minutes from OctConf2012: pkg - package system and structure

There's an Octave code sprint planned for the weekend of 17-18 of November with the purpose of improving the functionality of pkg(). Some of the improvements were discussed at OctConf2012 but more have since been discussed on the mailing list. At the time, I started writing a long report (in the style of a meeting minutes) about pkg() and Agora, the things that had a bigger impact for Octave Forge but only finished the part of pkg(). The comments I received were that the text was too long and detailed so I ended up writing a shorter text that covered both items.

But with the code sprint coming up soon, we do need a proper document stating what pkg() should be doing. When we looked at it during OctConf, things were intertwined in such a way that any changes required fixes everywhere else. My guess is that if a bunch of people start coding away on it at the same time, even if on different problems, we will keep stepping on each other toes. And if we do create 1 branch for each sprinter, merging them back together might not be so easy. The ideal would be to have something like a python's PEP.

In the mean time, I'll post here the minutes of the pkg() discussion during OctConf2012.

These were first dicussed between Carnë Draug (CD), Juan Carbajal (JC) and Carlo de Falco (CF) before being presented to the community present at OctConf2012 on the morning of July 19 for further discussion. During the rest of the event CD, JC and CF continued discussing the plans whose conclusions are now reported.

It was the opinion that the current problems with the pkg system are caused by the code complexity of pkg(), itself caused by the path of its development, slow, as new features were added as they were needed, one at a time on top of the previous ones. Also, the nature of the problem, mostly string processing and directories content, is not solved with the Octave functions with clean code. As such, a list of problems with the current system and new desired features was made to have a clear design of what the system should support.

It was proposed by CD to rewrite pkg() in Perl. Despite the language fame for being hard to read, it would allow for shorter and fast code. It would be much easier to maintain by someone familiar with Perl than the current code is for Octave programmers. Even for someone unfamiliar with Perl, it should take less effort to fix bugs. Plus, perl is already an Octave dependency so Octave users will already have it installed on their systems. CF pointed out it is just a building dependency and therefore not necessarily present on the user system. While it is true that pretty much all Linux distributions require perl, it does not hold for Windows. pkg() is currently faulty on Windows so it wouldn't be a problem but the hope is to make it work for them too. The idea to use perl was then rejected.

CD, CF and JC were of the opinion that the autoload option was not good and that pkg() should not support it. Packages can shadow core Octave functions, and even other packages functions. On the later case, no warning is given. Code meant to run with specific packages, or even with no packages at all, may catch others by surprise. Also, some users are not aware that some functions they use come from packages. Forcing them to load a packages as needed will make them know what they are doing. No other programming language has packages, modules or libraries loaded by default (with the exception of special cases such as python implementations). JC gave the example of a practical class where the teacher gives commands for the students in front of their pre-installed octave systems. The first command they should run should be pkg load and the professor should not have installed the package with autoloading by default. Any user would still be free to configure his .octaverc file to load packages at startup. That is the objective of .octaverc not of a package system, to configure startup of octave sessions. CD pointed that loading of packages is also not completely safe at the moment. When loading a package, its dependencies are loaded at the same time. However, these dependencies can be unloaded leaving those dependent on them loaded and not issuing a warning. The discussed options were: unload all other packages at the same time, refuse to unload the package, keep the current behaviour. The verbosity level for attempting to unload such package was also discussed but no conclusion was reached.

A frequently requested option is to automatically identify, download and install all package dependencies. All CD, CF and JC agreed that this should be implemented. It shouldn't be too much work since the dependencies are already being identified and can be downloaded with the -forge flag. All code is already in place, it should only require a few extra lines. This is obviously only possible for dependencies on other packages. A check on the availability of external libraries and software can be performed with Makefiles but pkg() can't solve them.

CF suggested to add two more options to the install option that would allow installing a package given a URL and another to install the development version of a package. As with the option to automatically solve dependencies, and for the same reasons, it should be easy to implement the URL. CD argueed that the dev option should not be implemented because it would stop packages from being released as users become more used to it and start installing possibly broken packages. CF said it would still be very useful for package maintainers preparing a new release. JC suggested to use releasePKG() on the admin section which already does it. It requires for a local clone of the repository which should already be available if it is for a developer preparing and testing a new package release. It was agreed that the url, but not the dev option would be added to pkg().

CF and JC were of the opinion that the package system should not support both local and global installs and that all installations performed by pkg() should be local. CF reported that on Mac systems global installs were made local even when running octave with sudo. CD mentioned that on Windows systems the opposite happens, and all installs are global (such being checked with isppc() on the code). The two types of installations are exclusive to Linux systems. CF and JC said that global installs should be kept for the distribution maintainers and pkg() should deal with local installs only. CD argueed that this would mean that system administrators, even the ones compiling the latest octave version, would be dependent on their distro maintainers for obtaining the latest version of packages. CF and JC replied that supporting both types complicates the system and that packages are more user specific. It was agreed that the option was then going to be removed. After discussing this option with Jordi Hermoso, it was discovered that at least the Debian maintainers actually use pkg() to prepare the packages. It was then decided that pkg() would deal with both installation types.

All CD, CF and JC were of the opinion that the -local and -global install flags were still useless and should be removed since the type of installation was already being decided based on whether the user was root, this flags only useful to force the other type. CD proposed changing the default for a global installation if there was write permissions rather than being root as to permit an octave system user to make octave global installs. This also allows for a local installation of octave (a regular user compiling and installing octave on its home directory for example), to make a global package install. Global relative to the local octave installation, the packages on the octave tree rather than on a separate directory. This should allow for a cleaner organization. These two changes were made and commit before the end of OctConf2012.

The current list of installed packages, local and global, is a simple struct saved in the -text format. CD was of the opinion this should be made a binary format to discourage users from manually editing the file and accidentally breaking the system. CF argued the opposite, that such editing may be necessary. It was decided to simply leave a warning on the file header.

CD noticed that it is not possible to use packages in a system that has more than one octave version installed. While .m functions will work normally, .oct files compiled at installation time are version specific and will therefore fail. These are placed in API specific directories to avoid their incorrect loading but reinstalling the package removes them, forcing a reinstallation of the package everytime a different octave version is to be used. CF also pointed out that a system to perform reinstalls should be made and the packages source kept so as to reinstall packages with new octave versions. CD noted that this would also allow for use of %!test of C++ functions after install. Similarly, it was noted that currently is not possible to have more than one version of the same package installed.

List of conclusions: