cuGraph SQL API

Algeon exposes NVIDIA RAPIDS cuGraph algorithms as GPU UDTFs (user-defined table functions) over DataFusion SQL sessions. The session can run in a Rust process or be served through Flight SQL. Each cugraph_* UDTF reads an edge relation, builds a GPU graph, runs an algorithm, and returns rows you can compose with ordinary SQL. The cuGraph UDTFs index lists every function; Discover & Validate GPU UDTFs and the GPU UDTF Catalog API cover discovery and validation.

Questions to explore

Each question opens a captured example on its UDTF page. The examples read the demo datasets. The MerRec queries ran over the full 1.27 billion events in native_required mode.

MerRec

The MerRec investigation turns Mercari US likes and purchases from May through October 2023 into brand graphs:

• Which brands moved toward Barbie's audience after the July film? (cugraph_personalized_pagerank, cugraph_pagerank)
• Which 5,000 shoppers should a Pokemon campaign reach? (cugraph_personalized_pagerank, then cuvs_brute_force_knn and cuml_forest_predict)
• Which brands belong on the same discovery page? (cugraph_jaccard, cugraph_leiden)
• Which brands connect the shopping worlds? (cugraph_betweenness_centrality, cugraph_leiden)
• Where do brand neighborhoods appear when shared fans determine the map? (cugraph_force_atlas2)
• Which catalog entries are one brand listed twice? (cugraph_jaccard over 1,999,000 pairs)
• Is session_id a per-user session key? (cugraph_weakly_connected_components)

Citation network

DBLP / AMiner V12: 4.9M papers linked by 45.6M citations.

• Which papers rank highest when citations from influential papers count more? (cugraph_pagerank)
• What should I read to understand BERT beyond its own bibliography? (cugraph_personalized_pagerank)
• Do citation communities match the research fields in our catalog? (cugraph_louvain)
• How are the papers around the Transformer connected? (cugraph_force_atlas2)
• Which papers act as reading guides, and which works are their shared foundations? (cugraph_hits)
• Which papers connect the most citation routes across AI research? (cugraph_betweenness_centrality)
• Which papers should I read after the Transformer if I want familiar references? (cugraph_jaccard)
• Which research pockets are disconnected from the main citation network? (cugraph_weakly_connected_components)
• What is the shortest citation chain from BERT back to LSTM? (cugraph_bfs)

EMBER2024

Win32 malware files from the challenge split, linked to the import symbols they share.

• Which import patterns should an analyst compare around an unlabeled file? (cugraph_personalized_pagerank, cugraph_jaccard)

The remaining cuGraph pages, from degree counts to spectral clustering, read the citation network.

Use functions_only execution to keep scans and surrounding SQL on DataFusion CPU while graph algorithms run on GPU. The same page covers backend configuration and host staging limits.

Graph input and call shape

Every cuGraph call uses named arguments. The required edges argument is a parenthesized relation subquery whose projection supplies canonical src and dst columns. Optional graph and algorithm values use the same named syntax:

SELECT *
FROM cugraph_pagerank(
  edges => (SELECT src, dst, weight FROM citation_edges),
  alpha => 0.9,
  directed => true
);

Algorithms that need additional relations name them explicitly, for example personalization => (SELECT vertex, value FROM ppr_seeds) for personalized PageRank or vertex_pairs => (SELECT first, second FROM candidate_pairs) for similarity UDTFs. BFS source selectors and SSSP's source vertex are named value arguments as well. Each UDTF page gives the complete signature and the canonical columns required for every relation role.

Relation subqueries can read tables, views, CTEs, joins, and filters. Register a derived relation as a temporary view when it must also be checked through gpu_validate_call, because the metadata envelope resolves registered names and does not contain SQL text. See Integration for source and session setup, and Graph Inputs and Construction for vertex-ID contracts, directedness, vertex renumbering, and edge normalization.

The GPU table functions design traces cuDF edge ownership, cuGraph completion and the return to native SQL operators. It distinguishes device copies from host materialization.

Result column names

Result columns follow one naming convention across UDTFs. The edge input columns (src, dst, weight, edge_id) are separate and keep their names.

Result	Columns
Vertex identity	vertex
One score per vertex	value
Edge	src, dst, plus value for an edge score
Community assignment	community
Connectivity label	component
Vertex pair similarity	first, second, similarity

Each UDTF page lists its exact result schema, including algorithm-specific columns such as hub_score and authority_score for HITS.

Reuse in bounded servers

A bounded Flight SQL server can retain constructed graphs for repeat calls with a strongly identified, eligible input. The current reusable path covers BFS without include_edges, PageRank, Jaccard, Sorensen, Overlap, and Cosine. Other calls bypass the cache. Configure and observe it through Configuration.