090a5ca69c* perf(search, metadata): add KG write buffer, score cache, adaptive budgets * perf(indexing): wire KGWriteBuffer into indexing service Converts per-document addEdgesUnique calls into batched flush via KGWriteBuffer. 500-doc ingest: 500 SQL round-trips → 1 batch. Verified: 14.2× edge insert gap (14.8K→195K edges/s) in kg_edge bench. Tests: 10/10 Catch2 pass (105 assertions). * perf(storage): add fsync durability, parallel batch, read integrity * test(storage): add stress, crash-sim, cold-cache, batch benchmarks * fix: CI optimizations and work coordinator shutdown fix * fix(testing): stabilize TSAN coverage gate * fix: harden KG flush, score cache, and benches - keep KG buffers intact when batch commit fails - count flushed edges per batch instead of lifetime totals - guard the Simeon score cache and clear it on rebuild - fall back to fsync when F_FULLFSYNC fails on macOS - remove interactive sudo from cold-cache benchmarks |
||
|---|---|---|
| .. | ||
| longmemeval_s_baseline.md | ||
| performance_report.md | ||
| README.md | ||
| storage_backends.md | ||
Benchmarks
This section is the public benchmark index for YAMS.
Each page answers three things:
- what workload or dataset was measured,
- what the latest published results are,
- what the historical comparison or regression signal is.
Benchmark Docs
- Performance Report - canonical local benchmark summary, live-mirror ingestion/retrieval refresh, and historical comparisons.
- Tree Builder - tree construction throughput.
- WriteCoordinator - ingest apply profile.
- LongMemEval_S Retrieval Quality Baseline - dataset statistics and retrieval-quality baselines.
- Storage Backends Benchmark - local vs R2 CLI CRUD and multi-client benchmark results.
Scope
- Public benchmark docs focus on benchmark data, workload context, and historical comparisons.
- Internal harness plumbing and generated outputs are intentionally omitted from public docs.
testedmeans the path is validated in this repository.supportedmeans the path exists but is not currently part of automated validation here.
Use tested paths as the default reference point.