After PyCon last month, Python packaging saint Donald Stufft generously tweeted:
If you’re interested to see Py2 vs Py3 breakdowns for a Python package, let me know while I still have the data set loaded (~300GB DB).
He had the database loaded for his diverting article on a Year of PyPI Downloads, but I was curious about three packages I own or contribute to.
Here's PyMongo's downloads this year, by Python version:
Python 2.7 dominates, and it is growing at the expense of 2.6. Python 2.4 is absent, and PyMongo downloads for Python 2.5 vanished last year, which validates our decision to drop Python 2.4 and 2.5 from the latest release, PyMongo 3.0.
Donald made a second chart isolating the Python 3 downloads:
As one expects, people who use Python 3 use the latest one, Python 3.4. When we released PyMongo 2.8.0 at the end of January, the download spike was entirely Python 3.4 users.
I expect Python 3.5 will soon dominate among Python 3 users, and Python 2.6 will continue to decline asymptotically, while the Python 2 versus 3 ratio overall will stay steady for a few more years.
Something is wrong with Donald's chart for Motor, but it suggests that Motor users are like PyMongo users: 10% run the latest Python 3, 50% run Python 2.7, and the others run a smattering of other Pythons. (Motor has never supported Python before 2.6.)
In Donald's Python 3 chart, it seems only Python 3.4 is represented:
Motor's potential user base includes all Tornado users, so I was curious about Tornado's overall distribution:
I have no explanation for the spike of Python 2.6 downloads last fall; Tornado's release schedule doesn't strongly correlate with it. Tornado's user base is distributed similarly to PyMongo's, though more inclined to stay on Python 2.6.
Again, Tornado's Python 3 users hold steady at ten percent, but they switched to Python 3.4 quickly after it was released.
The takeaways are: Python 2.4 is dead, and 2.5 is effectively dead for PyMongo, Motor, and Tornado users. It's important to keep packages working in Python 2.6—mainly for enterprises with long-term support contracts for Linux versions that shipped with 2.6—but adding features or optimizations that only work in Python 2.7 is reasonable now.
It's critical we keep testing Python 3.5 alphas and betas as they come out, because the 10% of people who run Python 3 will migrate to 3.5 rapidly. Python 3.2 is nearly dead, and 3.3 will also vanish soon.