written on Thursday, January 7, 2010
I was briefly expressing my disagreement with the Python 3 development decisions, so I want to elaborate on that a bit. While I was previously addressing some of the problems I have with Python 3, I took the time to create a list of things that were solved in a way other than I expected or hoped.
Let's start with the biggest grief of mine...
If you look at any of the pocoo libraries, they all use unicode. In fact, Jinja2 and Werkzeug even enforce unicode so you can't even use them with byte strings internally unless you do the encode dance. Why that? Because I believe in unicode and that is not too surprising because German, the official language of Austria uses some non ASCII letters and yet tons of systems deployed force me to substitute Umlauts with latin letters just because someone had a limited horizon.
But the unicode world is complex and Python does not care about unicode too much. And that neither in Python 2 nor Python 3. So what does not work about unicode in Python? So in German for example, there are words like "Fuß" (which means foot). The last letter there is a so-called "Scharfes Es" or "Eszett". The former means "sharp-s" and is usually represented as a "ſs" ligature, the other one is a "ſz" ligature and used in blackletter fonts (The letter "ſ" is no longer in use but worked similar to a latin "s"). Another common way was using an "ſs" litature. Because this letter will never occur at the beginning of words there was never an uppercase character for it (there was one introduced lately, but nobody uses it). However, it is pretty common to use title case or uppercase for emphasis, so there is a need for that letter to exist in uppercase. The common replacement you see is a doubled "s" or "sz". So "Fuß" becomes "FUSS", "Maße" became "MASZE" etc. There are some variations, but basically it means that one letter becomes two.
However, that does not work in Python. The Python unicode implementation cannot do two things: neither can it replace one letter with two when changing the case, nor does it allow a locale information for character mappings. The latter is necessary for languages like Turkish where an uppercase "I" is lowercased to "ı" and not "i". (And I will not complain about the shared state of the locale library which of course stayed in Python 3).
Another problem is that Python uses UCS2 or UCS4 internally and that shines through. So if you have an UCS2 build, len() called on strings does not give you the number of characters in the string, but the number of UCS2 interpreted characters which might not be the same. In fact, every letter outside the basic plane will be wrongly counted. Because UTF-16 (UCS2 with support for surrogate pairs that allow you to use characters outside the basic plane) is a variable length encoding it has the same problems as utf-8 as an internal encoding: namely, making slicing a non trivial operation. Another problem arises: binary extensions have to be compiled for UCS2 and UCS4 pythons separately. And last time I checked, setuptools did not allow you to publish both builds on pypi and pull the correct one. In fact, by default there is no such information in the filename which would make it possible to provide both extensions.
So what they did "improve" with Python 3 was making unicode strings the default. And they did that in a very backwards incompatible and IMO problematic way: they degraded bytestrings from strings to glorified integer arrays and enforced unicode on non-unicode protocols.
Just to give you an example: When you iterate over bytestrings in Python 2, the iterator will yield you a bytestring of length 1 for each character with that character in. While I was always a harsh opponent of strings being iteration, it was something everybody relied on. In Python 3, bytestrings are bytes objects, which are basically arrays of integers which look like strings in the repr, but yield the bytes as integers. So if you had code that relied on the iteration returning chars, your code will break. And yes, Python 3 breaks backwards compatibility, but this is something that 2to3 does not pick up and most likely you will not either. At least in my situation it took me a long time to track down the problem because some other implicit conversion was happening at another place.
Now at the same time unicode strings continue to yield unicode strings on iteration with the char in. That means suddenly bytes and unicode objects have different semantics, making it impossible to provide an interface for both bytestrings and unicode. There were tons of places libraries accepted both unicode and bytes in Python 2 because it "made sense". A good example is URL handling. URLs are encodingless. Some schemes hint a default charset, but in reality no such thing exists. However, applications themselves knew what encoding the URL would use, so they happily pass unicode strings to the URL encoder and that would use the application URL encoding to ensure the URL is properly quoted. In Django/Werkzeug/and probably many more libraries, if you passed unicode to the URL encoder, it would by default encode to UTF-8. However if the URL was came from another source with unknown encoding, it was possible to transparently pass the URL on. Also the decoding of URLs from somewhere else usually happened encoding-less. Many applications for example check the referrer of the page to check if the user came from a search engine and if yes, grab the search keywords from the referrer and highlight them on the current page. In that situation you can keep a list of known encodings of referrers in the URL and decode the referrer URL accordingly. In Python 3 the URL module in many situations uses an UTF-8 default encoding, or requires the URL to be UTF-8 encoded or provides a completely different and limited interface for byte URLs.
