Clean up percentiles_summary logic (#11094)

dask/dataframe/partitionquantiles.py

@@ -78,8 +78,10 @@
 from tlz import merge, merge_sorted, take
 
 from dask.base import tokenize
+from dask.dataframe._compat import PANDAS_GE_150
 from dask.dataframe.core import Series
 from dask.dataframe.dispatch import tolist_dispatch
+from dask.dataframe.utils import is_series_like
 from dask.utils import is_cupy_type, random_state_data
 
 
@@ -430,16 +432,13 @@ def percentiles_summary(df, num_old, num_new, upsample, state):
     elif is_datetime64_dtype(data.dtype) or is_integer_dtype(data.dtype):
         interpolation = "nearest"
 
-    # FIXME: pandas quantile doesn't work with some data types (e.g. strings).
-    # We fall back to an ndarray as a workaround.
     try:
-        vals = data.quantile(q=qs / 100, interpolation=interpolation).values
+        # Plan A: Use `Series.quantile`
+        vals = data.quantile(q=qs / 100, interpolation=interpolation)
     except (TypeError, NotImplementedError):
-        try:
-            vals, _ = _percentile(array_safe(data, like=data.values), qs, interpolation)
-        except (TypeError, NotImplementedError):
-            # `data.values` doesn't work for cudf, so we need to
-            # use `quantile(..., method="table")` as a fallback
+        # Series.quantile doesn't work with some data types (e.g. strings)
+        if PANDAS_GE_150:
+            # Fall back to DataFrame.quantile with "nearest" interpolation
             interpolation = "nearest"
             vals = (
                 data.to_frame()
@@ -451,6 +450,18 @@ def percentiles_summary(df, num_old, num_new, upsample, state):
                 )
                 .iloc[:, 0]
             )
+        else:
+            # Fall back to ndarray (Not supported for cuDF)
+            vals, _ = _percentile(array_safe(data, like=data.values), qs, interpolation)
+
+    # Convert to array if necessary (and possible)
+    if is_series_like(vals):
+        try:
+            vals = vals.values
+        except (ValueError, TypeError):
+            # cudf->cupy won't work if nulls are present,
+            # or if this is a string dtype
+            pass
 
     if (
         is_cupy_type(data)