[Bug] Java Parquet reads drop matching rows when a filtered nested field is pruned from the read type #10378

Search before asking

• I searched in the issues and found nothing similar.

Paimon version
master @ b0af456

Compute Engine
Java API (ReadBuilder). Flink/Spark SQL are not affected.

Minimal reproduce step
Create a Parquet table id INT, s ROW<a INT, b INT> and write (0, {10, 7}), (1, {20, 8}). Read with a nested filter on s.b and a read type that prunes s to ROW<a>, without executeFilter():

Predicate filter = new PredicateBuilder(rowType)
        .equal(new NestedFieldTransform(new FieldRef(1, "s", sType), singletonList("b")), 7);
table.newReadBuilder().withFilter(filter).withReadType(/* id, s ROW<a> */);

What doesn't meet your expectations?
No rows are returned. Row 0 must be returned, as it is when the read type keeps s.b.

Anything else?
#9858 drops filters on unprojected columns in the ParquetReaderFactory constructor, but checks only top-level names. Since #9423 pushes nested predicates to Parquet, s.b = 7 passes because s is projected; Parquet reads the unrequested s.b as all-null and skips the row group. executeFilter() is safe because it widens the read type to the filter fields.

Are you willing to submit a PR?

• I'm willing to submit a PR!