I think the behavior introduced in DRILL-4264 needs to be fixed:
- Improve the metadata to flag a column as missing (no back-tick business)
- All readers should apply the same fix
- This will give an opportunity to upstream operators to provide explicit 
handling (e.g., Filter)

@ilooner, ideally the RecordBatchSizer should be using SchemaPath as keys to 
handle nested column names; I guess the Planner is injecting a project step to 
disambiguate such use-cases.

For now, I let you guys decide; we either:
a) Accept this temporary workaround and create a new JIRA to implement the 
right fix
b) Fix the Parquet reader (remove the back-tick) with the side-effect that the 
schema-on-change behavior for filter based queries is broken
c) Implement the right fix for DRILL-4264 as discussed above

[ Full content available at: https://github.com/apache/drill/pull/1445 ]
This message was relayed via gitbox.apache.org for [email protected]

Reply via email to