openstack team mailing list archive

[Eric Day <eday@xxxxxxxxxxxx>]

On Mon, Jul 11, 2011 at 06:31:14PM +0200, Soren Hansen wrote:
> >> This is only a real problem if you insist on generating them in real
> >> time rather than pre-allocate them. Each compute node could have pool
> >> of thousands of ID's it could use as it pleased. That would still
> >> allow for millions of compute nodes. The ID's could be centrally
> >> assigned. Even if the central component that hands them out goes away,
> >> a couple of thousand ID's should provide ample time for a replacement
> >> to be spun up (or for the original one to come back).
> > This assumes you still have something you can call "central". What
> > about hybrid or peer-to-peer clouds?
> 
> How would you share these keyspaces anyway (be they uuid, arbitratry
> strings, urls, 32 bit integers, whateveR)? You can only rely on their
> global uniqueness if you actually trust the entities generating them.

Agreed, anyone could inject UUIDs that collide. UUIDs alone are not
sufficient, you need a namespace prefix as well (something I brought
up many times before on other ID threads). The full ID needs to be
something like:

nova-<account>-<instance uuid>

Or something along those lines (service and account/namespace
can be another part of a URL, it doesn't need to be the ID string
itself). Swift already does this (account/container/object), so we
have a pretty good example to follow here.

-Eric