performance-engineering

# Performance Engineering Note: This skill is Node.js-focused. The mental model (V8 internals, GC tuning, event loop forensics) applies broadly, but specific commands and profilers differ on Python/Go/Rust runtimes. ## Fail-Fast Rule If a performance metric exceeds its budget, **stop and investigate**. Never accept "it's fast enough" without numbers. Never optimize without profiling first. Never profile with small data sets and claim production readiness. ## {{PROJECT_NAME}} Performance Budgets | Metric | Budget | Current | Alert At | |--------|--------|---------|----------| | Event loop p50 | < 5ms | 1.26ms | > 10ms | | Event loop p99 | < 100ms | ~100ms | > 200ms | | Heap used | < 2GB | ~604MB | > 1.5GB | | RSS (main) | < 4GB | ~2-4GB | > 6GB | | RSS (adapter) | < 24GB | ~16-24GB | > 28GB | | IPC latency | < 50ms | ~50ms coalesce | > 100ms | | Event throughput | > 200 events/s | ~400+ events/s | < 150 events/s | ## V8 Internals That Matter ### Hidden Classes (Shapes) - Objects with same property order share hidden classes → fast property access - Adding properties in different orders creates new hidden classes → SLOW - **Pattern**: Hot-path state objects must always have the same property shape - In a `storeFastRecord()`-style function: in-place mutation preserves hidden class (good) - Creating new objects per event would thrash hidden classes (bad) ### Inline Caches (ICs) - V8 caches property lookup locations per call site - Polymorphic ICs (>4 shapes at one call si...