ParquetExec::statistics::is_exact likely wrong/misunderstood

[crepererum]

A ParquetExec is created from a FileScanConfig and an optional filter predicate¹. These two are different, independent parameters -- at least the documentation is not implying that the predicate should be considered when constructing the FileScanConfig. Now the statistics for the ParquetExec are calculated by FileScanConfig::project:

https://github.com/apache/arrow-datafusion/blob/0f6931caa6f8b48e116a8e77e989c404f31f3f8d/datafusion/core/src/physical_plan/file_format/mod.rs#L213-L219

This forwards is_exact from the input which might have been set to true. However there is a predicate, is_exact should likely be false because some data may be removed which will mess up the exact statistic. So either the forwarding is wrong (at least when a predicate is given) or the docs are imprecise.

Note that this is unrelated to #5613 because this issue here is about the is_exact=true case.

Footnotes

1. And a metadata size hint, but this is irrelevant here. ↩