refactor: add agg_ops.MedianOp compiler to sqlglot

Author: chelsea-lin

bigframes/core/compile/sqlglot/aggregations/unary_compiler.py

@@ -56,6 +56,21 @@ def _(
     return apply_window_if_present(sge.func("MAX", column.expr), window)
 
 
+@UNARY_OP_REGISTRATION.register(agg_ops.MedianOp)
+def _(
+    op: agg_ops.MedianOp,
+    column: typed_expr.TypedExpr,
+    window: typing.Optional[window_spec.WindowSpec] = None,
+) -> sge.Expression:
+    # TODO(swast): Allow switching between exact and approximate median.
+    # For now, the best we can do is an approximate median when we're doing
+    # an aggregation, as PERCENTILE_CONT is only an analytic function.
+    approx_quantiles = sge.func("APPROX_QUANTILES", column.expr, sge.convert(2))
+    return sge.Bracket(
+        this=approx_quantiles, expressions=[sge.func("OFFSET", sge.convert(1))]
+    )
+
+
 @UNARY_OP_REGISTRATION.register(agg_ops.MinOp)
 def _(
     op: agg_ops.MinOp,

tests/system/small/engines/test_aggregation.py

@@ -12,11 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from google.cloud import bigquery
 import pytest
 
 from bigframes.core import agg_expressions, array_value, expression, identifiers, nodes
 import bigframes.operations.aggregations as agg_ops
-from bigframes.session import polars_executor
+from bigframes.session import direct_gbq_execution, polars_executor
 from bigframes.testing.engine_utils import assert_equivalence_execution
 
 pytest.importorskip("polars")
@@ -84,6 +85,24 @@ def test_engines_unary_aggregates(
     assert_equivalence_execution(node, REFERENCE_ENGINE, engine)
 
 
+@pytest.mark.parametrize(
+    "op",
+    [agg_ops.MedianOp],
+)
+def test_sql_engines_unary_aggregates(
+    scalars_array_value: array_value.ArrayValue,
+    bigquery_client: bigquery.Client,
+    op,
+):
+    # TODO: this is not working??
+    node = apply_agg_to_all_valid(scalars_array_value, op).node
+    left_engine = direct_gbq_execution.DirectGbqExecutor(bigquery_client)
+    right_engine = direct_gbq_execution.DirectGbqExecutor(
+        bigquery_client, compiler="sqlglot"
+    )
+    assert_equivalence_execution(node, left_engine, right_engine)
+
+
 @pytest.mark.parametrize("engine", ["polars", "bq", "bq-sqlglot"], indirect=True)
 @pytest.mark.parametrize(
     "grouping_cols",

tests/unit/core/compile/sqlglot/aggregations/snapshots/test_unary_compiler/test_median/out.sql

@@ -0,0 +1,12 @@
+WITH `bfcte_0` AS (
+  SELECT
+    `int64_col` AS `bfcol_0`
+  FROM `bigframes-dev`.`sqlglot_test`.`scalar_types`
+), `bfcte_1` AS (
+  SELECT
+    APPROX_QUANTILES(`bfcol_0`, 2)[OFFSET(1)] AS `bfcol_1`
+  FROM `bfcte_0`
+)
+SELECT
+  `bfcol_1` AS `int64_col`
+FROM `bfcte_1`

tests/unit/core/compile/sqlglot/aggregations/test_unary_compiler.py

@@ -56,6 +56,15 @@ def test_max(scalar_types_df: bpd.DataFrame, snapshot):
     snapshot.assert_match(sql, "out.sql")
 
 
+def test_median(scalar_types_df: bpd.DataFrame, snapshot):
+    col_name = "int64_col"
+    bf_df = scalar_types_df[[col_name]]
+    agg_expr = agg_ops.MedianOp().as_expr(col_name)
+    sql = _apply_unary_agg_ops(bf_df, [agg_expr], [col_name])
+
+    snapshot.assert_match(sql, "out.sql")
+
+
 def test_min(scalar_types_df: bpd.DataFrame, snapshot):
     col_name = "int64_col"
     bf_df = scalar_types_df[[col_name]]