What is built. What is Enterprise.

LSM + PAX columnar storage, ART primary index, Monkey-optimal Bloom filters

NUMA-aware buffer pool, Lazy Leveling compaction with MVCC GC

WAL with crash recovery and write-stall mitigation

Durable catalog: table definitions, storage modes, and constraints survive restart

Automatic memory/configuration tuning (buffer pool, memtable, compaction) from host RAM/CPU

RGABH gradient-driven adaptive buffer pool: per-block heat gradient with W-TinyLFU admission and velocity-based speculative prefetch. HotSet hit rate +16 pp on skewed workloads.

io_uring backend with graceful tokio fallback

AES-256-GCM encryption at rest, pluggable KMS: local file, env var, external command, HashiCorp Vault, and native AWS KMS / GCP Cloud KMS / Azure Key Vault over REST (no vendor SDKs)

Joins, aggregates (COUNT/SUM/AVG/MIN/MAX), GROUP BY / HAVING, DISTINCT, ORDER BY / LIMIT / OFFSET via an embedded DataFusion engine

Native fast path for point reads, filtered single-table scans, and vector search

FROM-less scalar queries (SELECT 1+1, version(), current_database())

Real per-row expression evaluation in INSERT/UPDATE (e.g. SET bal = bal - 30)

PRIMARY KEY uniqueness enforced (SQLSTATE 23505), never a silent overwrite

Typed arithmetic errors (22012 division-by-zero, 22003 overflow, 42804 type mismatch)

Single-column secondary indexes (CREATE/DROP INDEX) with full-scan fallback

Explicit BEGIN / COMMIT / ROLLBACK with snapshot isolation and read-your-writes

SAVEPOINT / ROLLBACK TO / RELEASE

Write-write conflict detection (SQLSTATE 40001)

Serializable Snapshot Isolation: opt-in commit-time certifier that aborts write-skew anomalies with SQLSTATE 40001. Default stays snapshot isolation.

Mutable HNSW with crash-safe delta buffer, SQ8/FP16/RaBitQ quantization, parallel build, and on-disk index persistence (reload on open, no re-embed on restart)

Disk-resident DiskANN (Vamana) index for larger-than-RAM vector sets with FreshDiskANN incremental insert/consolidate. HNSW stays the default; DiskANN is opt-in.

EMBEDDING MODEL columns with a local HuggingFace sidecar (any model, no API key): all-MiniLM-L6-v2, BGE-M3, Qwen3-Embedding, EmbeddingGemma-300M, LFM2.5-Embedding-350M

SEMANTIC_MATCH in any WHERE clause, combinable with SQL filters, joins, aggregates

Semantic result caching: CREATE SEMANTIC CACHE FOR TABLE ... SIMILARITY ... TTL ... — repeated queries answered from cache, skipping HNSW search

AT VERSION time-travel (including over flushed on-disk/SST data) plus version tags and exact historical vector search via CONSISTENCY 'SEMANTIC_SNAPSHOT'

MinHash near-duplicate detection via WHERE NOT DUPLICATE

FOR TRAINING Lance export to PyTorch (incl. embedding columns), zero-copy mmap

EU AI Act Article 13 data lineage

SCRAM-SHA-256 wire authentication

TLS 1.2/1.3 transport encryption (rustls, no OpenSSL)

RBAC with table-level GRANT / REVOKE (SQLSTATE 42501)

JSONL security audit log (authN / authZ / admin events)

PostgreSQL wire protocol: simple and extended query / prepared statements, parsed-statement cache

COPY FROM STDIN / COPY TO STDOUT bulk ingestion

Python client: embedded mode (no server) and remote mode

pg_catalog compatibility (WHERE / projection / COUNT honored)

Backup/restore to a local path or object storage over REST: S3 (+ S3-compatible like MinIO/R2), GCS, Azure Blob, checksum-validated

Observability: /health and /metrics (Prometheus) on port 9090

Pre-built binaries: Linux x86_64/aarch64, macOS Intel/Apple Silicon, Windows x86_64

Docker image, Homebrew tap, pip wheel (galaxdb-client)