4 ms·

I'd super appreciate if you have sources (mail archive links or whatever), I'd love to read through Python history stuff like this and the PEP you were referrin

I'd super appreciate if you have sources (mail archive links or whatever), I'd love to read through Python history stuff like this and the PEP you were referring to? (read through the release notes and couldn't find anything, maybe it relates to PEP 538 in Python 3.7?) It's pretty clear that there were a lot of mistakes in the Python 2 to 3 transition, I'm certainly not trying to defend every choices of the developers.

Jasper_ 5y ago
The PEP I was thinking of was actually PEP-393 [0], which was done for Python 3.3 (well, I did say I mostly left the community, and that my dates were probably wrong). It seems it didn't include UTF-8 as a representation; only Latin-1 and UCS-2, and it seems like it only picks a representation if it can maintain the O(1) indexing property. So I was wrong on that one, I guess that's the one thing they haven't given up on :/ As for the rest, I'll see if I can dig up some logs, but a lot of the discussion I was involved in took place on IRC, and my logs were lost a few server wipes ago. You might find some old posts of mine on python-ideas. I've been planning on making a fuller blog post on why I think Python's text model is the wrong one, but Manish has a very good post [1] which covers most of the reason. Basically, Unicode code points don't net you anything over bytes other than O(1) indexing, which is completely useless, and in fact, is in many ways a worse representation. There is no operation where having a list of Unicode code points helps you more than having a list of bytes; well, one that isn't completely broken in practice. More drastically, the team wanting "Unicode everywhere" meant forcing things that clearly weren't Unicode into Unicode. File paths have not been valid Unicode on any popular system. Windows uses its "UTF-16" which is really more UCS-2 and allows unpaired surrogates, and POSIX says node names are bytes but cannot use ASCII NUL or /. These limitations were hit days into the prototyping process on real-world use cases, which is why there's an algorithm to shove arbitrary binary data into unpaired surrogate characters (see PEP-383 [2]), to be used on file paths. This should have blew a huge hole in the whole scheme, but they pushed onwards. The utopian future of "Unicode everywhere" was more important than dealing with the practical reality of existing systems. Similar things happened with console output, where it's now assumed that LANG will be set correctly. But no, when I ssh into a server in Japan that's set to use EUC-JP, Python will output EUC-JP-encoded bytes, which gets shovelled over ssh as bytes, which the terminal emulator on my laptop misinterprets as UTF-8. Yes, I know ssh is supposed to tunnel certain LANG envvars as well. No, I don't remember why it didn't in this case. But a lot of people were hitting this, that in Python 3.7 they finally added UTF-8 mode (PEP-450 [3]). There's still a lot of people for whom this general "Unicode console" idea breaks [4] [5]. Pretty much every part of the standard library was riddled with bugs when they rolled this out (as mentioned, I hit bugs in zipfile, others found bugs in email/mime). For a long time, the only way to know if handed a file-like object is going to be in bytes or not was isinstance(file.read(0), bytes), which is generally ugly, so a lot of modules didn't support bytes even when they should have. Practically everything after has been walking back this text model in favor of something where bytes and str are not even that different anymore. Python 3.5 brought with it PEP-461 [5], which added % formatting back to bytes (after a very long and very tiring discussion with the core maintainers [6]). I could go on and find more PEPs, but I'm done for now. [0] https://www.python.org/dev/peps/pep-0393/ https://www.python.org/dev/peps/pep-0393/ [1] https://manishearth.github.io/blog/2017/01/14/stop-ascribing-meaning-to-unicode-code-points/ https://manishearth.github.io/blog/2017/01/14/stop-ascribing... [2] https://www.python.org/dev/peps/pep-0383/ https://www.python.org/dev/peps/pep-0383/ [3] https://www.python.org/dev/peps/pep-0540/ https://www.python.org/dev/peps/pep-0540/ [4] On POSIX system, https://stackoverflow.com/questions/11741574/how-to-print-utf-8-encoded-text-to-the-console-in-python-3 https://stackoverflow.com/questions/11741574/how-to-print-ut... [5] On Windows, https://stackoverflow.com/questions/17918746/print-unicode-string-to-console-ok-but-fails-when-redirect-to-a-file-how-to-fix https://stackoverflow.com/questions/17918746/print-unicode-s... [5] https://www.python.org/dev/peps/pep-0461/ https://www.python.org/dev/peps/pep-0461/ [6] https://bugs.python.org/issue3982 https://bugs.python.org/issue3982

