[Notes]

[heathersherry]

• ML models for DB tasks [pic here to show different types of integration]

• Zero-shot cardinality estimation problem (query plan --> graph, can use GNN or graph transformer), CardBench arxiv – a good pretrained model for cardinality estimation [pic here] – conclusion: learning based methods are much better than heuristic methods, zero-shot learning is challenging but instance-specific FT can improve accuracy, GNN is better than graph transformer (in accuracy)
• Foundational Database Models [two-three pics here] main idea: involved experts in data --> logical plan --> physical plan, then run on the downstream tasks. Conclusions: task-independency, initial results, and pretrained models are important.