Parquet MR 1.9.0 and 1.8.2 both have a bug,
PARQUET-1217, that means that pushing certain predicates to Parquet scanners can return fewer results than they should.
The bug triggers in Spark when:
- The Parquet file being scanner has stats for the null count, but not the max or min on the column with the predicate (Apache Impala writes files like this).
- The vectorized Parquet reader path is not taken, and the parquet-mr reader is used.
- A suitable <, <=, > or >= predicate is pushed down to Parquet.
The bug is that the parquet-mr interprets the max and min of a row-group's column as 0 in the absence of stats. So col > 0 will filter all results, even if some are > 0.
There is no upstream release of Parquet that contains the fix for
PARQUET-1217, although a 1.10 release is planned.
The least impactful workaround is to set the Parquet configuration parquet.filter.stats.enabled to false.