Open Stack

On 16 April 2014 03:33, Jay Pipes <jaypipes at gmail.com> wrote:
>> While I agree with the message, if cloud provider A has VM restarts
>> every hour, and B has restarts every 6 months, all other things being
>> equal I'm going to go with B.
>
> Pretty sure James wasn't saying that he restarts VMs every hour. The
> idea is that applications that run on a utility cloud should be
> resilient and take into account failure as an expected part of
> operating.

A certain amount of reduxio absurdium involved in my post, but the point stands.

>> Restarts are a pain point for most
>> systems, requiring data resynchronisation etc, so looking to minimise
>> them is a good aim as long as it doesn't conflict much with other
>> concerns...
>
> I'm actually not entirely sure what restarts and data resync have to do
> with vm-level HA? What am I missing here?

Not so much to do with VM level H/A as the general danger of the
cattle .v. pets argument taken too far. I've been in summit sessions
before where people have seriously argued that we shouldn't bother
with rolling upgrade or live migration of VMs because people should
expect their VMs to die at any time, and shutting down all of the VMs
on a cloud for an upgrade is a totally reasonable thing to do. While I
agree that unexpected VM death should be something you expect
occasionally, I am also aware that this approach has a cost for the
customer, in terms of complexity, reliability and performance, and a
good cloud provider should look at what steps can reasonably be taken
to avoid VMs dying, where practical, else another cloud provider who
does take such steps will end up with more business.


--
Duncan Thomas