feat: expose lambda and higher-order array functions#1561
Conversation
Add a Pythonic API for DataFusion's higher-order array functions and the lambda expressions they consume. - Rust: lambda_, lambda_var, array_transform, and array_any_match pyfunctions, plus a ResolveLambdaVariables analyzer rule so expression-builder plans (which emit unresolved lambda variables) resolve before optimization. - Python: array_transform / array_any_match (with list_transform, any_match, list_any_match aliases) accept either a Python callable or an explicit lambda built with lambda_ / lambda_var. Callables are introspected so their parameter names become the lambda parameters. - Tests and docs (expressions guide + agent skill), noting v1 limits: lambda expressions are not serializable, and SQL arrow syntax needs the DuckDB dialect.
Combine the eight higher-order function result tests into a single parametrized test_higher_order_function_results, and the two to_lambda rejection tests into test_to_lambda_rejects_invalid_arg. Each case keeps a readable id via pytest.param. Co-Authored-By: Claude <noreply@anthropic.com>
Add array_filter, the remaining lambda-based higher-order array function in DataFusion (alongside the already-exposed array_transform and array_any_match). Includes the list_filter alias matching upstream, tests, and documentation in the expressions guide and skill. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lead user-facing array-lambda docs with "lambda function" instead of "higher-order function," which is less recognizable to users. Drop the alias list, serialization caveat, and DuckDB-dialect note from the skill to keep it lean; those details already live in the docstrings. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
| contains a lambda raises ``Lambda not implemented``. SQL lambda syntax | ||
| (``x -> x * 2``) is only parsed by dialects that support lambdas; set | ||
| ``datafusion.sql_parser.dialect`` to ``DuckDB`` to use it. The Python |
There was a problem hiding this comment.
DuckDB will remove this syntax in v2.1. Perhaps add other dialects that support this syntax (spark, databricks, clickhouse, snowflake), and/or add the new duckdb syntax (``lambda x: x *2`) ?
There was a problem hiding this comment.
Do you think it makes sense to just remove this whole discussion until the syntax stabilizes?
There was a problem hiding this comment.
Sure, I see now that duckdb is the only dialect in PyDialect that supports lambdas. Just clarifying my comment, the new syntax is already stabilized and supported in duckdb and sqlparser-rs, alongside the old arrow syntax.
duckdb v2.1 will only remove lambda support for the old syntax, and likely use it for json operators. My main point was to also document syntax+dialect pairs that will not break in the future, since sqlparser-rs may follow the duckdb new syntax when it get's released, so it can correcly parse json operators.
There was a problem hiding this comment.
I've updated the documentation. I appreciate the feedback!
Other dialects (ClickHouse, Snowflake, Databricks) also enable lambda parsing via sqlparser-rs. Document the full set and recommend the ``lambda x: x`` keyword form, since DuckDB will drop the ``x -> x`` arrow form in v2.1. Parametrize the SQL test over the four dialects using the keyword syntax. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
gstvg
left a comment
gstvg
left a comment
There was a problem hiding this comment.
This is looking great @timsaucer. The support for callable is very nice
2 participants
Which issue does this PR close?
Closes #811
Rationale for this change
DataFusion gained lambda expressions and higher-order array functions (apache/datafusion#21172). The Python binding already had the read-only
Lambda,HigherOrderFunction, andLambdaVariablewrapper classes, but offered no way to construct a lambda or call the higher-order functions. This PR adds a Pythonic API for both.What changes are included in this PR?
Adds the following:
array_transformandarray_any_matchthat accept either a Python callable or an explicit lambda built withlambda_/lambda_varAre there any user-facing changes?
There are new APIs but they are all new additions.