# spg-storage perf budgets
Per stone-of-truth (see workspace `PERFORMANCE.md`): ~10× headroom over
the criterion bench median; gate catches order-of-magnitude regressions
only.
| `catalog_roundtrip_100rows` (≤) | 5 ms | **5.3 µs** ser+deser (v3.0.0) | ~940× |
| `hnsw_search_top10_dim8_n200` (≤) | 1 ms | **397 ns** (v3.0.1) | ~2500× |
Run: `cargo test -p spg-storage --test perf_gate`.
The catalog roundtrip exercises every disk format path on a small
100-row, 3-col table; on M-series the real number sits in the
hundreds-of-microseconds band. HNSW top-10 search across a built
200-node index is sub-millisecond on M-series; the 1-ms budget trips
on real regressions (e.g. accidentally rebuilding the graph per
search, or an `O(n)` per-neighbour reallocation).