Mirrored system disks should be trivial to set up

June 9, 2008

I have a simple request for people putting together installers for modern systems, especially systems generally aimed at servers: it should be dirt simple to do an install with mirrored system disks.

With most modern servers having two drive bays (often hotswap ones) and disk space being so cheap, going to mirrored system disks make a lot of sense. But most people won't move to this configuration until it is somewhere between easy and trivial to set up, much like many people did not move to LVM-based system setups, despite their advantages, until installers made it trivially easy to install the system that way.

What I'd like to see for mirrored system disks is something similar to the LVM approach. If the system detects two identically sized disks, it offers 'standard mirrored system disks' as a partitioning option, and then does all of the magic necessary to make everything work nicely. (These days, probably using LVM on top of a single software RAID partition.)

But really, the specifics don't matter: what matters is that it is long since time for mirrored system disks to get first class support as an installation option, because they are (or should be) so common these days.

(Why yes, I was installing a Red Hat Enterprise 5 system today and cursing yet again the backwards way that the Red Hat installer approaches this. But my co-workers are wrestling with this for Solaris 10, and we have a pile of Ubuntu 6.06 machines that would have mirrored system disks if it was easily done in the installer.)

Comments on this page:

From at 2008-06-10 14:20:35:

I could not agree with you more that the installer should make it easier to enable mirroring. This has been a failure in the Linux distro installers for quite a while now. But, for some reason this post really annoys me. The fact that you have systems that aren't mirrored simply because someone was lazy?

Perhaps I'm being optimistic by thinking that because doing something is hard, but the correct thing to do that the person would put the effort into it (particularly if it's a one time thing). Or perhaps I'm being overly judgmental, but I was really taken aback when I read this.

Sure we all have deadlines and tight budgets to follow, but this is really cutting your toes off so your shoes will fit.

Thanks for all your posts and please ignore the trolls (like myself) who feel compelled to comment without all the information. :)

- Andrew.

By cks at 2008-06-10 17:15:20:

These particular non-mirrored machines are generic machines of which we have several already and which we can build new instances of in under an hour. Often we have a full dd'd image of the first drive on the second drive, so we can recover from a drive failure even faster if we're okay with having to apply pending updates or the like again.

(There are also a number of non-essential servers that we only have one of but could take a while to rebuild without anyone really caring.)

Note that there is no strong guarantee that a typical running system can survive a drive dying; a lot depends on just how the drive decides to die and what else its creativity manages to take with it. (Although I believe that life is better than back in the days of shared SCSI and IDE channels.)

From at 2008-06-11 14:21:23:

I just installed Solaris Express snv_90 from scratch. You just tell the installer to use ZFS, choose 2 drives, and you get a mirrored ZFS root.

I hope this feature will make it to Solaris 10 update 6.


Written on 09 June 2008.
« Thinking about Python's inheritance model
Tabs versus windows, or why I usually want windows »

Page tools: View Source, View Normal, Add Comment.
Login: Password:
Atom Syndication: Recent Comments.

Last modified: Mon Jun 9 23:43:29 2008
This dinky wiki is brought to you by the Insane Hackers Guild, Python sub-branch.