Enable caching for memory_usage calculation in Column (NVIDIA#10345)

We previously used to cache the `Column.memory_usage` output in `Column._cached_sizeof` but it probably was missed to be included in the recent refactors. This PR re-enables caching of `memory_usage`.

`Column.memory_usage` should be a no-op on consecutive calls and **60% faster**:
```python
In [1]: import cudf

In [2]: s = cudf.Series(["abc ", " d e", None, "10 11 234355"] * 10000000)

In [3]: s
Out[3]: 
0                   abc 
1                    d e
2                   <NA>
3           10 11 234355
4                   abc 
                ...     
39999995    10 11 234355
39999996            abc 
39999997             d e
39999998            <NA>
39999999    10 11 234355
Length: 40000000, dtype: object



# branch-22.04
In [3]: %timeit s.memory_usage()
2.86 µs ± 53.4 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

# THIS PR

In [3]: %timeit s.memory_usage()
1.77 µs ± 10.8 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
```

Authors:
  - GALI PREM SAGAR (https://github.com/galipremsagar)

Approvers:
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: rapidsai/cudf#10345

python/cudf/cudf/_lib/column.pyx

@@ -74,7 +74,6 @@ cdef class Column:
     ):
 
         self._size = size
-        self._cached_sizeof = None
         self._distinct_count = {}
         self._dtype = dtype
         self._offset = offset
@@ -204,7 +203,11 @@ cdef class Column:
 
     def _clear_cache(self):
         self._distinct_count = {}
-        self._cached_sizeof = None
+        try:
+            del self.memory_usage
+        except AttributeError:
+            # `self.memory_usage` was never called before, So ignore.
+            pass
         self._null_count = None
 
     def set_mask(self, value):

python/cudf/cudf/core/column/categorical.py

@@ -4,6 +4,7 @@
 
 import pickle
 from collections.abc import MutableSequence
+from functools import cached_property
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -1335,8 +1336,9 @@ def copy(self, deep: bool = True) -> CategoricalColumn:
                 size=self.size,
             )
 
+    @cached_property
     def memory_usage(self) -> int:
-        return self.categories.memory_usage() + self.codes.memory_usage()
+        return self.categories.memory_usage + self.codes.memory_usage
 
     def _mimic_inplace(
         self, other_col: ColumnBase, inplace: bool = False

python/cudf/cudf/core/column/column.py

@@ -5,6 +5,7 @@
 import builtins
 import pickle
 import warnings
+from functools import cached_property
 from types import SimpleNamespace
 from typing import (
     Any,
@@ -297,6 +298,7 @@ def _get_mask_as_column(self) -> ColumnBase:
             self.base_mask, self.offset, self.offset + len(self)
         )
 
+    @cached_property
     def memory_usage(self) -> int:
         n = 0
         if self.data is not None:

python/cudf/cudf/core/column/lists.py

@@ -1,6 +1,7 @@
-# Copyright (c) 2020-2021, NVIDIA CORPORATION.
+# Copyright (c) 2020-2022, NVIDIA CORPORATION.
 
 import pickle
+from functools import cached_property
 from typing import List, Sequence
 
 import numpy as np
@@ -42,6 +43,7 @@ def __init__(
             children=children,
         )
 
+    @cached_property
     def memory_usage(self):
         n = 0
         if self.nullable:

python/cudf/cudf/core/column/string.py

@@ -6,6 +6,7 @@
 import pickle
 import re
 import warnings
+from functools import cached_property
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -5024,7 +5025,6 @@ class StringColumn(column.ColumnBase):
 
     _start_offset: Optional[int]
     _end_offset: Optional[int]
-    _cached_sizeof: Optional[int]
 
     def __init__(
         self,
@@ -5102,6 +5102,7 @@ def end_offset(self) -> int:
 
         return self._end_offset
 
+    @cached_property
     def memory_usage(self) -> int:
         n = 0
         if len(self.base_children) == 2:

python/cudf/cudf/core/frame.py

@@ -360,7 +360,7 @@ def memory_usage(self, deep=False):
                 "The deep parameter is ignored and is only included "
                 "for pandas compatibility."
             )
-        return {name: col.memory_usage() for name, col in self._data.items()}
+        return {name: col.memory_usage for name, col in self._data.items()}
 
     def __len__(self):
         return self._num_rows

python/cudf/cudf/core/multiindex.py

@@ -1415,7 +1415,7 @@ def memory_usage(self, deep=False):
                 usage += level.memory_usage(deep=deep)
         if self.codes:
             for col in self.codes._data.columns:
-                usage += col.memory_usage()
+                usage += col.memory_usage
         return usage
 
     def difference(self, other, sort=None):

python/cudf/cudf/tests/test_dataframe.py

@@ -5195,8 +5195,8 @@ def test_memory_usage_cat():
     gdf = cudf.from_pandas(df)
 
     expected = (
-        gdf.B._column.categories.memory_usage()
-        + gdf.B._column.codes.memory_usage()
+        gdf.B._column.categories.memory_usage
+        + gdf.B._column.codes.memory_usage
     )
 
     # Check cat column
@@ -5209,8 +5209,7 @@ def test_memory_usage_cat():
 def test_memory_usage_list():
     df = cudf.DataFrame({"A": [[0, 1, 2, 3], [4, 5, 6], [7, 8], [9]]})
     expected = (
-        df.A._column.offsets.memory_usage()
-        + df.A._column.elements.memory_usage()
+        df.A._column.offsets.memory_usage + df.A._column.elements.memory_usage
     )
     assert expected == df.A.memory_usage()