performance

Performance

Measurable baselines for every hot path in worker-kmp. CI fails on >20% regression vs the latest release baseline. See perf-baselines/ for raw JMH JSON per release.

Metrics tracked (initial — v2.2.0)

Surface	Metric	Target (acceptable)	Regression threshold
TestWorkManager.enqueue(OneTimeWorkRequest) (no constraints)	wall-clock μs (avg)	<500μs	>20%
TestWorkManager.enqueue(OneTimeWorkRequest) (CONNECTED constraint)	wall-clock μs (avg)	<500μs	>20%
TestWorkManager in-memory state ops at N=100	wall-clock μs (avg)	<100μs	>20%

Additional metrics ship at later milestones as the relevant surfaces land:

• Observer chain overhead → v2.2.0 + alpha07 (Phase 4 WorkObserver SPI)
• Daemon cold-start + roundtrip → v3.0.0-alpha05 (Phase 8 Desktop daemon)
• Web Push wake → SW dispatch → v3.0.0-alpha06 (Phase 9 Web Push)
• Per-platform enqueue (Android / iOS / Web Wasm) → as each lands

How we benchmark

JMH (Java Microbenchmark Harness) 1.37 with:

• Mode: AverageTime + Throughput
• Forks: 3
• Warmup: 5 iterations × 2s
• Measurement: 5 iterations × 3s
• Output unit: microseconds (μs)

Runs on GitHub-hosted ubuntu-latest runners pinned to JDK temurin 21 + Gradle 9.5.1.

Reading the numbers

JMH reports score (mean) ± scoreError (confidence interval, 99.9%). Headline: score in scoreUnit. Lower is better for AverageTime; higher is better for Throughput.

Capturing a new baseline

bash scripts/perf-capture-baseline.sh <version>
git add perf-baselines/<version>.json
git commit -m "perf(bench): capture v<version> baseline"

Running benchmarks locally

./gradlew :cmp-worker-bench:jmh                           # full run (~15 min)
./gradlew :cmp-worker-bench:jmh -Pjmh.iterations=2 \
    -Pjmh.warmupIterations=2 -Pjmh.fork=1                  # quick run (~3 min)