8500 ns/op	1082 B/op, 22 allocs/op
MemoryManager/set	5548 ns/op	977 B/op, 18 allocs/op
CacheMiddleware/hit	17287 ns/op	2662 B/op, 58 allocs/op
CacheMiddleware/disabled	236 ns/op	144 B/op, 3 allocs/op
PrepareERPCall/legacy_get_query	2193 ns/op	613 B/op, 11 allocs/op
PrepareERPCall/generated_locations	5211 ns/op	1625 B/op, 24 allocs/op
PrepareERPCall/complete_body	1329 ns/op	240 B/op, 6 allocs/op

Concurrent cache-hit stress test

The stress benchmark shares one pre-seeded in-memory cache entry across parallel callers. It exercises cache-key creation, the cache middleware, JSON decoding, the lifecycle read lock, and the memory LRU lock. It does not call Redis, an ERP, plugins, or an HTTP server.

Shared-cache contention

Stress concurrent in-memory cache hits

Run the cache-hit path for five seconds at one and eight CPUs.

make stress

This command runs CacheMiddleware/hit_parallel for five seconds at one and eight CPUs. On the same local environment, five-sample medians were:

1 CPU77,206 ops/s12,952 ns/op · 2,656 B/op

8 CPUs176,903 ops/s5,653 ns/op · 2,664 B/op

The eight-CPU samples varied from 171757 to 255476 ops/s because callers contend for the shared in-memory LRU lock. Use the median for local regression comparison. Do not use these values to size a production deployment.