refactor: add ToArrayOp to the sqlglot compiler

Author: chelsea-lin

bigframes/core/compile/sqlglot/expressions/array_ops.py

@@ -22,13 +22,10 @@
 from bigframes import operations as ops
 from bigframes.core.compile.sqlglot.expressions.typed_expr import TypedExpr
 import bigframes.core.compile.sqlglot.scalar_compiler as scalar_compiler
+import bigframes.dtypes as dtypes
 
 register_unary_op = scalar_compiler.scalar_op_compiler.register_unary_op
-
-
-@register_unary_op(ops.ArrayToStringOp, pass_op=True)
-def _(expr: TypedExpr, op: ops.ArrayToStringOp) -> sge.Expression:
-    return sge.ArrayToString(this=expr.expr, expression=f"'{op.delimiter}'")
+register_nary_op = scalar_compiler.scalar_op_compiler.register_nary_op
 
 
 @register_unary_op(ops.ArrayIndexOp, pass_op=True)
@@ -66,3 +63,27 @@ def _(expr: TypedExpr, op: ops.ArraySliceOp) -> sge.Expression:
     )
 
     return sge.array(selected_elements)
+
+
+@register_unary_op(ops.ArrayToStringOp, pass_op=True)
+def _(expr: TypedExpr, op: ops.ArrayToStringOp) -> sge.Expression:
+    return sge.ArrayToString(this=expr.expr, expression=f"'{op.delimiter}'")
+
+
+@register_nary_op(ops.ToArrayOp)
+def _(*exprs: TypedExpr) -> sge.Expression:
+    do_upcast_bool = any(
+        dtypes.is_numeric(expr.dtype, include_bool=False) for expr in exprs
+    )
+    if do_upcast_bool:
+        sg_exprs = [_coerce_bool_to_int(expr) for expr in exprs]
+    else:
+        sg_exprs = [expr.expr for expr in exprs]
+    return sge.Array(expressions=sg_exprs)
+
+
+def _coerce_bool_to_int(typed_expr: TypedExpr) -> sge.Expression:
+    """Coerce boolean expression to integer."""
+    if typed_expr.dtype == dtypes.BOOL_DTYPE:
+        return sge.Cast(this=typed_expr.expr, to="INT64")
+    return typed_expr.expr

tests/system/small/engines/test_array_ops.py

@@ -26,7 +26,7 @@
 REFERENCE_ENGINE = polars_executor.PolarsExecutor()
 
 
-@pytest.mark.parametrize("engine", ["polars", "bq"], indirect=True)
+@pytest.mark.parametrize("engine", ["polars", "bq", "bq-sqlglot"], indirect=True)
 def test_engines_to_array_op(scalars_array_value: array_value.ArrayValue, engine):
     # Bigquery won't allow you to materialize arrays with null, so use non-nullable
     int64_non_null = ops.coalesce_op.as_expr("int64_col", expression.const(0))
@@ -46,7 +46,7 @@ def test_engines_to_array_op(scalars_array_value: array_value.ArrayValue, engine
     assert_equivalence_execution(arr.node, REFERENCE_ENGINE, engine)
 
 
-@pytest.mark.parametrize("engine", ["polars", "bq"], indirect=True)
+@pytest.mark.parametrize("engine", ["polars", "bq", "bq-sqlglot"], indirect=True)
 def test_engines_array_reduce_op(arrays_array_value: array_value.ArrayValue, engine):
     arr, _ = arrays_array_value.compute_values(
         [

tests/unit/core/compile/sqlglot/expressions/snapshots/test_array_ops/test_to_array_op/out.sql

@@ -0,0 +1,26 @@
+WITH `bfcte_0` AS (
+  SELECT
+    `bool_col`,
+    `float64_col`,
+    `int64_col`,
+    `string_col`
+  FROM `bigframes-dev`.`sqlglot_test`.`scalar_types`
+), `bfcte_1` AS (
+  SELECT
+    *,
+    [COALESCE(`bool_col`, FALSE)] AS `bfcol_8`,
+    [COALESCE(`int64_col`, 0)] AS `bfcol_9`,
+    [COALESCE(`string_col`, ''), COALESCE(`string_col`, '')] AS `bfcol_10`,
+    [
+      COALESCE(`int64_col`, 0),
+      CAST(COALESCE(`bool_col`, FALSE) AS INT64),
+      COALESCE(`float64_col`, 0.0)
+    ] AS `bfcol_11`
+  FROM `bfcte_0`
+)
+SELECT
+  `bfcol_8` AS `bool_col`,
+  `bfcol_9` AS `int64_col`,
+  `bfcol_10` AS `strs_col`,
+  `bfcol_11` AS `numeric_col`
+FROM `bfcte_1`

tests/unit/core/compile/sqlglot/expressions/test_array_ops.py

@@ -15,6 +15,7 @@
 import pytest
 
 from bigframes import operations as ops
+from bigframes.core import expression
 from bigframes.operations._op_converters import convert_index, convert_slice
 import bigframes.pandas as bpd
 from bigframes.testing import utils
@@ -60,3 +61,24 @@ def test_array_slice_with_start_and_stop(repeated_types_df: bpd.DataFrame, snaps
     )
 
     snapshot.assert_match(sql, "out.sql")
+
+
+def test_to_array_op(scalar_types_df: bpd.DataFrame, snapshot):
+    bf_df = scalar_types_df[["int64_col", "bool_col", "float64_col", "string_col"]]
+    # Bigquery won't allow you to materialize arrays with null, so use non-nullable
+    int64_non_null = ops.coalesce_op.as_expr("int64_col", expression.const(0))
+    bool_col_non_null = ops.coalesce_op.as_expr("bool_col", expression.const(False))
+    float_col_non_null = ops.coalesce_op.as_expr("float64_col", expression.const(0.0))
+    string_col_non_null = ops.coalesce_op.as_expr("string_col", expression.const(""))
+
+    ops_map = {
+        "bool_col": ops.ToArrayOp().as_expr(bool_col_non_null),
+        "int64_col": ops.ToArrayOp().as_expr(int64_non_null),
+        "strs_col": ops.ToArrayOp().as_expr(string_col_non_null, string_col_non_null),
+        "numeric_col": ops.ToArrayOp().as_expr(
+            int64_non_null, bool_col_non_null, float_col_non_null
+        ),
+    }
+
+    sql = utils._apply_ops_to_sql(bf_df, list(ops_map.values()), list(ops_map.keys()))
+    snapshot.assert_match(sql, "out.sql")