Skip to content

A reactive notebook for Python — run reproducible experiments, query with SQL, execute as a script, deploy as an app, and version with git. Stored as pure Python. All in a modern, AI-native editor.

tests: fix flakey dataframe tests (#7738)

Myles Scolnickebd663412b83
diff --git a/tests/_plugins/ui/_impl/dataframes/test_handlers.py b/tests/_plugins/ui/_impl/dataframes/test_handlers.py
index b9090957e7..0a696f749c 100644
--- a/tests/_plugins/ui/_impl/dataframes/test_handlers.py
+++ b/tests/_plugins/ui/_impl/dataframes/test_handlers.py
@@ -82,9 +82,18 @@ def assert_frame_equal(a: DataFrameType, b: DataFrameType) -> None:
assert nw_a.to_dict(as_series=False) == nw_b.to_dict(as_series=False)
-def assert_frame_equal_with_nans(a: DataFrameType, b: DataFrameType) -> None:
+def assert_frame_equal_with_nans(
+ a: DataFrameType, b: DataFrameType, allow_nan_equals_zero: bool = False
+) -> None:
"""
Assert two dataframes are equal, treating NaNs in the same locations as equal.
+
+ Args:
+ a: First dataframe
+ b: Second dataframe
+ allow_nan_equals_zero: If True, treat NaN and 0.0 as equivalent values.
+ This is useful for pivot operations where missing aggregations may
+ be filled with 0.0 or NaN depending on the backend.
"""
import math
@@ -109,7 +118,21 @@ def assert_frame_equal_with_nans(a: DataFrameType, b: DataFrameType) -> None:
and math.isnan(val_a)
and math.isnan(val_b)
)
- if not (val_a == val_b or both_nan):
+ # For pivot operations, treat NaN and 0.0 as equivalent
+ nan_or_zero_match = (
+ allow_nan_equals_zero
+ and isinstance(val_a, (float, int))
+ and isinstance(val_b, (float, int))
+ and (
+ (math.isnan(val_a) if isinstance(val_a, float) else False)
+ or val_a == 0.0
+ )
+ and (
+ (math.isnan(val_b) if isinstance(val_b, float) else False)
+ or val_b == 0.0
+ )
+ )
+ if not (val_a == val_b or both_nan or nan_or_zero_match):
raise AssertionError(
f"DataFrame values differ at column '{col}', row {idx}: {val_a} != {val_b}"
)
@@ -1385,7 +1408,11 @@ class TestTransformHandler:
df: DataFrameType, expected: DataFrameType, transform: PivotTransform
) -> None:
result = apply(df, transform)
- assert_frame_equal_with_nans(result, expected)
+ # Allow NaN and 0.0 to be treated as equivalent for pivot operations
+ # since different backends may fill missing aggregations differently
+ assert_frame_equal_with_nans(
+ result, expected, allow_nan_equals_zero=True
+ )
@staticmethod
@pytest.mark.parametrize(
diff --git a/tests/_plugins/ui/_impl/dataframes/test_print_code.py b/tests/_plugins/ui/_impl/dataframes/test_print_code.py
index 307dcd58e2..848b9b36f7 100644
--- a/tests/_plugins/ui/_impl/dataframes/test_print_code.py
+++ b/tests/_plugins/ui/_impl/dataframes/test_print_code.py
@@ -572,6 +572,14 @@ def test_print_code_result_matches_actual_transform_polars(
# TODO: unimplemented
if transform.type == TransformType.AGGREGATE:
assume(False)
+ # Don't group by unhashable types (lists, dicts) as ordering is non-deterministic
+ if transform.type == TransformType.GROUP_BY:
+ assume(
+ not any(
+ column_id in {"lists", "dicts"}
+ for column_id in transform.column_ids
+ )
+ )
if transform.type == TransformType.PIVOT:
assume(
(
@@ -681,8 +689,9 @@ def test_print_code_result_matches_actual_transform_polars(
# For group_by transforms, the row order might differ even with maintain_order=True
# Sort both dataframes by all columns before comparing
+ # Note: We exclude grouping by unhashable types (lists, dicts) via assume() above,
+ # so all columns should be sortable
if transform.type == TransformType.GROUP_BY:
- # Filter out columns with unhashable types (dicts, lists)
import polars.datatypes as pl_dtypes
sortable_cols = [