On 18/05/17 20:19, Matt Riedemann wrote:
I just wanted to blurt this out since it hit me a few times at the
summit, and see if I'm misreading the rooms.
For the last few years, Nova has pushed back on adding orchestration to
the compute API, and even define a policy for it since it comes up so
much [1]. The stance is that the compute API should expose capabilities
that a higher-level orchestration service can stitch together for a more
fluid end user experience.
I think this is a wise policy.
One simple example that comes up time and again is allowing a user to
pass volume type to the compute API when booting from volume such that
when nova creates the backing volume in Cinder, it passes through the
volume type. If you need a non-default volume type for boot from volume,
the way you do this today is first create the volume with said type in
Cinder and then provide that volume to the compute API when creating the
server. However, people claim that is bad UX or hard for users to
understand, something like that (at least from a command line, I assume
Horizon hides this, and basic users should probably be using Horizon
anyway right?).
As always, there's a trade-off between simplicity and flexibility. I can
certainly understand the logic in wanting to make the simple stuff
simple. But users also need to be able to progress from simple stuff to
more complex stuff without having to give up and start over. There's a
danger of leading them down the garden path.
While talking about claims in the scheduler and a top-level conductor
for cells v2 deployments, we've talked about the desire to eliminate
"up-calls" from the compute service to the top-level controller services
(nova-api, nova-conductor and nova-scheduler). Build retries is one such
up-call. CERN disables build retries, but others rely on them, because
of how racy claims in the computes are (that's another story and why
we're working on fixing it). While talking about this, we asked, "why
not just do away with build retries in nova altogether? If the scheduler
picks a host and the build fails, it fails, and you have to
retry/rebuild/delete/recreate from a top-level service."
(FWIW Heat does this for you already.)
But during several different Forum sessions, like user API improvements
[2] but also the cells v2 and claims in the scheduler sessions, I was
hearing about how operators only wanted to expose the base IaaS services
and APIs and end API users wanted to only use those, which means any
improvements in those APIs would have to be in the base APIs (nova,
cinder, etc). To me, that generally means any orchestration would have
to be baked into the compute API if you're not using Heat or something
similar.
The problem is that orchestration done inside APIs is very easy to do
badly in ways that cause lots of downstream pain for users and external
orchestrators. For example, Nova already does some orchestration: it
creates a Neutron port for a server if you don't specify one. (And then
promptly forgets that it has done so.) There is literally an entire
inner platform, an orchestrator within an orchestrator, inside Heat to
try to manage the fallout from this. And the inner platform shares none
of the elegance, such as it is, of Heat itself, but is rather a
collection of cobbled-together hacks to deal with the seemingly infinite
explosion of edge cases that we kept running into over a period of at
least 5 releases.
The get-me-a-network thing is... better, but there's no provision for
changes after the server is created, which means we have to copy-paste
the Nova implementation into Heat to deal with update.[1] Which sounds
like a maintenance nightmare in the making. That seems to be a common
mistake: to assume that once users create something they'll never need
to touch it again, except to delete it when they're done.
Don't even get me started on Neutron.[2]
Any orchestration that is done behind-the-scenes needs to be done
superbly well, provide transparency for external orchestration tools
that need to hook in to the data flow, and should be developed in
consultation with potential consumers like Shade and Heat.
Am I missing the point, or is the pendulum really swinging away from
PaaS layer services which abstract the dirty details of the lower-level
IaaS APIs? Or was this always something people wanted and I've just
never made the connection until now?
(Aside: can we stop using the term 'PaaS' to refer to "everything that
Nova doesn't do"? This habit is not helping us to communicate clearly.)
cheers,
Zane.
[1] https://review.openstack.org/#/c/407328/
[2]
http://lists.openstack.org/pipermail/openstack-dev/2014-April/032098.html
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: [email protected]?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
