mirror of
https://github.com/langchain-ai/datafusion.git
synced 2026-07-19 23:35:20 -04:00
Xander
d59cdfe999
## Which issue does this PR close? - Closes #17508 ## Rationale for this change The previous implementation used UUID-based aliasing as a workaround to prevent duplicate names for literals in Substrait plans. This approach had several drawbacks: - Non-deterministic plan names that made testing difficult (requiring UUID regex filters) - Only addressed literal naming conflicts, not the broader issue of name deduplication - Added unnecessary dependency on the `uuid` crate - Didn't properly handle cases where the same qualified name could appear with different schema representations ## What changes are included in this PR? 1. Enhanced NameTracker: Refactored to detect two types of conflicts: - Duplicate schema names: Tracked via schema_name() to prevent validate_unique_names failures (e.g., two Utf8(NULL) literals) - Ambiguous references: Tracked via qualified_name() to prevent DFSchema::check_names failures when a qualified field (e.g., left.Utf8(NULL)) and unqualified field (e.g., Utf8(NULL)) share the same column name 2. **Removed UUID dependency**: Eliminated the `uuid` crate from `datafusion/substrait` 3. **Removed literal-specific aliasing**: The UUID-based workaround in `project_rel.rs` is no longer needed as the improved NameTracker handles all naming conflicts consistently 4. **Deterministic naming**: Name conflicts now use predictable `__temp__N` suffixes instead of random UUIDs Note: This doesn't fully fix all the issues in #17508 which allow some special casing of `CAST` which are not included here. ## Are these changes tested? Yes: - Updated snapshot tests to reflect the new deterministic naming (e.g., `Utf8("people")__temp__0` instead of UUID-based names) - Modified some roundtrip tests to verify semantic equivalence (schema matching and execution) rather than exact string matching, which is more robust - All existing integration tests pass with the new naming scheme ## Are there any user-facing changes? Minimal. The generated plan names are now deterministic and more readable (using `__temp__N` suffixes instead of UUIDs), but this is primarily an internal representation change. The functional behavior and query results remain unchanged.