The Pillow functions Image.open() and Image.save(), among others, use os.path.realpath() on the input path to transform it to a symlink-free canonical path.

• Pillow/src/PIL/Image.py

  Line 2553 in d642d2c

  filename = os.path.realpath(os.fspath(fp))

• Pillow/src/PIL/Image.py

  Line 2562 in d642d2c

  filename = os.path.realpath(os.fspath(fp.name))

• Pillow/src/PIL/Image.py

  Line 3466 in d642d2c

  filename = os.path.realpath(os.fspath(fp))

There is no reliance on this path's "realness" afterwards that is obvious to me, and realpath() requires an lstat() system call on every single path segment from / on down to the file in question (or more, if symlinks are detected along the way). On clusters with distributed filesystems such as BeeGFS, this leads and is seen as a multiplication (by a factor greater than or equal to the depth of the image file within the hierarchy) of the number of metadata I/O operations, and causes DDoS of the metadata servers for those filesystems.

Pillow is heavily utilized in AI workloads with image datasets and data-loaders, a popular usecase, making Pillow's use of realpath() especially painful. PyTorch's ImageFolder is the classic example.

Please reconsider all uses of realpath() in Pillow. Preferentially remove it entirely. In the alternative that you absolutely need the last segment of the path to be translated to acquire a "true" file-extension, consider readlink().