Support reusing the physical plan in prepared statements #26706

Is your feature request related to a problem or challenge?

Building and optimizing the physical plan can be relatively expensive when we execute small queries. Currently, DataFusion's prepared statements only reuse the logical plan.

In a proof of concept, I found that we can improve the performance of small queries in >3x if we just reuse the physical plan.

This was done in a hacky way that first created the physical plan using placeholder arguments and at runtime replaced nodes like DataSourceExec and FilterExec with the passed arguments.

Describe the solution you'd like

For a clean implementation, I think we need the following:

• Build ExecutionPlan from queries with placeholder values.
• Add a ExecutionPlan::set_placeholder (or similar) method that is implemented by every operator that uses placeholder values, changing only what is necessary and returning a new plan that can be consumed.

This would certainly be a breaking change.

Describe alternatives you've considered

Hacky downstream implementation.

Additional context

Related:

• Prepared physical plan reusage #14342
• Stateless execution plans for plan caching #19351

Willingness to contribute

None