SQL path pruning turns a 17,644-second RelBench epoch into 1,494 seconds

A new RelBench result shows that SQL can make relational machine learning cheaper before the model trains. On RelBench’s beer-churn test split, RelGT with random three-hop sampling took 17,644 seconds per epoch and reached 0.777 test AUCROC. Replacing random sampling with MetaSieve cut the epoch to 1,494 seconds—an 11.8× measured speedup—while test AUCROC rose to 0.799.

That pairing matters. This is not a speed claim bought by accepting a worse test score: for this task and backbone, both the reported runtime and held-out metric improved. It is also not an NL2SQL result. RelBench evaluates predictive learning over related tables; MetaSieve uses SQL joins and aggregations to decide which schema paths a graph model should sample.

What the system prunes

MetaSieve computes statistics for candidate foreign-key “metapaths,” scores them using label dependence, fanout cost and coverage, then removes paths judged uninformative before GNN training. The authors evaluate three backbones—HeteroGraphSAGE, HGT and RelGT—on classification and regression tasks from five RelBench databases. The largest evaluated database has about 21 million rows; the ratebeer database used for the headline result has about 14 million rows.

The preparation cost is visible rather than hidden. For beer-churn, MetaSieve processed 300,000 training samples in 32 batches with four SQL workers. SQL work took 2,318 seconds and total preprocessing took 2,340 seconds, reported as 0.13 of one random-sampling RelGT epoch. Because the selected paths can be reused across training epochs, that one-time cost is small beside the per-epoch reduction in this setup.

The benchmark boundary

The result still needs its full experimental label. Runs used one Nvidia H100 with 80 GB of memory and 2 TB of RAM. MetaSieve’s task-specific uncertainty parameter was selected from {0.1, 0.2, 0.3, 0.4} using validation performance, with epoch time used as a secondary criterion when validation results were close. The method also assumes the database contents, schema and task definition remain fixed; a change requires regenerating its sampling rules.

Those constraints do not erase the result, but they define it. The defensible takeaway is narrower than “SQL makes graph learning 11.8× faster”: on the RelBench beer-churn test task, with RelGT, three hops and k=50, SQL-based metapath pruning reduced the measured epoch time from 17,644 to 1,494 seconds while moving test AUCROC from 0.777 to 0.799. The authors have published the implementation, making this unusually testable for a new benchmark claim.