Sure, it might be sufficient for 98% of all users, but there are non obvious implications: a library that wraps urllib/urlparse and whatnot cannot reuse the same code for Python 2 and 3. When I started supporting IRIs in Werkzeug (basically the URL successors with proper encoding, already somewhat used by browsers) I chose to abandon the urllib module altogether and write my own simple decoder to make it easier to later port that thing over to Python 3 without changed semantics.
There are other examples as well: filesystem access. Python 3 assumes your filesystem has an encoding, but many linux systems do not. In fact, not even OS X enforces an encoding. You can happily use fopen to create a file that does not look like UTF-8 at all. And even there, the situation is a lot more complex because on OS X, different unicode normalization rules apply for the filesystem than for the applications themselves. So even if you are using Python 3, you still will have to manually normalize the filename to a different encoding when you want to compare filenames on the filesystem.
When I looked at the unicode stuff in Python 3 I did not see much value over nicely written libraries in Python 2 that enforced unicode usage. In fact, the update makes it especially hard to convert such libraries (that required unicode) to Python 3, because 2to3 assumes you are using byte strings and not unicode.
I wonder how this slipped past everybody and why Guido is okay with that, but the new super non-keyword-keyword in Python 3 is just wrong. The fact that this code works is alarming already:
class Foo(Bar): def foo(self): super().foo()
Assuming you know Python used to work function invocation wise the fact that this works smells. But it gets a lot worse because this code does not work:
_super = super class Foo(Bar): def foo(self): _super().foo()
That's just wrong. The use of the name of a global function (which btw I can reassign!) should never affect the bytecode generated, that's what keywords are for! Python also did not optimize while True loops because someone could reassign True, but suddenly it's sortof okay to do that. Also, why have self explicit when some magic in the compiler is now suddenly able to inject new symbols in the code? From that point onwards it is a one-liner to make the self implicit and suddenly there is no reason for that self being the explicit first parameter any more.
From what I remember, this was done to optimize the code. That's true, they do optimize something, but at the same time a function call of a global function in a method in a loop, will do a dict lookup every time the thing is invoked. Another thread could reassign the global function and suddenly the code would no longer call the new function because the old one was pulled into a local "register" (fastlocal or similar). And if you think "that's undefined behavior", I beg you to look into the mimetypes library. That will explain that no where in the world a Python implementation could be conforming if it avoids global lookups by optimizing them.
What I wished for Python 3 was to remove really useless dynamic features like pulling in functions on every call to allow more compiler/interpreter optimizations, easier multithreading support and everything.
Also what I was wishing for, for Python 3 was a better interpreter interface, and a revoked GIL or no GIL at all. I would love to be able to use multiple Python interpreters per application. Some sort of reentrant interpreter. That would simplify embedding Python into other applications and expand the possibilities. Just look at how V8 works internally to get an idea of what I was hoping for. I also wished there was a builtin support for freezing objects (no longer a frozenset, just freeze the set, and then finally be able to do the same for lists etc.). Also builtin support for proxing would be nice. The hack thread local libraries and the weakref module to proxy objects is just wrong, wrong, wrong (and unreliable as well). Imports are still horrible implemented, the standard library is still inconsistent or limited (and now even broken, cgi.FieldStorage in Python 3 anyone?)
What I really like is the new non local stuff. I was longing for that for a long, long time. booleans being a keyword, that should have been in there for longer, finally easier division semantics, improved metaclasses, class decorators, no more classic classes, dict views, the builtins returning iterables instead of lists etc (Though they should have added improved repr support that would allow me to introspect those iterators and freezing them at the same time [which I guess would once again require a cleaner and improved interpreter design to get right]).
So yes, I am disappointed how Python 3 worked out. They could have done so much more or skipped Python 3 altogether and get the cool stuff into an optional strict mode in Python 2.