How closely have you examined the live migration requirement for specific tolerances? If you absolutely need seamless, uninterruped-client-connection live migrations, your options are fairly limited—a moment of downtime and client reconnections are hard to avoid (and impossible, afaik, without some support or tolerance for it in the application). If a moment of interruption and reconnection is acceptable, then you may be able to meet your requirements through scripting and ZFS snapshots. Joyent has published notes on an unsupported procedure for this:
https://docs.joyent.com/private-cloud/instances/compute-nodes In broad strokes, it's a matter of provisioning the instance on the target host, snapshotting the source dataset and sending it to the target, stopping the source instance, and bringing up the target instance. With a high-capacity fabric and incremental snapshots, downtime is effectively reduced to the time it takes the target instance to boot. *John Burwell* Systems Administrator 16945 Northchase Drive, Suite 1610, Houston, Texas 77060 *M* 281-874-2110 *D* 281-875-7997 *C* 713-269-8605 crystaphase.com On Tue, Dec 8, 2015 at 9:43 AM, Bryan Horstmann-Allen <[email protected] > wrote: > > +------------------------------------------------------------------------------ > | On 2015-12-08 17:17:09, George Mamalakis wrote: > | > > The lower the level the better. For example, FiFo's "solution" with LeoFS > > is very close to what we need (even though it's not at the hypervisor's > > level), but the problem with it needs us to set the redundancy > > configuration during the design phase which is somewhat difficult for us, > > since our setup is expected to grow. As I stated in my original email, I > > haven't found the SDC's section on how to accomplish high availability, > so > > that's what I'm trying to understand through this list. > > qemu in illumos/SmartOS does not support live migration. Nor do we have > live > migration with zones (as does Solaris 11.) > > The abstraction for HA in SmartOS is assumed to be higher up -- at the app > level. You should be architecting your aps in such a way they are > resilient to > the loss of a zone/VM _and_ a physical node. > > SDC itself has HA components, but still assumes you are implementing a load > balancing tier (which you can run on your compute fleet) for pieces of it. > > https://docs.joyent.com/private-cloud/resilience > > > Which means that the boot order is random? Or do they start based on when > > they were created (eg. older first)? > > They all effectively boot at once, see /lib/svc/method/svc-zones. If your > apps > require boot time orchestration, use something like ZK/Consul/etcd (or > whatever > you prefer) to manage that. > > Cheers. > -- > bdha > ------------------------------------------- smartos-discuss Archives: https://www.listbox.com/member/archive/184463/=now RSS Feed: https://www.listbox.com/member/archive/rss/184463/25769125-55cfbc00 Modify Your Subscription: https://www.listbox.com/member/?member_id=25769125&id_secret=25769125-7688e9fb Powered by Listbox: http://www.listbox.com
