nordisk/skade

skade — bench worklist (loop-until-times-down)

Prompt item: "skade — gå igenom källkod, hitta tidigare analyzer, skapa bencher, loopa tills du får ner tiderna." skade = the embedded single-file Iceberg engine (default warehouse impl). This is the triage of what's already benched vs the un-benched core hot paths + concrete zero-copy candidates. No heavy runs here — heavy benches run on the quiet bench box (Loki/Odin) per bench policy; bench data → /home.

What's ALREADY benched (the bench/ crate — nornir-Bencher harness, migrated OFF criterion)

• catalog throughput — catalog-bench embedded|nessie|polaris, RedbCatalog vs competitor backends (Nessie/Polaris/RustFS-S3/Iceberg-REST); emits nornir Bencher JSON.
• encode/write pipeline — pipeline.rs (zstd codec), OSM/TPC-H synthetic workloads (data.rs).
• tpch_warehouse (scalable) — first scalable-dimension bench (see .nornir/scalable-bench-design.md).
• containers — bench-containers bin brings competitor backends up/down in Rust (zero-shell).
• throughput — bench/src/skade_throughput.rs = the skade.* benchers (this is where new micro-benches belong).

Harness rule: NOT criterion (deliberately removed). New benches implement nornir::bench::api::Bencher (id()+run()) and emit BenchRun JSON, added to bench/src/skade_throughput.rs.

Un-benched CORE hot paths (skade src/) — GAPS to add

Zero-copy / algo candidates (the loop levers)

1. create/register key dedup — ✅ DONE. create_table and register_table each rebuilt table_key(name, ident) three times (redb insert · commit-log record · L1 write-through). Now built once per op and reused by reference: 3 allocs → 1 on the table-create/register commit path. Trivially byte-identical (same builder), covered by tests/catalog.rs (table_create_load_list_drop, register/rename). The optimistic commit hot path (persist_commit) was already single-key.
2. commit-key zero-copy — ✅ DONE (table_key) and EXTENDED to the whole builder family (namespace_key, namespace_prop_key, namespace_prop_prefix, table_prefix, catalog_prefix) via zero-alloc helpers ns_path_len/push_ns_path: 2 allocs → 1 each, byte-identical (guard ns_keys_are_byte_identical_to_legacy_format). Light bench: ~8–10× (see catalog-key-zerocopy.md). ✅ Harness bencher landed: bench-feature pub hook skade_katalog::bench_keys::* (gated, zero-cost in normal/published builds) lets the nornir-Bencher harness reach these pub(crate) fns — bencher skade_catalog_key_build in bench/examples/nornir-bench.rs (batched, black_box-pinned, BENCH_CATALOG_KEY_ITERS for the Loki/Odin sweep). The in-crate #[ignore] micro-test stays as the local sanity check. 1b. spatial radius O(k·n)→O(k) — ✅ DONE. GeoIndex::query_radius re-found every candidate's coords with entries.iter().find(|e| e.id == id) (an O(n) linear scan per candidate, so a k-hit radius was O(k·n)); query_bbox had already thrown those coords away. Extracted a shared for_bbox(min,max, |&Entry|) covering-cell walk that hands each matching Entry (with its lat/lon) straight to the caller — query_bbox collects e.id, query_radius computes haversine from the coords in hand. Byte-identical for unique ids (guard tests/spatial.rs::radius_matches_brute_force_after_zero_copy_rewrite compares id+bit-exact distance vs a brute-force scan across 4 query points). Light bencher landed: skade_spatial_radius in bench/examples/nornir-bench.rs (BENCH_SPATIAL_POINTS / BENCH_SPATIAL_QUERIES for the Loki/Odin sweep) + timing guard tests/spatial.rs::light_radius_timing_is_sublinear.
3. cache alloc audit — confirm meta_cache/table_cache/pointer_cache hits don't clone the value; hand back Arc/& on the hit path.
4. snapshot-append batching — coalesce commit-log writes; measure single vs batched append.
5. parquet window read — ParquetWindows streaming reader is already all-core (gatling); bench a zero-copy/mmap window path.

Next action (concrete)

Add skade.catalog_key + skade.cache_lookup (pure in-proc, no containers/S3) to bench/src/skade_throughput.rs, get a light local baseline, then run the heavy sweep on the bench box and loop the zero-copy commit-key change until the number drops (keep only if faster and byte-identical output).

Not added here as runnable code: a new criterion bench would reintroduce the removed pattern, and the catalog hot paths need the async+iceberg+redb setup that already lives in the bench/ crate — so the micro-benches belong in skade_throughput.rs, specced above, not a throwaway example.