Step to reproduce:
1. create an partitioned avro table, "avro1"
2. insert some data with snappy compression
3. add a column to the table
4. insert more data including new column
5. create a second table "avro2" with the old table schema (not include the new added column)
6. msck repair table "avro2"
7. invalidate metadata and query table "avro2", got weird data at the end.
8. disable codegen and query table again, it shows proper result.
seems like when codegen enabled, it gets row offset wrong from the data with extra column.
sometimes impalad hits DCHECK and crashes.