I like Python; code written in Python tends to be very easy-to-read, and the massive number of libraries make it easy to make useful programs.
|It was a mistake to have not implemented both Python 2 and Python 3 in a single CPython executable.|
Python 3 is in many ways an improvement over Python 2, but it was a mistake to have not implemented both Python 2 and Python 3 in a single executable widely used by most people (CPython). This mistake makes it very difficult to transition from Python 2 to a Python 3 implementation. If your program is small and doesn't use any libraries, transitioning tends to be easy; just use the "2to3" conversion program, fix up some problems manually, and you are off and running. But that's a weird special case, because Python is popular and has spawned a massive number of libraries. Most Python programs depend on libraries, which depend on other libraries, which depend on other libraries (you get the point). If any library anywhere doesn't support Python 3, then nothing else that depends on it can use Python 3 either. So that means that every library you depend on (transitively!) must simultaneously switch to support Python 3 if you want to use Python 3. You also have to translate your entire program to Python 3, all at once; if your program is big, this is often impractical even with tools like 2to3.
Because of the library bottleneck, and the basic incompatibilities between Python 2 and Python 3, Python 3 uptake has been slow. Linux distributors find it painful to support Python 3 because of this (see this Fedora thread and the Fedora Python3 F13 page).
On 2014-04-13 Guido van Rossum had to add another 5 years to the Python 2.7 end-of-life date (to 2020), and in the comments admitted that many users "cannot yet migrate to Python 3". Even the infrastructure supporting Python can't switch to Python 3. Mercurial, a popular program written in Python, has tried to move to Python 3 twice, and abandoned the effort because it determined that it would cost an additional man-year of effort to make the switch. In case you missed the irony, the Python developers agreed to move Python development to mercurial (per PEP 385), most Python.org repositories have now moved from the Python.org subversion repositories to the Python.org mercurial Repositories. So Python.org is using a tool in Python to develop Python 3 that itself has abandoned switching to Python 3... because the transition to Python 3 is far too difficult.
The Python Wall of Superpowers (formerly Python Wall of Shame) shows many libraries support Python 3, but remember, if 99% of your Python 2 libraries also support Python 3, then you cannot use Python 3 and must use Python 2. Some libraries (such as Twisted, Zope2, and meld3) have declared that they will never support Python 3. Through herculean effort, a number of libraries and programs do now work on Python 3. But even as of 2014, every developer I know who uses Python uses Python 2, never Python 3 (and Python 3.0 was released in December 2008!).
I think it was a terrible mistake to have combined changing the programming language with switching to a different C implementation. It is perfectly possible to have a single "python" executable that implements both version 2 and version 3 semantics to allow mixing of the different notations. Then, any program could be written with the version 3 semantics, yet call on libraries that were written with the version 2 semantics. There is a reason that people don't like backwards-incompatible changes; it makes transition rediculously hard.
The mailing lists sometimes talk about "carrots" to use Python 3, but I think that is missing the point. The problem is not the lack of carrots. The problem is the landmines, barbed wire, and guns shooting at you if you try to transition existing Python 2 code to a Python 3 implementation. The transition is broken, not the destination.
|The transition is broken, not the destination.|
In a lot of cases it would better if we write Python programs that worked, without change, on both Python 2 and Python 3 (where "Python 3" really means version 3.3 or later). Then we don't have to muck with 2to3 (or 3to2) and other nonsense. This approach has been called forwards-compatible Python.
Python 2.6 includes many capabilities that make it easier to write code that works on both 2.6 and 3. As a result, you can program in Python 2 but using certain Python 3 extensions... and the resulting code works on both. Python 2.6 has been out for a while, so for many people, requiring 2.6 is a reasonable precondition. (Some people can mandate version 2.7, which is even easier.) Where that's too difficult, write in Python 2.6+, but add some of the syntactic and semantic niceties of Python 3. Mark Summerfield has a nice summary of Python 2 and 3 idiom differences.
Although it's taken 5 years, it looks like the Python developers are finally starting to see the light and have moved closer to my position. In 2014, Guido van Rossum summarized the PyCon conference, "... I do think we should make things better for people who have to straddle Python 2 and 3 in a single codebase, by developing more tools, and by security and possibly installer updates to 2.7 (PEP 466)... Some suggestions that were made: PSF financial support for tool development and/or porting... additional 2to3 fixers to help convert Python-2-only code to Python-2-and-3-single-source code, a separate linter, a sumo 2.7 distribution that includes all known backported-from-Python-3-stdlib packages,... The recommended and least painful way to develop for Python 2 and 3 is definitely to use a single source that runs under both without translation; we no longer recommend auto-generating Python 3 compatible source code using 2to3, for a variety of reasons."
A simple way to do this is to use Python 2.6+ for development and begin each of your Python .py files with the following:
from __future__ import print_function, unicode_literals from __future__ import absolute_import, division
These switch to Python 3 meanings for key constructs. Now you use print(...) instead of a print statement, unicode strings, imports will always be absolute, and division will create floating-point values as needed (i.e., 1/2 now returns 0.5).
Python 2.6 includes a number of Python 3 features by default (like support for "bytes"), so you can just use them directly. In some cases, you should avoid using certain constructs and replace them with another (my thanks to Running the same code on Python 2.x and 3.x which points out some of these). For example:
|d.has_key(k)||k in d|
Some code constructs require a little extra work to make them work the same way in Python 2 and Python 3. For example, Python 3's "range" is the same as Python 2's "xrange". We can do this by inserting after the "from __future__" statements the following:
try: xrange = xrange # We have Python 2 except: xrange = range # We have Python 3
Now we can use "xrange(...)" in the rest of the file, and it will work correctly. (You could use "range()" directly, but in Python 2 this can be very inefficient.)
