Purpose

Linked issue: close #146

Add column projection support to ReadBuilder, allowing users to specify which columns to read from Parquet data files. This reduces unnecessary I/O for wide tables, especially on remote storage (S3/OSS), by leveraging the existing ProjectionMask in the Parquet reader layer.

Brief change log

• crates/paimon/src/table/read_builder.rs

  • Add ReadBuilder::with_projection(&[&str]) — accepts column names, stores them for deferred resolution.
  • Add resolve_projected_fields — resolves names to DataFields in caller-specified order; rejects unknown columns (ColumnNotExist) and duplicates (ConfigInvalid).
  • Change TableRead::read_type from &[DataField] to Vec<DataField> to support projected subsets.

• crates/paimon/src/arrow/reader.rs

  • Always apply ProjectionMask to clip Parquet schema to read_type columns (previously only applied conditionally).
  • Add column reorder step after Parquet read to guarantee RecordBatch schema matches read_type order (Parquet ProjectionMask returns columns in file-schema order, not projection order).
  • Use fail-fast error propagation instead of silent fallback when projected columns are missing from batch schema.

Tests

• cargo test -p paimon-integration-tests test_read_with_column_projection
• cargo test -p paimon-integration-tests test_read_projection_empty
• cargo test -p paimon-integration-tests test_read_projection_unknown_column
• cargo test -p paimon-integration-tests test_read_projection_all_invalid
• cargo test -p paimon-integration-tests test_read_projection_duplicate_column

API and Format

• New public API: ReadBuilder::with_projection(&[&str]) -> Self
• TableRead::read_type() return type changed from &[DataField] (borrowed from table schema) to &[DataField] (borrowed from owned Vec<DataField>). No breaking change for callers.
• No storage format changes.

Documentation