Skip to content

BENCHMARKS.md

Latest commit

 

History

History
85 lines (64 loc) · 3.67 KB

BENCHMARKS.md

File metadata and controls

85 lines (64 loc) · 3.67 KB
Raw

Benchmarks

Baseline numbers for the gstate engine hot path. Run with:

just bench
# or: go test -bench=. -benchmem -count=3 -benchtime=1s -run='^$' ./...

Environment
Go 1.26.2 darwin/arm64
CPU Apple M1 Pro
GOMAXPROCS 10

Results

Each benchmark sends a batch of events through the engine and waits for all transitions to complete. Numbers are per-iteration (one batch).

Engine Hot Path (100 transitions per iteration)

Benchmark ns/op B/op allocs/op Notes
SendTransition_TrulyNoObserver 64,000 8,060 220 Baseline: pure engine overhead, zero observers registered
SendTransition_TransitionOnlyObserver 86,000 54,920 632 1 transition observer, triggers lazy event context
SendTransition_ThreeTransitionObservers 95,000 55,976 652 3 transition observers, shows scaling with multiple observers
SendTransition_RecordingObserver 115,000 123,025 953 Mutex append + context copies + state recording
SendTransition_BaseObserver 87,000 54,936 632 Registers BaseObserver (unused/filtered) + 1 transition observer

Context Snapshot Cost (100 transitions per iteration)

Benchmark ns/op B/op allocs/op Notes
ContextSnapshot_ValueCopy 86,000 54,920 632 Plain struct copy
ContextSnapshot_Cloner 89,000 74,184 630 Cloner interface: +4% time, ~1.35× bytes

Hierarchical Transitions (50 transitions per iteration)

Benchmark ns/op B/op allocs/op Notes
depth=1 45,000 30,104 332 Flat — no LCA walk
depth=4 110,000 71,576 736 2.4× depth=1
depth=8 222,000 127,104 1,193 4.9× depth=1 — linear in depth

Parallel Region Entry (50 transitions per iteration)

Benchmark ns/op B/op allocs/op Notes
regions=2 90,000 59,992 608
regions=4 144,000 90,848 836
regions=8 247,000 153,593 1,266 ~linear in region count

Always-Transition Chain (20 rounds per iteration)

Benchmark ns/op B/op allocs/op Notes
steps=5 95,000 65,160 772 7 transitions/round (5 always + GO + RESET)
steps=25 295,000 225,161 2,772 27 transitions/round
steps=50 548,000 425,162 5,272 52 transitions/round

Lifecycle Operations

Benchmark ns/op B/op allocs/op Notes
InvokeStartCancel (20 rounds) 54,000 34,842 414 Goroutine spawn + cancel per round
Snapshot/simple 92 64 2 1 active state, 0 history
Snapshot/parallel_4regions 200 192 2 9 active states
Hydrate/simple 1,780 4,712 14 Includes goroutine start
Hydrate/parallel_4regions 2,540 5,168 17

Allocation Check

Benchmark ns/op B/op allocs/op Notes
SendTransition_ZeroAlloc 64,000 8,060 220 Same as TrulyNoObserver — per-batch (100 events), not per-event

Note: The ~220 allocs/iteration in the TrulyNoObserver baseline include actor Start/Stop overhead (goroutine, channels, maps). Per-event allocation cost is 0 allocs/event when no observers are registered (pure engine hot path). Note that the benchmarks consistently include the actor.Stop() WaitGroup drain as part of the timed execution loop, meaning that actor lifecycle teardown cost is factored into the reported ns/op. The TestSendTransition_ZeroAllocCheck test logs the per-run breakdown.