neo4j-snowflake-graph-analytics-skill

Install the Neo4j Graph Analytics Native App from the Snowflake Marketplace, create node and relationship projection views with cast key columns, then call algorithms like Neo4j_Graph_Analytics.graph.page_rank with a compute pool name and a JSON project-compute-write config. Results are written back to Snowflake output tables.

Available algorithms include WCC, Louvain, Leiden, Label Propagation, K-Means, Triangle Count, PageRank, Article Rank, Betweenness, Degree, Dijkstra, Delta-Stepping, BFS, Yen's, Max Flow, FastPath, Node Similarity, KNN, FastRP, Node2Vec, HashGNN, and GraphSAGE training and prediction procedures.

The graph engine only accepts BIGINT, DOUBLE, ARRAY, and VECTOR node properties, and BIGINT, DOUBLE, or INT relationship properties. Snowflake views inherit source column types, so every property column needs an explicit CAST, and VARCHAR, DATE, BOOLEAN, and VARIANT columns must be converted or dropped.

No, this Native App only runs graph algorithms through SQL CALL procedures and does not support Cypher or Neo4j DBMS queries. For Cypher workloads use a Neo4j database skill such as neo4j-cypher instead.

Algorithm output tables contain numeric NODEID or SOURCENODEID and TARGETNODEID keys because VARCHAR labels are dropped during projection. Join the result table back to the original source table on the key column to retrieve human-readable names and attributes.

GPU pools such as GPU_NV_S are appropriate for GraphSAGE training jobs, which are slow on CPU. For most other algorithms and small graphs, the default CPU_X64_XS pool is sufficient, with larger CPU or HIGHMEM pools for bigger graphs.