GalaxDB: A Unified AI-Native Storage Engine for Transactional, Analytical, and Vector Workloads

Modern AI applications require simultaneous access to transactional row data, analytical column scans, and vector similarity search – capabilities that today demand three or more separate systems. We present GalaxDB, a unified storage engine that serves all three workload classes from a single binary (galaxdb-server: 7.9 MB, galaxdb-sidecar: 7.6 MB on Linux x86-64), while introducing AI-native primitives unavailable in any existing database: MinHash LSH near- duplicate detection at write time, per-row embedding model version tracking, training-precision-aware Lance exports with 4–32× I/O reduction per training epoch, and EU AI Act Article 13-compliant data lineage. The v1 storage engine combines a PAX columnar block format, an Adaptive Radix Tree (ART) primary index, Monkey-optimal Bloom filters, a NUMA-aware dual-partition buffer pool, Lazy Leveling compaction with MVCC garbage collection integrated into the compaction pass, BVLSM WAL-time key-value separation, AEGIS-256 transparent encryption, and a three-mechanism write-stall mitigation system. The vector index implements parallel mutable HNSW with SQ8 quantization and a WAL- backed delta buffer for crash-safe incremental updates without index rebuilds. Benchmarked on AWS c6id.4xlarge (Intel Xeon Platinum 8375C, 32 GiB DDR4, 884 GB NVMe) against the SIFT1M ground-truth dataset, GalaxDB achieves: recall@10 of 0.983 at ef=100 (matching hnswlib 0.8 at 0.980); build speed of 15,114 vectors/second (10.7% faster than hnswlib 0.8’s 13,656 vec/sec, measured during development on the same hardware; raw hnswlib log not retained in this release); 258,555 storage-level writes/second with write p99 of 377 μs under sustained compaction load (vs. 1–10 s for naive LSM); wire-protocol SELECT throughput of 7,390 QPS (2.3× PostgreSQL 16)†; column scan at 4.49 GB/s with 80% zone-map block skip rate; and cold-cache point read p50 of 147 μs on a 50M-row dataset exceeding RAM. All results are fully reproducible from named commands against named datasets on named hardware, except where noted. The v2 distributed design – covering RGABH gradient-driven adaptive storage, causal-consistency clustering with two-level IVF+HNSW, and an active learning feedback loop – is fully specified and implementation-ready.