On Saturday March 23, this year's pkgsrc conference (pkgsrccon 2013) took place in Berlin. Julian Fagir organized it with unending energy, even though pkgsrc is not the primary focus of his NetBSD work. He just took matters in his hands because no one else stepped forward. A big thanks for that!
The flight from Zurich to Berlin was uneventful. It was my first flight to TXL airport (I normally arrive at SXF), and arriving there is incredibly quick and convenient compared to the latter. The terminal is very small, and it takes just five minutes to go from the plane to a bus to the city.
Now for the conference itself: we started at 12pm on Saturday with a program of talks but no fixed schedule. Due to this, the conference took a long time (we finished only at 9pm or so) but on the other hand, it allowed for lots of interesting and fruitful discussion. At no point did we have to cut a question short because of a lack of time. Overall, I think that this was an excellent choice and made the conference more useful and productive.
We were about 21 people – mostly pkgsrc developers (of course) but also a Debian Developer (Ralf Treinen, who presented his work on Mancoosi), a FreeBSD dev and some interested users. I won't give an exhaustive recollection of all talks here but simply comment on a few ones that I found particularly interesting.
The most important theme of the conference was virtualization and cloud computing. Jonathan Perkin and Filip Hajny gave a talk about their company's product, SmartOS, and how it uses pkgsrc. SmartOS is a "cloud OS" based on OpenSolaris. It boots from a read-only medium (such as a CD) into a lean system that only does the administration of all the zones that it runs. All useful work happens in zones, which are a sort of lightweight VM solution specific to OpenSolaris. The zone images include access to a very complete set of pkgsrc packages for things such as a compiler. They can also run other OSes (NetBSD!) by setting up a zone that runs KVM. Joyent runs a large public cloud with SmartOS, where customers purchase virtual machines by the hour. This is similar to Amazon EC2 but with a focus on high performance.
Hubert Feyrer gave another talk about a similar theme. He described the use of Ansible for provisioning and setting up VMs. Ansible can automatically create VMs on EC2, gather the necessary information (such as the IP address) and do various setup tasks without further user interaction. This was all very impressive, even though the live demo failed. This was for two reasons: Somebody deleted the sudo package for amd64 from the NetBSD ftp server (boo), and the i386 VM failed to come up, the kernel paniced on startup. Joerg speculated that this was due to _some_ machines in their DC not having PAE enabled, while the i386 kernel uses PAE. This was interesting, as I had noticed the very same problem when I set up the netbsd-386-bsiegert continuous builder for Go.
Amitai Schlair alias Schmonz came out in a passionate defense of the venerable pkglint. He put the source on github and started refactoring the code and adding tests. He calls this approach TED for "Test Eventually Development" ;) and advocated a similar approach for the pkgsrc infrastructure: Every time a developer takes five minutes to understand a part of the infrastructure (when making a change, for instance), he or she should write a test for it. This is a very pragmatic and doable approach, in my opinion, and we should all do this.
I gave a slightly amended version of the "Go on NetBSD" talk I had given at FOSDEM 2013. There were a lot of valuable questions and discussion, both about the language and about how to package software written in it.
Aleksej Saushev ended the day with a talk that was not in the program about the Google Code-In and the problems that developers and particularly new contributors face. If pkgsrc can get more contributors, it gets more fixes, which in turn makes it more useful to users. More usefulness leads to more users, leading to more contributors. We should do more to get into this virtuous circle. There are about five different mechanisms to build and/or deploy packages in pkgsrc: build directly with "make package", pkg_chk, pkg_comp, the old bulk build scripts and pbulk. The basic frustration that should be overcome is the following: you want to upgrade a set of packages, the old ones are removed, new ones are rebuilt, and the build fails. Rolling back is difficult in general. pbulk could be a valuable solution to this, but its standard config is heavily tailored for a different use case, and its _two_ separate pieces of documentation are contradictory, incomplete and confusing. So the talk contained a call for action to fix those minor annoyances and generally document things better, which makes it easier for everybody.
My take-home message – and my next project idea – is the following: each time that I do a MirBSD bulk build using pbulk, I have to do a lot of painful steps to set up the right build environment on all my machines. This time, I will try to automate this process with Ansible, making up the recipes as I go along, and then (more importantly) publish these recipes for others to use and to share.