vtgate: bind recursive CTE column filters as arguments in unmerged term queries by GrahamCampbell · Pull Request #20631 · vitessio/vitess

GrahamCampbell · 2026-07-19T15:31:41Z

Description

When a recursive CTE does not merge into a single route, WHERE predicates in the recursive term that reference CTE columns are never rewritten to per iteration bind variables. Filters over CTE columns alone are emitted as references to a table that does not exist in the term query, so every execution fails with an unknown column error. Predicates mixing CTE and term table columns fail the same way, including through comma join linkage, and OR predicates spanning both sides are silently dropped, which returns wrong results. A column referenced only in the WHERE clause is also never registered in the RecurseCTE primitive's join variables.

The term's WHERE path now binds these predicates at translation time, exactly as the join ON path always has. They become reserved arguments supplied per iteration, the referenced columns are registered in the join variables, and the tracked original form is restored when the whole CTE later merges into a single route, where the column reference is correct inside the pushed WITH RECURSIVE. Routes gain the same binding as a safety net for predicates arriving through other paths. Offset planning and the projection path both resolve recursion columns positionally through a declared column list; each previously failed with an internal error, the former reachable through the ON clause and the latter whenever the outer query selected a declared name. A declared list that reaches this resolution is guaranteed well formed by #20633, since only seed projections with a statically known cardinality plan this far.

Eight planner cases pin the behavior, covering the plain filter, the comma join, mixed, and OR forms, a filter on a column absent from the term select list, a declared column list selected through its declared names, a declared list whose names swap the seed aliases, and a fully merged CTE keeping its legal column form. Two end to end tests walk a chain of rows across two shards and compare results against MySQL, one through seed names and one through a declared column list, with the data sized so a dropped filter changes the result. Predicates containing subqueries settle after term translation and are not covered by this change.

Related Issue(s)

This PR depends on #20633 being merged first. Its first commit is that change.

Checklist

"Backport to:" labels have been added if this change should be back-ported to release branches
If this change is to be back-ported to previous releases, a justification is included in the PR description
Tests were added or are not required
Did the new or modified tests pass consistently locally and on CI?
Documentation was added or is not required

Deployment Notes

Recursive CTEs whose terms filter on CTE columns previously failed with unknown column errors or silently dropped those filters unless the whole CTE merged into a single route. Direct term predicates are now bound per iteration, and selecting the declared column names of such a CTE previously failed with an internal error and now works. Predicates containing subqueries remain affected in the unmerged form. No migrations or configuration changes.