wombatu-kun opened a new pull request, #19456:
URL: https://github.com/apache/hudi/pull/19456

   ### Describe the issue this Pull Request addresses
   
   Closes #19387, which carries the analysis, the trigger conditions and the 
measurements.
   
   ### Summary and Changelog
   
   Resolve pushed-down predicate columns the same way the projection is already 
resolved, so a predicate prunes row groups on the column it was written for 
instead of on whichever column sits at its stale metastore ordinal.
   
   - New `remapPredicateColumnIndicesToPhysical`, applied through a private 
`getPushdownPredicate`. The name-based mode and every path where pushdown is 
off are untouched.
   - A column the file does not carry is dropped from the predicate. That costs 
pruning, never a row, because `HudiMetadata.applyFilter` returns the whole 
regular predicate as the remaining filter.
   - Handles are deduplicated on the physical index, not on the rebuilt handle: 
two names differing only by case resolve to one file column while staying 
unequal, and pushing both down would fail the split on 
`getParquetTupleDomain`'s duplicate `ColumnDescriptor`.
   - `remapColumnIndicesToPhysical` is re-expressed on shared private helpers 
so the two remaps cannot drift apart. One difference is folded in: the handle 
rebuild now passes `getBaseType()` rather than `getType()`, since the 
constructor's fourth argument is the base column's type. Unreachable today 
because `HudiMetadata` has no `applyProjection`, and now covered by a test.
   
   Tests: 8 unit tests on the new helper, plus a new 
`TestHudiPredicatePushdownColumnOrdinals` with 5 end-to-end tests through 
`createPageSource` over a synthetic base file, covering the stale ordinal, 
pruning still landing on the right column, a stale ordinal arriving through a 
dynamic filter, a column added after the base file was written, and agreement 
between the positional and name-based modes.
   
   ### Impact
   
   None with `hudi.parquet.use-column-names=true`, the connector default, and 
none where predicate pushdown is off. With `false`, affected queries stop 
losing rows and read fewer of them. Two side effects worth naming: a predicate 
on a column the base file lacks is no longer pushed down at all, and a promoted 
metastore type now reaches the real column, so positional mode inherits 
whatever the name-based mode already does with that combination.
   
   ### Risk Level
   
   low
   
   ### Documentation Update
   
   none
   
   No new or changed config property. `hudi.parquet.use-column-names` already 
exists and its description is unchanged.
   
   ### Contributor's checklist
   
   - [x] Read through [contributor's 
guide](https://hudi.apache.org/contribute/how-to-contribute)
   - [x] Enough context is provided in the sections above
   - [x] Adequate tests were added if applicable
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to