Bug in Image.get_rendition() overwrites thumbnails of other images with similar prefix

We’ve noticed that wagtailimages overwrites thumbnails of previous image versions, when they have a similar prefix. Depending on the filter settings, this may happen after 17 similar characters.

You can test this by running the Image.get_rendition() function with an image that has a similar prefix. In the debugger, we found that the following values are constructed in AbstractImage.get_rendition():

spec = 'fill-720x720-c100|format-jpeg'
input_filename = 'our-story-verhaal-3.jpg'
cache_key = '2e16d0ba'   # the result of sha1('None-None-None-None').hexdigest()[:8]

then the magic bug happens at https://github.com/torchbox/wagtail/blob/2c9805dc3fa37e9bf27fc7593de4465673faffca/wagtail/wagtailimages/models.py#L301:

output_filename_without_extension = input_filename_without_extension[:(59 - len(output_extension))]

Which produces:

output_filename_without_extension = 'our-story-verhaal'   # <-- SILENT TRUNCATION!!
output_filename = 'our-story-verhaal.2e16d0ba.fill-720x720-c100.format-jpeg.jpg'

Which means that any previous image that starts with the same 17 characters (“our-story-verhaal”) is overwritten with thumbnails of the next image. In our case for example, the tester uploaded 3 images:

• our-story-verhaal-1.jpg
• our-story-verhaal-2.jpg
• our-story-verhaal-3.jpg

In the end, all images had the same thumbnail.

[gasman]

Sounds good to me. While I personally think it would be nice for django-storages to not blatantly violate the Django storages spec… hashing on file content would be a quick and effective way to solve this and several other problems in one go (including the far-future caching).

I think it would make sense to drop the filter-spec bit of the filename at the same time - that was originally put in place in the hopes that we could keep the filenames somewhat meaningful, but that’s a fairly marginal benefit and probably incompatible with far-future expiry.

[gasman]

See #4502 - S3BotoStorage’s failure to handle duplicate filenames also breaks documents (and so any workaround we apply should probably also be applied to documents).