[NoahKusaba]

Which issue does this PR close?

• Follows up on feat: make the plan nodes inspectable and rebuildable #20 (comment) and feat: make the plan nodes inspectable and rebuildable #20 (comment).

Rationale for this change

#20 made IcebergDataSource::new_with_predicate public so that a codec can rebuild a scan from its accessors. It takes every part of the scan as a positional argument. A scan option added later would mean adding an argument, which is a breaking change. A builder lets new options be added without breaking callers.

What changes are included in this PR?

• Replaces IcebergDataSource::new_with_predicate with IcebergDataSource::builder(table, schema), exported from iceberg_datafusion::read as IcebergDataSourceBuilder. It has these setters:

  • with_projection(Option<&Vec<usize>>): selects columns of schema by index, as TableProvider::scan passes them.
  • with_snapshot_id(Option<i64>)
  • with_filters(&[Expr]): converts DataFusion filters to an Iceberg predicate.
  • with_predicate(Option<Predicate>): sets an already-converted predicate.
  • with_fetch(Option<usize>)

  The Option setters let a codec pass the accessors' values straight through. build() returns a Result, because a projection index can be out of range.

  IcebergDataSource::builder(scan.table().clone(), scan.schema())
      .with_snapshot_id(scan.snapshot_id())
      .with_predicate(scan.predicate().cloned())
      .with_fetch(scan.fetch())
      .build()?

• Removes the crate-internal positional constructor IcebergDataSource::new. Both table providers now build their scans with the builder.

• Renames the accessor predicates() to predicate(), since it returns a single Predicate.

Are there any user-facing changes?

Yes. IcebergDataSource::new_with_predicate is removed in favor of the builder, and IcebergDataSource::predicates() is renamed to predicate(). Both were added in #20 and aren't in any release.

Are these changes tested?

Yes.

• The integration tests that rebuild a scan from its accessors now use the builder, including the scan built over part of the table's columns.
• A unit test checks that with_projection returns columns in the projection's order and rejects an out-of-range index.
• Every provider scan goes through the builder, so the existing scan tests cover it too.

[gabotechs]

👍 nice, thanks @NoahKusaba!