2 ms·
If your systems are cattles [0] your case holds. If not, your method won't work on the long term. Also, preseeding can do much more than ansible/salt/puppet. I
by bayindirh 6y ago
If your systems are cattles [0] your case holds. If not, your method won't work on the long term.
Also, preseeding can do much more than ansible/salt/puppet. It can produce %100 reproducible systems under 5 minutes.
Case in point:
I manage a cluster and our deployments are managed by XCAT [1]. I just enter the IP addresses, the OS I want (CentOS or Debian), and some small details in 5 minutes. With three commands, I power up the machines and in 15 minutes ~200 servers are installed the way I want, with the OS I want, identically and with no risk to drift into anything.
The magic? XCAT generates Kickstart/Preseed files and instead of pushing a minion and blasting through the installation, it configures them properly. Some compilation and other stuff is done with off the shelf scripts we have written over the years. More stable and practical than trying to stabilize a salt recipe/pillar set or an ansible playbook.
I only re-install a server if its disks are shot or it somehow corrupts itself (which happens once a year per ~500 servers due to some hw glitch or something).
The new ways are cool, but they don't replace the old methods. You're free to choose the one you like most.
[0]: https://www.engineyard.com/blog/pets-vs-cattle https://www.engineyard.com/blog/pets-vs-cattle
[1]: https://xcat-docs.readthedocs.io/en/stable/ https://xcat-docs.readthedocs.io/en/stable/
- gorgoiler 6y agoWhen your user data / logs / spools / db files are on a filer or confined to one local place that’s easily backed up and can be deployed — I tend to use /data — there’s no reason to have ‘pets’ any more. What’s the non-cattle use case you are referring to? Systems that can’t be wiped because of important persistent user processes? That’s the only volatile state I can think of that would be lost by reimaging.
- bayindirh 6y agoIn normal enterprise operations, it's easily attainable. I presume your servers are not extremely under load and they mount a remote storage system. In our system, there's tens of storage servers under heavy load (while they are redundant they are generally actively load-balancing) and more than 700 nodes which are cattle but under max load 24/7/365. The whole thing doesn't have any time or space to breath basically. While losing nodes doesn't do any harm, we lose some processes hence lose time and, we don't want that. Even if we can restore something under 15 minutes, avoiding this allows us to complete a lot of jobs. We don't want to rewind and re-execute a multi-day, multi-node job just because something decided to go south randomly. Our servers are rarely rebooted and most reboots mean we have a hardware problem.