memory leak when loading images in webp format with Pillow 5.0.0

[Dobiasd]

With Pillow 4.3.0 the following code works fine with constant memory usage.

# download example image here:
# http://www.gstatic.com/webp/gallery/1.webp
from PIL import Image
while True:
    img = Image.open('1.webp') # no problem with png
    img = img.convert('L') # just a dummy operation on the image
    print(img.size)

When using Pillow 5.0.0 however, the memory usage of the process constantly grows quite fast.

I'm using Python 3.5.2 and Ubuntu 16.04.

[wiredfool]

I'm not seeing that here, running under valgrind --tool=massif:

I've tried this with and without the convert line, and it looks like a steady state after the initial ramp.

from PIL import Image
for _ in range(100):
  im = Image.open('issue_2972.webp')
  im.load()
  im.convert('L')
  print (im.size)

(note this was 2.7, but the 3.5 graph is similar, with a bit more junk in the initial ramp up.)

Are you installing from source or from a wheel?

[Dobiasd]

I'm using the wheels from pip3.

Right now I'm at a Windows 7 machine with Python 3.6.4, and the problem also occurs here.

#issue_2972.py
from PIL import Image
for _ in range(1000): # 1000 instead of 100 to have an even bigger effect
  im = Image.open('issue_2972.webp') # http://www.gstatic.com/webp/gallery/1.webp
  im.load()
  im.convert('L')
  print (im.size)

When running pip3 install pillow==4.3.0 and then python issue_2972.py the process is constantly at about 8 MB in the task manager.

But with pip3 install pillow==5.0.0 and python issue_2972.py it reaches over 700 MB at the end.

[wiredfool]

Ok, with the binary wheels I get a slightly different view:

[wiredfool]

Ok, now with patch:

[Dobiasd]

Wow, very nice analysis and a lightning fast fix.
10/10, would open issue here again ;-)

[wiredfool]

Don't take this the wrong way, but just because I'm good at locating memory leaks doesn't mean that I want to be finding more of them here.