Is your feature request related to a problem or challenge?

Joins are the largest missing piece of the DataFrame API, and they
force a design decision that the smaller methods have so far avoided:
how do we represent DataFusion Expr values from Java?

DataFusion's join surface:

• join(right, JoinType, left_cols, right_cols, filter: Option<Expr>)
  — equi-join on named columns, optional residual predicate.
• join_on(right, JoinType, on: impl IntoIterator<Item = Expr>) —
  arbitrary join predicates as a list of expressions.

join (the column-name form) is buildable today via the SQL-string
pattern (parse_sql_expr for the optional filter). join_on is not —
it fundamentally takes Expr values.

Describe the solution you'd like

Split into two phases.

Phase 1 — join (column-name form). Ship this first, no Expr
model required:

• DataFrame.join(DataFrame right, JoinType type, String[] leftCols, String[] rightCols)
• Overload accepting a String filter parsed via parse_sql_expr for
  the residual predicate.
• JoinType Java enum mirroring DataFusion's (INNER, LEFT,
  RIGHT, FULL, LEFT_SEMI, RIGHT_SEMI, LEFT_ANTI,
  RIGHT_ANTI, LEFT_MARK).

Phase 2 — joinOn (Expr form). Requires picking one of:

1. SQL strings. df.joinOn(right, INNER, "l.a = r.b", "l.c > r.d").
   Each predicate parsed via parse_sql_expr after both sides are
   stitched into a synthetic schema. Cheapest; no Java-side model.
2. Typed Expr builder. Java class hierarchy mirroring DataFusion's
   Expr (Column, BinaryExpr, Literal, ScalarUDF…). Discoverable but a
   large, ongoing maintenance surface that has to track DataFusion's
   Expr enum.
3. Defer. Keep only join (Phase 1) until a concrete user needs
   non-equi joins.

Recommendation: do (1) for symmetry with filter(String) and the
proposed withColumn(String) / sort(String). (2) is its own
multi-PR effort that the project may or may not want to commit to.

Describe alternatives you've considered

SQL joins via ctx.sql(...). Works for everything, but requires
naming both sides — losing the DataFrame composition story.

Additional context

Filing this as a discussion issue, not an implementation issue.
Phase 1 (column-name join) can be implemented immediately once the
JoinType enum and lifecycle (does the right-hand DataFrame get
consumed?) are settled. The Phase 2 decision is the real ask here.

Related: the right-hand DataFrame consume-vs-clone question is
identical to the set-operations issue and should be answered the same
way in both.