[bitfolk] To partition or not - ease of install vs. ease of resize

Hello, I've just been looking at the CentOS network install. It works great apart from one thing: anaconda/disk druid refuse to install directly onto a disk device without a partition table. There appears to be no way to work around this. It's been a common theme with the Debian and Ubuntu installers and with various other tools that they get a bit freaked out by the idea of a filesystem directly on a block device, but in those cases there have been workarounds. So, what is so bad about having a partition table? Well, assume that the customer has /dev/xvda with a filesystem directly on it, no partition table, and they want to grow that. The process is as follows: 1. We enlarge their volume on the host. 2. Some time later the customer shuts down and boot their VPS again to see the larger /dev/xvda. 3. Customer then does resize2fs /dev/xvda to online grow their root filesystem into the new space. You have to admire the simplicity of this; it's pretty hard to get a shut down / boot wrong, and then you just type "resize2fs" with no parameters and it's done. What a shame that's it's unsupported by every major Linux distribution's installer. Assume instead that the customer has a partitioned /dev/xvda with xvda1 as / and xvda2 as a swap partition. Now the process is more like: 1. We enlarge their volume on the host 2. Some time later the customer shuts down and boot their VPS again to see the larger /dev/xvda. 3. Customer can't just enlarge xvda1 because the swap partition xvda2 is in the middle. So, delete that. 4. Edit partition table to delete xvda1 and recreate it again larger. Note that it fails to tell the OS about the new partition layout because it's in use. 5. Reboot again to see the larger xvda1 device. 6. Online resize of xvda1 with resize2fs. 7. Recreate swap (or use swapfile). You can see that the second process is much more complicated than the first. Not only that but it contains a number of terrifying steps such as step 4 where the partition for the root filesystem is deleted. I shudder at the thought of inexperienced sysadmins doing things like this. I don't like doing it myself. Disk resizing is quite a common operation at BitFolk. It's a lot more common than reinstalling. The first process (without partition table) has been done by customers around 100 times according to my records, and has never gone wrong that I know of. So, another possible approach: every VPS is set up with LVM from the start. xvda partitioned, small / (containing /boot) on xvda1, rest of disk as xvda2 and used for LVM physical volume, broken up into small logical volumes for /usr, /var, /home and swap. If more space is needed it comes as xvdb, xvdc, xvd.., which the customer adds to their PV and puts in LVM. Advantages - The most flexibility for future growth. - Works in all major distribution installers. Disadvantages - I expect very few customers know how to drive LVM, and I don't really want to be in the business of selling or giving away LVM training. - Makes it a bit harder for me to poke around in a customer's VPS install from outside, which is sometimes requested. e.g. if it becomes compromised. Something between the last proposal and the one before it: Strongly suggest that people who partition their disks (through choice or because they want to use an installer that gives no choice like the CentOS one) use LVM, but set up VPSes ourselves directly on the block devices. At least then the CentOS installer is usable. In the CentOS case this could actually be enforced by using a kickstart recipe to automatically partition the disk in some suggested way, although this does rather take away some of the point of using an installer (flexibility). Anyone have any thoughts? Any tricks I've missed here? Constraints imposed by Xen: First device (xvda or xvda1) must be ext3, and must contain /boot. Cheers, Andy -- Of course not. Everyone knows you can't have your kayak and heat it. -- James Fidell