[yuzelin]

Purpose

Manifest entries may contain complete min/max statistics but an empty
null_counts array. Accessing this array in LeafPredicate causes an
ArrayIndexOutOfBoundsException during statistics filtering.

Treat an empty array as an unknown null count, allowing the existing
predicate logic to continue using available min/max bounds.

Tests

LeafPredicateTest#testEmptyNullCountsWithCompleteBounds

[JingsongLi]

Requirement fit: NEEDS-EVIDENCE. Implementation: no confirmed regression in the local guard.

Normal JDK 8 Maven verify passes all 15 LeafPredicateTest cases (including the added case), with the standard formatting/checkstyle/enforcer/license checks. Before treating this as a production fix, could you identify the supported writer/import path that emits nonempty min/max rows with an empty null-count array, and add a regression through its serialized manifest and an ordinary filtered table scan?

The Java SimpleStatsConverter paths I traced emit one null-count slot per emitted stats column (unknown counts use null), the Python DataWriter does the same, and SimpleStats.EMPTY_STATS is documented as zero-column statistics. The added unit test constructs the mismatched shape directly, so it demonstrates the local guard but does not yet establish a reachable table-read failure or verify schema/dense-stats adaptation. A concrete producer or existing affected table plus the normal scan reproduction would resolve the end-to-end applicability question.

[JingsongLi]

+1