I'm approaching the end of converting a large system from Python 2 to Python 3. Here's why you don't want to do this.
The language changes aren't that bad, and they're known and documented. It's the package changes that are the problem. Discovering and fixing all the new bugs takes a while. BeautifulSoup: BeautifulSoup 3 has been phased out. I had my own version of BeautifulSoup 3, modified for greater robustness. But that was years ago. So I converted to BeautifulSoup 4, as the documentation says to do. The HTML5parser module is claimed to parse as a browser does, with all the error tolerance specified in the HTML5 spec. (The spec actually specifies how to handle bad HTML consistently across browsers in great detail, and HTML5parser has code in it for that.) It doesn't deliver on that promise, though. Some sites crash BeautifulSoup 4/HTML5parser. Try "kroger.com", which has HTML with <head><head>. The parse tree constructed has a bad link, and trying to use the parse tree results in exceptions. Submitted bug report. Appears to be another case of a known bug. No workaround at this time. https://bugs.launchpad.net/beautifulsoup/+bug/1270611 https://bugs.launchpad.net/beautifulsoup/+bug/1430633 PyMySQL: "Pymysql is a pure Python drop-in replacement for MySQLdb". Sounds good. Then I discover that LOAD DATA LOCAL wasn't implemented in the version on PyPi. It's on Github, though, and I got the authors to push that out to PyPi. It works on test cases. But it doesn't work on a big job, because the default size of MySQL packets was set to 16MB. This made the LOAD DATA LOCAL code try to send the entire file being loaded as one giant MySQL packet. Unless you configure the MySQL server with 16MB buffers, this fails, with an obscure "server has gone away" message. Found the problem, came up with a workaround, submitted a bug report, and it's being fixed. https://github.com/PyMySQL/PyMySQL/issues/317 SSL: All the new TLS/SSL support is in Python 3. That's good. Unfortunately, using Firefox's set of SSL certs, some important sites (such as "verisign.com") don't validate. This turned out to be a complex problem involving Verisign cross-signing a certificate, which created a certificate hierarchy that some versions of OpenSSL can't handle. There's now a version of OpenSSL that can handle it, but the Python library has to make a call to use it, and that's going in but isn't deployed yet. This bug resulted in much finger-pointing between the Python and OpenSSL developers, the Mozilla certificate store maintainers, and Verisign. It's now been sorted out, but not all the fixes are deployed. Because "ssl" is a core Python module, this will remain broken until the next Python release, on both the 2.7 and 3.4 lines. Also, for no particularly good reason, the exception "SSL.CertificateError" is not a subclass of "SSL.Error", resulting in a routine exception not being recognized. Bug reports submitted for both OpenSSL and Python SSL. Much discussion. Problem fixed, but fix is in next version of Python. No workaround at this time. http://bugs.python.org/issue23476 Pickle: As I just posted recently, CPickle on Python 3.4 seems to have a memory corruption bug. Pure-Python Pickle is fine. So a workaround is possible. Bug report submitted. http://bugs.python.org/issue23655 Converting a large application program to Python 3 thus required diagnosing four library bugs and filing bug reports on all of them. Workarounds are known for two of the problems. I can't deploy the Python 3 version on the servers yet. John Nagle -- https://mail.python.org/mailman/listinfo/python-list
