About Legacy Python retirement

  |   Source

Note, I started this writing in September, it took me a while to finish, so things might have changed a bit during the writing.

This is a follow on previous post about retiring Legacy Python. I got some response, in particular thanks to Konrad Hinsen who took the time to ping me on Twitter, and who responded in a GitHub issue.

First I want to apologize if the post felt condescending, it was not my intention, by re-reading it I still do not feel it. Maybe this is because I'm still not skilled enough with words in English, and this is typically the kind of feedback I appreciate.

Don't break user code

One other thing that seem to leek through my post was the impression I was advocating breaking user code. What I am advocating is to make a step forward, by actually dropping compatibility with previous version. I really like semantic versioning, and as a tool maintainer, I would not drop support for APIs in between minor revisions. Keeping backward compatibility, is not always the easier job, and in a lot of case I fought hard and complain about changes that shouldn't have happen, and even reverted pull requests. As a developper I would have like these changes to go in though.

What I am advocating or is for libraries to start thinking of new API that would be Python 3 only. And potentially consider new major revision that does not support Legacy Python, or at least do not make efforts to support it. This is already the path that a growing number of tools are going through. Nikolas developers have decided that the next version (8.0) would be Python 3 only (more background). I suppose the necessary modification would be minimal, but the overhead and developer time to support Legacy Python at the same time was too much. Scikit-Bio is envisaging the same thing: whether or not they should stop trying to support python 2. Some project have even already started active removal of Legacy Python compatibility layer.

Last version of anaconda can be installed with the default root environment being Python 3.5. Some people do really awesome tools which are Python 3.4+ only.

An increasing number of distribution also now come with Python 3, and python 3.5 is now considered as default installed Python more and more. Ubuntu is considering having Legacy Python as persona non grata on next stable release, and so does Fedora apparently.

Changing major library versions would not break old code. Or at least not more than any other upgrade of major versions. User would still be able to pin version of the libraries in use, and if there are dependencies conflict, your package manager i there to resolve them. There is a specific requires_python metadata that you can give to your package to indicate whether or not it's Python2/3 compatible (including minor versions), which would just prevent non-compatible version to be installed, though it is not commonly use. This is one the reason the Py2/Py3 transition might look complicated to some people. One example is fabric, which is not compatible Python3, but will accept to install in a Python3 environment. In a perfect world you would just pip install your package and get the last compatible version without worrying about the Legacy Python/Python 3 compatibility, but we venture in Packaging, so here be dragons.

Python 3 does improve on Python 2

Even if, as a scientist, you might see Python as only a tool; this tool rely on a lot of work done by other. You might not feel the difference on the day to day basis when using Python, but , the feature that Python 3 provide do make a difference for the developers of software stack you use. During the last few month, we had the chance to get a new release of NumPy that support the new python 3 @ (aka __matmul__) operator, and a recent release of Pandas, that now support more operation with nogil. [Matplotlib] released its 1.5 version and should relatively soon release the 2.0 version that change the default color scheme. These are only some of the core packages that support the all Scientific Python infrastructure, and the support of both Legacy Python and Python 3 is a huge pain on developers. The support of a single major Python version for a code base make the life of maintainers much more easier, and Python 2.x branch is coming to end of life in 2020, we can do our best to allow these maintainer to drop the 2.x branch.

I really think that the features offered by Python 3 make maintaining packages easier, and if it takes me 10 minutes less to track a bug down because I can re-raise exceptions, or even get a bug report with a multiple-stage exception, it give me 10 more minutes to answer questions.

As a side note, If you think I'm exaggerating that Legacy Python support can be a huge overhead, I I've already spend a day on this bug and I'm giving up on writing a test that work on Python 2.

Also, for the record, re-raising exceptions, yield, keyword-only arguments are things that make the science I did during my PhD easier. Maybe no better, but easier for sure. I playing more and more with yield from, Python 3 Ast, generic unpacking since then, and they also do help (understand I wish they were there or I knew how to use them at the time). I strongly believe that the quality and the speed at which you do science is influenced by your environment, and Python 3 is a much nicer environment than Python 2 was. Of course there is a learning curve, and yes it is not easy: I went through it more than a year ago when it was harder than today. It was not fun nor was I convince after a week, but after spending some time with Python 3, I really feel impaired when having to go back to Legacy Python.

Just one Python

So while not all scientist are seeing the immediate effect of Python 3, I would really hope for all the core packages to be able to just forget about Legacy Python compatibility and have the ability to focus their energy on only pushing their library forward. It would really help improving the scientific stack forward if the all SciPy related libraries could be released more often. And one of the way to help that is to push Python 3 forward. We won't get end-user to migrate to Python 3 if the list of available features of package are identical. The carrots are new features.

