BUG: boundary_slice assumes sorted indexes

Changes `boundary_slice` to handle cases where

- the index is not sorted
- using label-based indexing (loc)
- the start or stop is missing

See pandas-dev/pandas#8613 for details on the pandas
side.

Author: TomAugspurger

dask/dataframe/methods.py

@@ -55,6 +55,12 @@ def boundary_slice(df, start, stop, right_boundary=True, left_boundary=True,
     2  20
     2  30
     """
+    # Pandas treats missing keys differently for label-slicing
+    # on monotonic vs. non-monotonic indexes
+    # If the index is monotonic, `df.loc[start:stop]` is fine.
+    # If it's not, `df.loc[start:stop]` raises when `start` is missing
+    if kind == 'loc' and not df.index.is_monotonic:
+        df = df.sort_index()
     result = getattr(df, kind)[start:stop]
     if not right_boundary:
         right_index = result.index.get_slice_bound(stop, 'left', kind)

dask/dataframe/tests/test_dataframe.py

@@ -2720,3 +2720,16 @@ def test_to_datetime():
 
     assert_eq(pd.to_datetime(s, infer_datetime_format=True),
               dd.to_datetime(ds, infer_datetime_format=True))
+
+
+def test_slice_on_filtered_boundary():
+    # https://github.com/dask/dask/issues/2211
+    x = np.arange(10)
+    x[[5, 6]] -= 2
+    df = pd.DataFrame({"A": x, "B": np.arange(len(x))})
+    pdf = df.set_index("A").query("B > 0")
+    ddf = dd.from_pandas(df, 1).set_index("A").query("B > 0")
+
+    result = dd.concat([ddf, ddf.rename(columns={"B": "C"})], axis=1)
+    expected = pd.concat([pdf, pdf.rename(columns={"B": "C"})], axis=1)
+    assert_eq(result, expected)