Hollow

v1.3.6 — Real Benchmark Suite

Every metric comes from a real API call against the live server. No mocks. No simulated latency.

Scenarios

• heap_alloc_throughput — alloc/free ops/sec against working memory kernel
• message_bus_latency — send→receive round-trip p50/p95/p99 (ms)
• transaction_commit_latency — begin→stage→commit round-trip (ms)
• checkpoint_roundtrip — save→restore→verify round-trip (ms)
• consensus_vote_latency — propose→vote×N→resolved wall time (ms)
• rate_limit_precision — verify 429 fires at correct bucket depth
• audit_write_throughput — operations/sec captured in audit log
• task_latency_c1/c3 — Ollama task submit→done wall time (optional)

Features

• Regression detection with 15% threshold per metric (higher/lower direction-aware)
• Results stored as timestamped JSON (last 20 runs)
• 3 REST routes: POST /benchmarks/run, GET /benchmarks/results, GET /benchmarks/compare
• 3 MCP tools: benchmark_run, benchmark_results, benchmark_compare
• 9 integration tests: all pass