Donnerstag, 18. Juli 2013

The way forward with Python on Qt 5 (with a bit of history)

As you have probably heard on Twitter from the official Jolla account, the first Jolla will ship with Wayland. In that discussion, some worries are brought up about Python support with Qt 5. Here are my personal thoughts of how I see mobile Python development moving forward with the new technology. So first some background information:

If you want to write Python applications on a mobile platform, you usually need some way to interface with the native graphical toolkit. On Maemo 4 and Maemo 5, this was done via PyGTK and some additional bindings for the Hildon UI elements. PyGTK is licensed under the terms of the GNU LGPL, so you could use it to develop open and closed source applications.

When MeeGo 1.2 Harmattan (the system running on the N9) came up, the toolkit with which third party application developers wrote their applications changed from Gtk to Qt (or to be more precise, QML). Just like pure Gtk didn't provide all the UI elements for mobile (and therefore Hildon was used on top of Gtk), pure QML provided very little, and so Qt Quick Components were used for the Harmattan user interface (that includes things like buttons, toolbars, menus, dialogs, etc...). These could only be used from QML, so you had to use QML if you wanted your application to look anything like native (ignoring the MeeGo Touch Framework here, as it wasn't really used for third party apps, with the notable exception of sociality-mtf).

With Qt and QML being the new toolkit of choice, Python developers needed a different "bridge" to go from the Python world to the toolkit world. PyQt already existed for some time then, it had two license options: GNU GPL (basically requiring all your code that uses PyQt to also become GPL'd) and commercial. As the Wikipedia article of PySide says, after Nokia failed to reach an agreement with Riverbank Computing (the developers of PyQt) to change the license to LGPL, they started their own Python bindings named "PySide", which are licensed under the terms of the GNU LGPL. There's a wiki page with PySide-PyQt differences on the Qt Wiki.

As MeeGo at Nokia was ramped down, the core team of PySide had less and less time to spend on improving it. PySide is still kept up to date (the latest release is from July 9, 2013), and it already supports Python 3, but nobody has yet ported it to Qt 5. The most recent post I found about PySide and Qt 5 from one of the core developers was from November 2011. For using Python with Qt 5, there are two possibilities at the moment: Use PyQt under the terms of the GNU GPL, or get a commercial PyQt license. Another option to get LGPL'd bindings would be to add support for Qt 5 to PySide, but that needs somebody to adapt PySide to work with Qt 5.

But let's step back a little now. In my experience with writing Python applications for mobile operating systems (starting with Maemo 4 all the way to MeeGo 1.2 Harmattan), there are really two parts: The frontend, which is specified in terms of whatever the toolkit of the day happens to be, and the backend, which (in case of Python applications) is written in Python.

In Qt 5 / QtQuick, the user interface language is QML, which includes a declarative language and JavaScript for scripting. Writing QtQuick applications with Python doesn't mean replacing JavaScript, it means replacing the "backend" - code that would usually be written in C++ and take care of storage and communication. In fact, as QML grew more powerful over the years, you had to "go down to the C++ level" (or Python in our case) less often. And compared to imperative user interfaces (where you still go through your tree of UI widgets with code and then set some button text to some value), in declarative user interfaces the backend really only provides services to the frontend (that button gets the label from the backend in this way, and when the data in the backend changes, so does the button's text).

So, what we really need is a way to provide services to the QML UI by using Python. There's no need for having access to all classes in all Qt modules - even in Qt4 / PySide times, the classes that I used most often were QApplication (for the event loop), QDeclarativeView (for displaying the QML UI) and QObject (for providing signals and slots to interface the backend to the view). Things like access to the contacts database or GPS location can usually be done directly from QML (and displayed there) - it might be that you don't even need to have that data in your backend (and if you do, the UI can for example send the phone number of the contact "down to" the Python backend for further processing).

But before I'll tell you where I think mobile Python development should be heading, let's bring up yet another disadvantage of using PySide (probably applies to PyQt as well) for mobile applications: Startup time and responsiveness. Startup time is slow, because PySide libraries are big (they are - at least on the N9 - bigger than the Qt libraries they are binding), and the dynamic linker has to resolve all Qt symbols at load time instead of just the ones you end up using. Instead of seeing the QML UI right away, your application first has to load up the Python interpreter, then load up the PySide modules and other modules and then finally load and display the QML UI. And because of Python's Global Interpreter Lock, function calls from QML to Python will make the UI block and not be totally responsive, even if you are using threads.

So, in my opinion, mobile Python applications with Qt 5 have to be fast to load, lightweight and responsive. Fast to load can be achieved by making sure the libraries that get loaded are small, and that ideally the QML UI can already be displayed before the Python interpreter is even loaded and/or initialized. Lightweight can be achieved by not binding all the Qt classes, but only the ones you really need for creating QML applications. And responsive can be achieved by making sure that the interface between QML and Python is asynchronous, so the UI never blocks even if Python is working hard in the background to fulfill the requests of the user interface.

I currently have a prototype running on Python 3.3 (Python 2.6 and 2.7 is also supported still, but for me, this is a great opportunity to migrate some of my old code to Python 3, and new code should be written in Python 3, because, as you know, Python 2 sucks) and Qt 5 (I've just ported it to Qt 5 today, but it runs just as well on Qt 4 - with all the advantages like startup time and asynchronousness still being valid). In some unscientific tests that I carried out a few weeks ago, I've brought down startup time of gPodder on the N9 down from ~ 12 seconds (using Python 2 and PySide) to ~ 3 seconds (using Python 3.3 and the lightweight "PyOtherSide" approach). No code release yet, as this is tailored towards my gPodder experiment right now and has many things hardcoded, but once I clean it up (maybe as a QML plugin) and define the API in more detail, I plan to release it.

In the mean time, here are some (relatively old) videos of PyOtherSide running with Python 3 and Qt 4 on MeeGo Harmattan, Blackberry 10 and Android (so yes, it is portable).