11–20 of 30 posts
I run large datacenters with thousands of boxes. I have a little app (written in golang) installed on each box that effectively is a task runner. Tasks can be written to do anything, including apt-get installing software. If apt-get fails to run, the task fails (context.WithTimeout) and is run again at a later date. No random hacks needed. Everything is built to be idempotent, self-healing and eventually consistent.
I run large datacenters with thousands of boxes. I have a little app (written in golang) installed on each box that effectively is a task runner. Tasks can be written to do anything, including apt-get installing software. If apt-get fails to run, the task fails (context.WithTimeout) and is run again at a later date. No random hacks needed. Everything is built to be idempotent, self-healing and eventually consistent.
cfengine3 does that (and much more) form me.
> This all started when I was looking into why instances in an auto scaling group were sometimes failing to bootstrap correctly. Please use Packer to build your images; don't do it on ASG instance deploy. If somebody in the company tells you you're not allowed to build your own images, tell them to go fuck themselves, and write an e-mail to that guy's boss's boss explaining how much engineering time you're wasting (a…
So you start using Packer to build your images. Your Packer script does "apt get install" and fails because something is holding the apt lock, and the author ends up writing more or less the same article.
Additionally, I work in Azure, and VM images in Azure are a world of absolute nightmarish pain there: a normal user literally cannot make the API call to bring up a VM with a custom image if the image is not in the same tenant. There is a way to do it with SP, but it is so completely and thoroughly undocumented as to be black magic. (Yes, if you Google, you'll be able to find Azure documentation on this exact subject. No, the instructions do not work.)
Yeah, I agree at the end of the day, it's the right way to do things, but it is an absolute, utter, PITA.
But then again, if you choose to not bring up a VM with a custom image, you get an unpinned image: "Ubuntu 20.04 LTS" is a moving target, and we once got one with a kernel that would BUG after ~5 minutes. Azure needed us to tell them what kernel we got from them.
Coming from Gentoo, encountering Debianesque systems only later, that all-out locking really surprised me. Portage, the Gentoo package manager, only takes a lock at the very end of the installation, when a file tree (result of compilation etc) is merged into the "live" filesystem - so that those merges are serialized and checked for collisions.
So whereas on Ubuntu, two competing apt commands cause one to fail, with no real recourse except a very nightmarishly long investigation resulting in the above article, in Gentoo, two competing commands … just … work.
It checks the dpkg lock, but I believe apt has its own lock, I believe at /var/lib/apt/lists/lock , and that option name strongly implies that it only checks the dpkg lock, still leaving you with a race condition on the apt one.
I run large datacenters with thousands of boxes. I have a little app (written in golang) installed on each box that effectively is a task runner. Tasks can be written to do anything, including apt-get installing software. If apt-get fails to run, the task fails (context.WithTimeout) and is run again at a later date. No random hacks needed. Everything is built to be idempotent, self-healing and eventually consistent.
cfengine3 does that (and much more) form me.
I can assure you that my use case is a bit more involved than what cfengine can provide.
Cheers.
But one of my colleagues figured out that it is probably because the apt-get is getting locked due to cloud-init and removed the flakiness by making packer wait[1] for cloud-init to complete before running the installation scripts that involved apt-get locking.
We too wished that there were more docs to help us, especially explaining how apt-get worked.
[1] https://github.com/hashicorp/packer/issues/2639#issuecomment...
> This all started when I was looking into why instances in an auto scaling group were sometimes failing to bootstrap correctly. Please use Packer to build your images; don't do it on ASG instance deploy. If somebody in the company tells you you're not allowed to build your own images, tell them to go fuck themselves, and write an e-mail to that guy's boss's boss explaining how much engineering time you're wasting (a…
> Please use Packer to build your images; don't do it on ASG instance deploy. So you start using Packer to build your images. Your Packer script does "apt get install" and fails because something is holding the apt lock, and the author ends up writing more or less the same article. Additionally, I work in Azure, and VM images in Azure are a world of absolute nightmarish pain there: a normal user literally cannot make…
Incidentally the author's suggestions are not that good, and not what I would suggest. He should be waiting for cloud-init to finish, not just apt:
cloud-init status --wait
while [ ! -f /var/lib/cloud/instance/boot-finished ]; do echo 'Waiting for cloud-init...'; sleep 1; done
Earlier quoted context omitted.
> Please use Packer to build your images; don't do it on ASG instance deploy. So you start using Packer to build your images. Your Packer script does "apt get install" and fails because something is holding the apt lock, and the author ends up writing more or less the same article. Additionally, I work in Azure, and VM images in Azure are a world of absolute nightmarish pain there: a normal user literally cannot make…
In the Packer build you can have your provisioning script wait for the cloud-init to finish because you're not delaying a production scaleout; you've got all the time in the world. It's true that this article still gets written, but I don't think OP was suggesting otherwise. Building fully baked images rather than installing things on production startup avoids all sorts of flakiness. Incidentally the author's suggest…
That just leaves unattended-upgrades, which is harder than the undead to kill, and Azure's WAAgent…