Fix compute_meta recursion in blockwise (dask#5048)

dask/dataframe/core.py

@@ -4759,7 +4759,7 @@ def map_partitions(func, *args, **kwargs):
         }
         graph = HighLevelGraph.from_collections(name, layer, dependencies=args)
         return Scalar(graph, name, meta)
-    elif not (has_parallel_type(meta) or is_arraylike(meta)):
+    elif not (has_parallel_type(meta) or is_arraylike(meta) and meta.shape):
         # If `meta` is not a pandas object, the concatenated results will be a
         # different type
         meta = make_meta(_concat([meta]), index=meta_index)
@@ -5841,7 +5841,7 @@ def new_dd_object(dsk, name, meta, divisions):
     """
     if has_parallel_type(meta):
         return get_parallel_type(meta)(dsk, name, meta, divisions)
-    elif is_arraylike(meta):
+    elif is_arraylike(meta) and meta.shape:
         import dask.array as da
 
         chunks = ((np.nan,) * (len(divisions) - 1),) + tuple(

dask/dataframe/utils.py

@@ -317,7 +317,7 @@ def make_meta_object(x, index=None):
     """
     if hasattr(x, "_meta"):
         return x._meta
-    elif is_arraylike(x):
+    elif is_arraylike(x) and x.shape:
         return x[:0]
 
     if index is not None:

dask/tests/test_utils.py

@@ -27,6 +27,7 @@
     derived_from,
     parse_timedelta,
     parse_bytes,
+    is_arraylike,
 )
 from dask.utils_test import inc
 from dask.highlevelgraph import HighLevelGraph
@@ -506,3 +507,15 @@ def test_parse_timedelta():
     assert parse_timedelta("1", default="seconds") == 1
     assert parse_timedelta("1", default="ms") == 0.001
     assert parse_timedelta(1, default="ms") == 0.001
+
+
+def test_is_arraylike():
+    assert is_arraylike(0) is False
+    assert is_arraylike(()) is False
+    assert is_arraylike((0)) is False
+    assert is_arraylike([]) is False
+    assert is_arraylike([0]) is False
+
+    assert is_arraylike(np.empty(())) is True
+    assert is_arraylike(np.empty((0,))) is True
+    assert is_arraylike(np.empty((0, 0))) is True

dask/utils.py

@@ -1051,8 +1051,9 @@ def is_arraylike(x):
     Examples
     --------
     >>> import numpy as np
-    >>> x = np.ones(5)
-    >>> is_arraylike(x)
+    >>> is_arraylike(np.ones(5))
+    True
+    >>> is_arraylike(np.ones(()))
     True
     >>> is_arraylike(5)
     False
@@ -1061,10 +1062,9 @@ def is_arraylike(x):
     """
     from .base import is_dask_collection
 
-    return (
+    return bool(
         hasattr(x, "shape")
         and isinstance(x.shape, tuple)
-        and x.shape
         and hasattr(x, "dtype")
         and not any(is_dask_collection(n) for n in x.shape)
     )