Mostly what I would like is only one Python. I honestly don't think that Python 2 has a long term future. New distribution

For the in house legacy code that scientist cannot update, I'm really sorry for them, and it would be nice to find a solution. Maybe something close to PyMetabiosis that would allow to run Legacy Python modules inside Python 3 ? I understand the lack of funding and/or technical competence/time to do the migration. Though asking a volunteers based project which also lack funding to maintain backward compatibility indefinitely seem also unreasonable. Not all project have finding like IPython/Jupyter, and even though, those which get funding, also have deliverable, if we don't ship the new feature we promise to our generous sponsors, we will likely not get out grants renewed. More generally this lead toward the question on software sustainability in science. Is it unreasonable to ask a specific software to work only in a restricted environment ? It is definitively convenient, and that's why many papers now rely on VMs for results to be reproduced. But if you want your software to be used you have to make it work on new hardware, on new OS, which often means drivers, newer library, so why not new version of the language you use ? How many user are still not upgrading from XP? Is it unreasonable to ask then to install the version of libraries that were distributed at that time?

A good single resource on how to operate the Python2 to 3 transition is likely needed. Matt Davis created an example repository on how to write a Python2 and 3 compatible extension. CFFI is in general something we should point user to as it is seem to become the way which, in the long term, will lead to the less pain for future upgrades, and even for more backward compatibility with 2.6. Cython also allows to write pseudo-python that can run closer to C-Speed. Cython also have a pure python mode, which I think is under-used. I suppose that with Python 3 type-hinting, function annotation could be used in the same way that Cython Pure Python mode magics atributes work. This to provide Cython with some of the needed type annotations to generate fast efficient code. How to tackle the human problem is more complicated. It is hard to organize a community to help scientific software being ported to Python 3, as everyone is missing time and/or money. There is a need to realise that in-house software written has a cost, and this cost at some point need to be financed by institution. The current cost is partially hidden as it goes into Graduate student and Post-Doc time, who in the end will write the software. The Graduate and Post-doc often lack the best practices of a good software engineer which leads to technical dept accumulation.

Which path forward

The Python Scientific community is still growing. Laboratories are still starting to adopt Python. Wether or not it is a good thing, Python 2 will reach it's end of life is a couple of year. And despite the many in-house libraries that still haven't been ported to Python 3, it is still completely possible and reasonable to do science with only a Python 3 stack. It is important to make this new generation of Python programmers to understand that the Python 3 choice is perfectly reasonable. In a perfect world the question of which python to use should get an obvious answer, which is Python 3. Of course with time passing, we will always have some necessity fro developers speaking Legacy Python, in the same way that Nasa is looking for programmer fluent in 60 Years-old languages


Reviving Python 2.7

I'm Playing devil advocate here, to try to understand how this could go forward. The scientific community is likely not the only one to use old version of Python. Google recommend to install Python 1.6 to run it's Real-time Google drive collaboration examples. So once 2.7 support will be officially dropped by the PSF, one can imagine trying to have Google (or any other company) taking over. I am not a lawyer, but I guess in a case of language revival, trademark might be an issue, so this new version might need to change name.

It is not the first time a fork of a project have became successful. From the top of my head I can think of Gnome 2, that gave birth to Mate, Gcc that was forked to Egcs, (which in the end was re-named Gcc), as well as Libre Office, successful alternative to it's "Open" sibling.

Seeing that the Scientific community has already a lack of time and funding, I find the chance of this happening slim.


One of the main pain in the transition to Python 3 is keeping compatibility with both Python. Maybe one of the solution that will come up is actually jumping over all this Python 2/3 problem to an alternative interpreter like PyPy, which got a brand new release. The promises of PyPy is to make your Python code as fast as C, removing most of the need to write lower level languages. Other alternative interpreter like Pyston and Pyjion are also trying to use Just-In-Time compilation to improve Python performance. For sure it still need some manual intervention, but if it can highly decrease the amount of work needed. I'm still unsure that this is a good solution as C-API/ABI compatibility CPython 2.x is a complex task that does hinder language development. More generally, CPython implementation details, do leak into the language definitions and make alternative implementation harder.

Conversion tools

The Julia community dealt with migrating code from Fortran by writing a Fortran-to-julia transpiler, Python have the 2to3 and now python-modernize tools, it might be possible to write better conversion tools that handle common use case more seamlessly, or even import-hooks that would allow to import packages cross versions ? Maybe having a common AST module for all version of Python and have tools working more together is the way forward.

Step forward

Make Python 3 adoption easier, or make it clearer what is not Python 3 compatible, both for humans and machine. There are some simple steps that can be taken: Make sure you favorite library run tests on latest python version, it's often not too much work.

Point the default docs at Python 3 using canonical links. I too often come across example that point to Python 2 docs, because they are better referenced by Google. Often the difference LPy/Py3 is small enough that example works, but when it does not, it give the impression that Python 3 is broken.

Tell us what you think, what you use, what you want, what you need, what you can't do.