Inconsistent behavior opening truncated JPEG file

[elbaro]

What did you do?

from PIL import Image
import numpy as np
a=Image.open('aa.jpg')
b1=np.array(a)
b2=np.array(a)

What did you expect to happen?

b1.shape == b2.shape == (H,W,3)

What actually happened?

Surprisingly, b1.shape == () and b2.shape == (H, W, 3).

b1:
array(<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=2000x2000 at 0x7FCD...48>, dtype=object)

I have a large RAM available (>50G).
This is consistently reproduced for a specific image file:
https://farm5.staticflickr.com/4360/36444958735_32c0410b64_o.jpg

What versions of Pillow and Python are you using?

5.1.0.dev0 && 4.3.0 (pip version)
Python 3.6.4

[elbaro]

Many viewers can see it but viewnior failed to open it.
If there is something wrong with the image, I expect an explicit raise or consistent behavior.

[wiredfool]

We do throw an error on the first try, but apparently it's silently caught and discarded in numpy. Though, it is a bit of a puzzle as to why it succeeds the second time after the initial failure due to a truncated file.

>>> im = Image.open('36444958735_32c0410b64_o.jpg')
>>> im.__array_interface__.keys()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/erics/vpy35/lib/python3.5/site-packages/Pillow-5.0.0-py3.5-linux-x86_64.egg/PIL/Image.py", line 673, in __array_interface__
  File "/home/erics/vpy35/lib/python3.5/site-packages/Pillow-5.0.0-py3.5-linux-x86_64.egg/PIL/Image.py", line 720, in tobytes
  File "/home/erics/vpy35/lib/python3.5/site-packages/Pillow-5.0.0-py3.5-linux-x86_64.egg/PIL/ImageFile.py", line 233, in load
OSError: image file is truncated (13 bytes not processed)
>>> im.__array_interface__.keys()
dict_keys(['version', 'shape', 'data', 'typestr'])
>>> 

This is consistent behavior with two load calls, so it's something in the ImageFile code handling truncated files where it's not cleaning up the failure properly, and allowing the attempt to do it again.

[radarhere]

With #3965, the behaviour is now consistent - the shape is empty for both calls.

[radarhere]

You may also be interested to know that if you use LOAD_TRUNCATED_IMAGES, the dimensions are returned -

from PIL import Image, ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
import numpy as np
a=Image.open('aa.jpg')
b1=np.array(a)
print(b1.shape)  # (2000, 2000, 3)
b2=np.array(a)
print(b2.shape)  # (2000, 2000, 3)

[radarhere]

The fact that the errors are silenced is actually acknowledged as a problem with numpy - numpy/numpy#11629. So I'm going to close this.