postgresml

By running ML workloads inside the database process with GPU acceleration, PostgresML can perform inference on data without extracting it to a separate ML platform, achieving latency and throughput characteristics impossible with network-separated architectures.

The extension supports training classical ML models (XGBoost, LightGBM, scikit-learn algorithms) on database tables, running inference with user-defined SQL functions, fine-tuning transformer models, generating text embeddings for semantic search, and performing vector similarity search with pgvector integration all through standard SQL queries.

This enables application developers to add ML features using existing database skills without learning Python ML infrastructure, and allows ML engineers to query production data directly without ETL pipelines.

PostgresML is open-source under the MIT license and available both as a self-hosted extension for existing PostgreSQL instances and as a managed cloud database service.

GPU-accelerated inference within the database makes it competitive with dedicated ML inference services for applications where the data already lives in PostgreSQL.

It is particularly attractive for applications requiring real-time personalization, fraud scoring, or recommendation features where the latency of round-tripping to an external ML service is prohibitive and the operational simplicity of a single database system is valuable.