feat: Add expression-keyed statistics for struct field pruning

[adriangb]

Summary

• Adds StatisticsKey enum and expression_statistics HashMap to Statistics for expression-keyed statistics lookup
• Teaches PruningPredicate to recognize get_field chains and rewrite them to synthetic dotted column names for statistics-based pruning
• Populates struct leaf column statistics from Parquet metadata into expression_statistics
• Enables file-level pruning for predicates on struct fields like WHERE col['a'] > 5

Motivation

Parquet stores struct fields (e.g., col.a inside struct col) as separate physical leaf columns with their own min/max/null statistics. DataFusion's Statistics struct uses column_statistics: Vec<ColumnStatistics> indexed 1:1 with schema fields, so struct field stats had nowhere to live. This blocked file-level pruning for predicates like WHERE get_field(col, 'a') > 5.

Design

By storing statistics keyed by expressions (StatisticsKey::FieldPath), the pruning system doesn't need to understand each expression type — it just checks "does this sub-expression have stats?" and uses them directly via dotted column name resolution.

Key changes

File	Changes
datafusion/common/src/stats.rs	StatisticsKey enum, expression_statistics field, updated helper methods
datafusion/common/src/pruning.rs	PrunableStatistics falls back to expression_statistics for dotted column names
datafusion/datasource-parquet/src/metadata.rs	Populates expression_statistics for struct leaf columns from Parquet metadata
datafusion/pruning/src/pruning_predicate.rs	get_field chain recognition, dotted name rewriting, schema hierarchy resolution
datafusion/proto-common/src/from_proto/mod.rs	Backward-compatible deserialization

Test plan

• All existing tests pass (cargo test -p datafusion-common && cargo test -p datafusion-pruning && cargo test -p datafusion-datasource-parquet)
• Full workspace compilation clean
• Integration test: Create Parquet file with struct column, verify expression_statistics is populated, verify file pruning based on struct field predicates

🤖 Generated with Claude Code