Is your feature request related to a problem or challenge?
DeleteFileIndex currently indexes every position delete by partition, so each data file in a partition receives the partition's entire position delete list during scan planning. Readers re-filter rows by path when applying deletes, so results are correct but overbroad: delete files are loaded in proportion to the partition's delete history rather than the data files actually scanned. This is noticeable on merge-on-read tables with many small file-scoped position delete files.
The spec says "The data file's file_path is equal to the delete file's referenced_data_file if it is non-null", and the code carries a TODO for exactly this in get_deletes_for_data_file. Java implements it via DeleteFileIndex.posDeletesByPath.
Describe the solution you'd like
Mirror Java's path-based indexing. A position delete references a single data file through the referenced_data_file field or through equal file_path column bounds. It is indexed by that path and returned only for the matching data file. All other position deletes keep the existing partition routing.
Willingness to contribute
None
Is your feature request related to a problem or challenge?
DeleteFileIndex currently indexes every position delete by partition, so each data file in a partition receives the partition's entire position delete list during scan planning. Readers re-filter rows by path when applying deletes, so results are correct but overbroad: delete files are loaded in proportion to the partition's delete history rather than the data files actually scanned. This is noticeable on merge-on-read tables with many small file-scoped position delete files.
The spec says "The data file's file_path is equal to the delete file's referenced_data_file if it is non-null", and the code carries a TODO for exactly this in get_deletes_for_data_file. Java implements it via DeleteFileIndex.posDeletesByPath.
Describe the solution you'd like
Mirror Java's path-based indexing. A position delete references a single data file through the referenced_data_file field or through equal file_path column bounds. It is indexed by that path and returned only for the matching data file. All other position deletes keep the existing partition routing.
Willingness to contribute
None