v0.2.2

Released: October 2026

v0.2.2 ships the vector search and AI engine (HNSW, FLAT, BM25 full-text, FT.HYBRID RRF/LINEAR fusion) alongside a round of memory and throughput optimizations: arena milestone-shrink, two-pass FLAT Int8 pre-filter, lock-free search pool dispatch, O(bytes) bitmap operations, zero-copy cursor reads, and O(output) complement filters. Zero clippy warnings across all crates.

Against a 12-node Redis Cluster, a single FyroDB node delivers 9.14M Pipeline-64 SET, 47.12M INCR, 46.69M Hot Key, 70.44M Pub/Sub msg/sec, 5.0K HNSW KNN-10 q/s (4.55×), and 42.6K text-term q/s (26.6×).

---

Highlights

• Vector Search (HNSW + FLAT): FT.CREATE, FT.SEARCH KNN, VECTOR_RANGE, tag/numeric/text pre-filters, QUANTIZATION INT8 | RABITQ | NONE, exact rerank from the stored hash field, persistence in RDB v3, incremental maintenance via a write hook.
• BM25 Full-Text: Unicode tokenizer, RediSearch stopwords, BM25 (k1 = 1.2, b = 0.75) ranking, FT.AGGREGATE GROUPBY/REDUCE with cursors.
• FT.HYBRID: Lexical and vector legs run in parallel on the pool and fuse with RRF (Cormack) or LINEAR weighting.
• HNSW Arena Milestone-Shrink: Arena buffers shrink at each power-of-two node count milestone (last_shrink: AtomicUsize), reducing cache-miss pressure on streaming HSET inserts as well as bulk builds.
• Thread-Local Scratch Zero-Alloc: Scratch gains q_int8, acorn_hop, and acorn_frontier — ACORN traversal and FLAT two-pass scans share pre-allocated working memory with zero hot-path allocation after warmup.
• Two-Pass FLAT Scan: Int8 pre-filter pass narrows candidates before exact f32 rerank; codes are held in thread-local scratch to avoid per-query allocation.
• Lock-Free Search Pool: Job dispatch uses crossbeam::SegQueue<Job> per priority tier; submission never takes a lock. Index builds run on the low-priority tier so queries stay responsive during background work.
• prepare_batch Burst HSET: N consecutive HSET writes on the same vector field take one write-lock instead of N, cutting ingestion lock overhead.
• cursor_read Zero-Clone: FT.CURSOR READ uses Vec::split_off instead of a full cursor-row clone, eliminating O(n²) work per page.
• complement_ids Sorted-Merge: Negated filter sets resolve as an O(output) sorted merge over the ascending id range instead of a 100 KB boolean array per query.
• ZSet No-Alloc Update: ZSetData::insert probes the member index before allocating a SmallStr, skipping all allocation on score-update paths.
• O(bytes) Bitmap Ops: BITCOUNT and BITPOS operate at the byte level using hardware count_ones() / leading_zeros() instead of per-bit loops.
• Zero Clippy Warnings: All 16 warning categories resolved across fyro-vector, customhash, rust-zmalloc, and src/.

---

Performance

Benchmark	v0.2.2	Redis Cluster (12 nodes)	Speedup
Pipelined SET	20.60M ops/sec	6.53M ops/sec	3.15×
Pipelined GET	29.52M ops/sec	11.62M ops/sec	2.54×
INCR (counters)	47.12M ops/sec	6.54M ops/sec	7.2×
LPUSH/RPOP (queue)	42.86M ops/sec	7.17M ops/sec	5.98×
Hot Key (contention)	46.69M ops/sec	1.61M ops/sec	29.0×
SETBIT/BITCOUNT	20.92M ops/sec	5.94M ops/sec	3.52×
PFADD/PFCOUNT	36.29M ops/sec	6.03M ops/sec	6.02×
Pub/Sub delivery	70.44M msg/sec	8.09M msg/sec	8.7×
Vector KNN-10 (FLAT)	1.0K q/s	104 q/s †	9.6×
Text term search	42.6K q/s	1.6K q/s †	26.6×
Text negation	21.7K q/s	361 q/s †	60.1×
Vector HNSW KNN-10	5.0K q/s	1.1K q/s †	4.55×
Vector HNSW M=32 KNN-10	2.9K q/s	1.1K q/s †	2.64×

† Redis Cluster does not distribute vector or graph operations across nodes — those benchmarks ran against a single Redis node.

See the full benchmark table in Introduction.