Abstract
Training Graph Neural Networks on graphs that exceed main memory often requires exporting data from graph database systems to external machine learning frameworks. We present a database-native, out-of-core GraphSAGE training pipeline integrated into MillenniumDB and exposed through four GQL procedures. The pipeline materializes reusable mini-batches, distributes features across GPU memory, pinned host memory, and NVMe storage based on access frequency, applies frequency-based tiering to the graph topology as well, and writes the learned embeddings back to MillenniumDB for graph pattern and similarity queries. We evaluate on Cora, ogbn-arxiv, ogbn-products, and ogbn-papers100M (111M nodes, 1.6B directed edges). On a commodity desktop (32 GiB RAM, 16 GB GPU), PyG, DGL, and Neo4j GDS run out of memory on ogbn-papers100M. MillenniumDB completes 50 training epochs in 41.2 minutes after a one-time 95.7-minute preparation, matching the accuracy of the state-of-the-art out-of-core system DiskGNN. An ablation shows that topology tiering speeds up offline sampling by 3.81× without altering the sampled mini-batches.
