PERF: Performance regression with memory_usage(deep=True) on `object` columns

[Lawrr]

Code Sample, a copy-pastable example if possible

import time

import pandas as pd
import numpy as np

df = pd.DataFrame({
    "a": np.empty(10000000),
    "b": np.empty(10000000),
})

df["a"] = df["a"].astype("object")

s = time.time()
mem = df.memory_usage(deep=True)
print("memory_usage(deep=True) took %.4fsecs" % (time.time() - s))

Problem description

Performance of memory_usage(deep=True) on object columns seems to have regressed significantly since v0.23.4. Once in v0.24.0, and another regression in v1.0.0 that remains in v1.0.3.

Output
v1.0.3

memory_usage(deep=True) took 26.4566secs

v0.24.0

memory_usage(deep=True) took 6.0479secs

v0.23.4

memory_usage(deep=True) took 0.4633secs

Removing df["a"] = df["a"].astype("object") reverts it back to the expected magnitude of speed in v1.0.3:

memory_usage(deep=True) took 0.0024secs

Output of pd.show_versions()

INSTALLED VERSIONS

commit : None
python : 3.6.5.final.0
python-bits : 64
OS : Linux
OS-release : 3.16.0-77-generic
machine : x86_64
processor : x86_64
byteorder : little
LC_ALL : None
LANG : en_AU.UTF-8
LOCALE : en_AU.UTF-8

pandas : 1.0.3
numpy : 1.18.2
pytz : 2019.3
dateutil : 2.8.1
pip : 20.0.2
setuptools : 46.1.1
Cython : None
pytest : None
hypothesis : None
sphinx : None
blosc : None
feather : None
xlsxwriter : None
lxml.etree : None
html5lib : None
pymysql : None
psycopg2 : None
jinja2 : None
IPython : None
pandas_datareader: None
bs4 : None
bottleneck : None
fastparquet : None
gcsfs : None
lxml.etree : None
matplotlib : None
numexpr : None
odfpy : None
openpyxl : None
pandas_gbq : None
pyarrow : None
pytables : None
pytest : None
pyxlsb : None
s3fs : None
scipy : None
sqlalchemy : None
tables : None
tabulate : None
xarray : None
xlrd : None
xlwt : None
xlsxwriter : None
numba : None

[TomAugspurger]

Was the previous (0.23.4) output correct?

Removing df["a"] = df["a"].astype("object") reverts it back to the expected magnitude of speed in v1.0.3:

deep=True only has an effect for object dtype. With object dtype we need to inspect each value for its memory usage.

[Lawrr]

The column sizes look the same. Looks like the index size changed in v0.25.0.

v1.0.3

memory_usage(deep=True) took 26.5421secs
Index          128
a        320000000
b         80000000
dtype: int64

v0.25.0

memory_usage(deep=True) took 4.7278secs
Index          128
a        320000000
b         80000000
dtype: int64

v0.24.0

memory_usage(deep=True) took 4.5771secs
Index           80
a        320000000
b         80000000
dtype: int64

v0.23.4

memory_usage(deep=True) took 0.4579secs
Index           80
a        320000000
b         80000000
dtype: int64

[TomAugspurger]

Thanks for posting all the timings. The slowdown from 0.25.0 to 1.0.3 looks concerning. It seems primarily driven by doing lib.memory_usage_of_objects(self.array) in

pandas/pandas/core/base.py

Lines 1388 to 1390 in 28e0f18

	v = self.array.nbytes
	if deep and is_object_dtype(self) and not PYPY:
	v += lib.memory_usage_of_objects(self.array)

That's now a PandasArray, and it involves doing a getitem on every element, which is slower that NumPy's ndarray.

@jorisvandenbossche do you know if we've discussed adding memory_usage to the EA interface? There's a check for it at

pandas/pandas/core/base.py

Lines 1385 to 1386 in 28e0f18

	if hasattr(self.array, "memory_usage"):
	return self.array.memory_usage(deep=deep)

. Ideally PandasArray would be able to do a .to_numpy() on the object passed to memory_usage_of_objects.

[TomAugspurger]

With that diff

diff --git a/pandas/core/arrays/base.py b/pandas/core/arrays/base.py
index 67e3807c47..a8d04fee8d 100644
--- a/pandas/core/arrays/base.py
+++ b/pandas/core/arrays/base.py
@@ -431,6 +431,14 @@ class ExtensionArray:
         # on the number of bytes needed.
         raise AbstractMethodError(self)
 
+    def memory_usage(self, deep=True):
+        from pandas.core.dtypes.common import is_object_dtype
+        from pandas.compat import PYPY
+        v = self.nbytes
+        if deep and is_object_dtype(self.dtype) and not PYPY:
+            v += lib.memory_usage_of_objects(self.to_numpy())
+        return v
+
     # ------------------------------------------------------------------------
     # Additional Methods
     # ------------------------------------------------------------------------

In [2]: %timeit df.memory_usage(deep=True)
89.4 ms ± 1.05 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

vs master

In [2]: %timeit df.memory_usage(deep=True)
2.56 s ± 114 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

[jorisvandenbossche]

I think the Series.memory_usage implementation should simply use self._values instead of self.array? ExtensionArrays can never be object dtype, so never take that path.

[jorisvandenbossche]

(I am actually surprised that memory_usage_of_objects works with passing a PandasArray, I would have though cython to raise an error when not getting an ndarray/memoryview)

[TomAugspurger]

I'm also surprised that worked.

ExtensionArrays can never be object dtype, so never take that path.

This is probably safe to assume... So then this requires Updating

pandas/pandas/core/base.py

Line 1390 in e7ee418

v += lib.memory_usage_of_objects(self.array)

to use self._values rather than self._array and then adding an ASV, probably in asv_bench/benchmarks/frame_methods.py that has a DataFrame with object-dtype columns.

[neilkg]

take