GraphStore is a PyTorch Geometric object that stores a graph’s structural information—its edge indices—outside the training process. It enables node sampling when the graph is too large for memory, typically as part of a remote backend. section 7.9
A remote GraphStore implementation typically connects these pieces:
- Connect to a graph database or other remote data source. section 7.9
- Implement CRUD operations such as put_edge_index(), get_edge_index(), and remove_edge_index(). section 7.9
- Link the GraphStore to a sampler that retrieves edge indices and produces subgraphs from input nodes. section 7.9
- Pass the GraphStore with a FeatureStore to NodeLoader for batch-based training. section 7.9
GeoGrid remote backend
GeoGrid stored graph structure in Neo4J and node features in RocksDB. Its team implemented secure database connections and GraphStore CRUD operations, then integrated a custom sampler and NodeLoader. The resulting pipeline supported large-graph training but required testing and optimization to reduce I/O latency. section 7.9.1
