Runtime-aware plan embeddings

Develop a method for extending query-plan embeddings to incorporate execution-time information for candidate plans in cost model parameter tuning.

Background

The embedding-based similarity method represents structurally similar plans with nearby vectors, but the current embeddings use plan-structure information without runtime observations. Execution-time information is available only after a candidate plan has run, creating a challenge for incorporating it into the embedding representation. The paper explicitly leaves this extension unresolved.

References

It is unclear how to extend the plan embeddings to also include execution time information, as this information is only available for plans that have been executed.

— Query Performance Tuning with Optimal Exploration of Optimizer Cost Model Parameter Space  (2610.02607 - Wu et al., 2 Oct 2026) in Section 3.2.1, “Similarity by Bottleneck Matching”