Semantic cache, DiskANN, SSI, adaptive buffer pool
The single-node open-source engine is now feature-complete as of v0.7.0. Semantic result caching, on-disk vector index persistence, exact historical vector search, Serializable Snapshot Isolation, RGABH adaptive buffer pool, and disk-resident DiskANN all ship in this release.
- FeatureSemantic result caching: CREATE SEMANTIC CACHE FOR TABLE ... SIMILARITY ... TTL ... serves repeated SEMANTIC_MATCH queries from cache, skipping the HNSW search. Cache config is durable; entries invalidated on row changes and model-version changes.
- FeatureOn-disk vector index persistence. Each table's HNSW state is persisted on flush and shutdown, and reconciled on open. Removes the re-embed-every-row-on-restart cost.
- FeatureExact historical vector search via SEMANTIC_SNAPSHOT. SELECT ... AT VERSION <v> CONSISTENCY 'SEMANTIC_SNAPSHOT' WHERE SEMANTIC_MATCH(...) runs semantic search restricted to rows visible at version v, joined exactly by key.
- FeatureSerializable Snapshot Isolation. An opt-in serializable isolation level with a conservative commit-time certifier that aborts write-skew anomalies with SQLSTATE 40001. Default stays snapshot isolation.
- FeatureRGABH gradient-driven adaptive buffer pool. Per-block heat gradient drives W-TinyLFU-style frequency admission and velocity-based speculative prefetch. HotSet hit rate improves from 0.639 to 0.803 on skewed workloads.
- FeatureDisk-resident DiskANN (Vamana) index for larger-than-RAM vector sets. HNSW remains the default; DiskANN is opt-in. Recall verified against exact brute-force ground truth.
- Featuregalaxdb_semantic_cache_hits_total metric. A restart-durable counter incremented once per semantic-cache hit, available on the /metrics endpoint.