You can also import packages and rename them. However, instead of trying to do all this yourself, try to use the "six" module. The six module packages up many of these Python2/3 differences and makes it easy to "import whichever library I need".
You can use python's "-3" flag when running Python 2 to warn you about code that will not cleanly transition to Python 2. Using that flag, and fixing what it finds over time, is a much more reasonable transition approach than trying to change everything simultaneously.
One of the advantages of Python is that it's a clean language to read; too much of this stuff makes it too complicated. There's a philosophical question as to whether or not you write in Python 2 (with some modifications), or in a Python 3 that happens to work in Python 2. For example, do you choose to use "xrange" or "range" as the name in the code? I prefer working in "python 2 with specific modifications" right now, for the following reasons:
Python 2 and Python 3 have some different library interfaces, and trying to deal with all of that in each file can be awful.
Writing code that works in both 2 and 3 can become a serious pain, so when it gets too difficult, I abandon it, and make the code work on a Python 2 that adds some features of 3. Over time, I can modify the code to be more 3-like, presuming that future versions of Python 2 add notation from Python 3. This gives me a practical way to transition to Python 3, gradually, and then use 2to3 when all libraries have made that final step. By that point, the code differences will be trivial instead of the current chasm.
One of the biggest challenges is handling Unicode, because Python 3 is stricter about Unicode than Python 2. Generally, you want to encode into Unicode as you read data in, and only encode back to bytes as you write it out (this is called the "Unicode sandwich" approach). See the Unipain presentation, e.g., at Pragmatic Unicode ~ or ~ How Do I Stop the Pain?, for more information. One nasty problem is that both Python 2 and Python 3 have a type called "str" but they are completely different; Python 2 "str" is for bytes (NOT Unicode), while Python 3 is for Unicode (and NOT bytes). You can write code that works on both, but the mixing of names is confusing.
The biggest problem with Python 3's approach to Unicode is that it works fine when given perfectly-encoded Unicode data, and that you always know what this encoding is. Sounds great, but this is almost never true. For example, Unix filenames are not strings, they are byte sequences that may not be legal character sequences. Later versions of Python 3 have papered over this simple case (see PEP 383). You normally have no idea what the encoding of the standard input is; the officially-claimed one is often wrong, and it may not even have one. In other cases there is often either no information about the encoding, or the encoding data is wrong (the latter is especially common in CJK settings). Data often has a nonstandard mixture of different encodings, e.g., many "UTF-8" datasets are actually mixtures of 1252 (particularly curly quotes) and UTF-8. This means that using chardet, or using the character set encoding as provided by response headers, is often worthless. Armin Ronacher in January 2014 argues that "Python 2's way of dealing with Unicode is error prone and I am all in favour of improving it... [but] Python 3 is a step backwards and brought so many more issues that I absolutely hate working with it." In many cases you can use BeautifulSoup's UnicodeDammit class to figure out the encoding (it also has a "detwingle" operator to handle data that is UTF-8 but also includes Windows-1252 characters for curling/smart quotes). A discussion on python-readability issue 42 noted that Unicode Dammit often worked, but not on Korean pages (where encodings are often wrong).
Pragmatic Unicode discusses Unicode issues as follows: "Here's Fact of Life #4: You can't determine the encoding of a byte string by examining it. You need to know through other means. For example, many protocols include ways to specify the encoding. Here we have examples from HTTP, HTML, XML, and Python source files. You may also know the encoding by prior arrangement, for example, the spec for a data source may specify the encoding. There are ways to guess at the encoding of the bytes, but they are just guesses. The only way to be sure of the encoding is to find it out some other way... Fact of Life #5: Sometimes you are told wrong. Unfortunately, because the encoding for bytes has to be communicated separately from the bytes themselves, sometimes the specified encoding is wrong. For example, you may pull an HTML page from a web server, and the HTTP header claims the page is 8859-1, but in fact, it is encoded with UTF-8." These may be facts of life, but Python 3's design doesn't actually support these facts of life very well... especially point #5. Python 3's design seems to presume that you always can know what the encoding is, even though there is often no way to know in a sure way what this encoding actually is. Indeed, Pragmatic Unicode presumes that there data uses some standard encoding, even though real-world data is sometimes a mishmash of different encodings. Armin Ronacher points out in May 2014 that these complexities make creating many real programs in Python 3 much harder than in Python 2 (the opposite of what was intended). It may be possible to work with real-world data in Python 3 by reading bytes of data, and then combining the data along with the claimed encoding information to infer the actual encoding (and to clean up data if it is not in any valid encoding). BeautifulSoup's UnicodeDammit class shows a plausible way to (potentially) get there.
I hope that the Python 2 C implementation will continue to be upgraded until it supports nearly all of the Python 3 features. In particular, I'd love to see "import __future__ python3", which would try to make it as python3-like as possible, including the new Python 3 names and interfaces. Then programs and libraries could easily switch to version 3 features at their own pace, instead of requiring a "flag day". It would also mean that code could be quite clean.
PEP 477 has added ensurepip to Python 2.7.9. A key reason is that this makes it easier to access third party modules for migration from Python 2 to Python 3, including six, modernize, and future.
I think it's plausible that Python users will eventually end up at Python 3, but the Python developers did a terrible job in handling the transition. Expecting everyone to instantaneously convert code, both their own and all libraries transitively, was a terrible and unworkable idea. Hopefully other language developers will learn from this situation and handle transition more gracefully.
Lots of other pages have similar info on making code work on both 2 and 3 directly. A lot of them include doing this for versions of Python before 2.6, which tends to be more work. These include:
Feel free to see my home page at http://www.dwheeler.com. You may also want to look at my paper Why OSS/FS? Look at the Numbers! and my book on how to develop secure programs.
(C) Copyright 2009 David A. Wheeler.