For Azure, it looks like the macOS images were updated from 20220124.1 to 20220201.2. I'm not seeing anything specifically about Ghostscript in the changes. Ghostscript seems to be installed via Homebrew and hasn't changed versions there, but possibly some of its dependencies have.

Based on the output from Azure, it appears that Ghostscript is compressing some streams with LZW in binary. This gets somewhat mangled by outputting to StringIO and then trying to encode it to compare against the BytesIO result. I'm not sure we really should be supporting StringIO, or at least disable compression if so.

I think I agree that saving distiller output to text streams does not make sense (as the postscript standard explicitly allows arbitrary binary sequences (section 3.14 "Binary encoding details"), we just don't output them ourselves. I guess the most backcompatible fix is the one you suggest (to disable distillers when outputting to text, perhaps logging a warning in that case).

I generally dislike hard upper pins, but we will want to do a hard-pin on the 3.5.x branch anyway and it sounds like we are not the only people with this issue.

I'm going to merge this over the GS related failures.

I generally dislike hard upper pins, but we will want to do a hard-pin on the 3.5.x branch anyway and it sounds like we are not the only people with this issue.

I'm following the upstream issue so I can test unpin when it's fixed,