Jasper_ 5y ago

The PEP I was thinking of was actually PEP-393 [0], which was done for Python 3.3 (well, I did say I mostly left the community, and that my dates were probably wrong). It seems it didn't include UTF-8 as a representation; only Latin-1 and UCS-2, and it seems like it only picks a representation if it can maintain the O(1) indexing property. So I was wrong on that one, I guess that's the one thing they haven't given up on :/ As for the rest, I'll see if I can dig up some logs, but a lot of the discussion I was involved in took place on IRC, and my logs were lost a few server wipes ago. You might find some old posts of mine on python-ideas. I've been planning on making a fuller blog post on why I think Python's text model is the wrong one, but Manish has a very good post [1] which covers most of the reason. Basically, Unicode code points don't net you anything over bytes other than O(1) indexing, which is completely useless, and in fact, is in many ways a worse representation. There is no operation where having a list of Unicode code points helps you more than having a list of bytes; well, one that isn't completely broken in practice. More drastically, the team wanting "Unicode everywhere" meant forcing things that clearly weren't Unicode into Unicode. File paths have not been valid Unicode on any popular system. Windows uses its "UTF-16" which is really more UCS-2 and allows unpaired surrogates, and POSIX says node names are bytes but cannot use ASCII NUL or /. These limitations were hit days into the prototyping process on real-world use cases, which is why there's an algorithm to shove arbitrary binary data into unpaired surrogate characters (see PEP-383 [2]), to be used on file paths. This should have blew a huge hole in the whole scheme, but they pushed onwards. The utopian future of "Unicode everywhere" was more important than dealing with the practical reality of existing systems. Similar things happened with console output, where it's now assumed that LANG will be set correctly. But no, when I ssh into a server in Japan that's set to use EUC-JP, Python will output EUC-JP-encoded bytes, which gets shovelled over ssh as bytes, which the terminal emulator on my laptop misinterprets as UTF-8. Yes, I know ssh is supposed to tunnel certain LANG envvars as well. No, I don't remember why it didn't in this case. But a lot of people were hitting this, that in Python 3.7 they finally added UTF-8 mode (PEP-450 [3]). There's still a lot of people for whom this general "Unicode console" idea breaks [4] [5]. Pretty much every part of the standard library was riddled with bugs when they rolled this out (as mentioned, I hit bugs in zipfile, others found bugs in email/mime). For a long time, the only way to know if handed a file-like object is going to be in bytes or not was isinstance(file.read(0), bytes), which is generally ugly, so a lot of modules didn't support bytes even when they should have. Practically everything after has been walking back this text model in favor of something where bytes and str are not even that different anymore. Python 3.5 brought with it PEP-461 [5], which added % formatting back to bytes (after a very long and very tiring discussion with the core maintainers [6]). I could go on and find more PEPs, but I'm done for now. [0] https://www.python.org/dev/peps/pep-0393/ https://www.python.org/dev/peps/pep-0393/ [1] https://manishearth.github.io/blog/2017/01/14/stop-ascribing-meaning-to-unicode-code-points/ https://manishearth.github.io/blog/2017/01/14/stop-ascribing... [2] https://www.python.org/dev/peps/pep-0383/ https://www.python.org/dev/peps/pep-0383/ [3] https://www.python.org/dev/peps/pep-0540/ https://www.python.org/dev/peps/pep-0540/ [4] On POSIX system, https://stackoverflow.com/questions/11741574/how-to-print-utf-8-encoded-text-to-the-console-in-python-3 https://stackoverflow.com/questions/11741574/how-to-print-ut... [5] On Windows, https://stackoverflow.com/questions/17918746/print-unicode-string-to-console-ok-but-fails-when-redirect-to-a-file-how-to-fix https://stackoverflow.com/questions/17918746/print-unicode-s... [5] https://www.python.org/dev/peps/pep-0461/ https://www.python.org/dev/peps/pep-0461/ [6] https://bugs.python.org/issue3982 https://bugs.python.org/issue3982