On Wed, Jul 08, 2009 at 05:46:43PM +0100, Daniel P. Berrange wrote:
On Wed, Jul 08, 2009 at 12:10:18PM -0400, Hugh O. Brock wrote:
Hello all.
I spent quite a while yesterday in a meeting with a client interested in using libvirt for a large project.
They are quite happy in general with the direction things are going. However they have objections to a couple of libvirt design points that have come up in multiple cases in the past, including in developing the other project I'm involved in, oVirt. Specifically, they would really prefer that libvirtd be a one-stop shop for everything they need to do on a virtualization host, including things we have traditionally held out-of-scope for libvirt. A partial list of those things would include:
* In-depth multipath config management * Hardware lifecycle management (power-off, reboot, etc.) * HA configuration
... and pretty much anything else you might want to do on a piece of hardware that is hosting virtual machines.
My initial reaction of course was to tell them "Well you should just use a separate agent for that sort of thing." But of course this means making a separate connection to the node depending on what operation you want to perform, which is cumbersome.
So I guess what I'm asking is, why *not* expand the scope of libvirtd to be a one-stop shop for managing a node? Is there a really good reason it shouldn't have the remaining capabilities libvirt users want? Is that reason good enough to balance the headache our users have to deal with in coming up with a separate package to handle the tasks libvirtd does not?
This is essentially suggesting that libvirtd become a general purpose RPC layer for all remote management tasks. At which point you have just re-invented QPid/AMQP or CIM or any number of other general purpose message buses.
libvirtd has a core well defined goal:
- Provide a remote proxy for libvirt API calls
if you want todo anything more than that you should be considering an alternative remote management system. We already have 2 good ones to choose from supported with libvirt
- QPid/AMQP, with libvirt-qpid agent + your own custom agents - CIM, with libvirt-CIM + your own custom CIM providers
Both of these offer other benefits besides just pluggable support for other functionality. In particular
- Non-blocking asynchronous RPC calls - Assured delivery for RPC calls - Scalable network architecture / topology - Inter-operability with plugins written by other projects/vendors
Furthermore, adding more plugins to libvirtd means we will never be able to reduce its privileges to an acceptable level, because we'll never know what capabilities the plugins may want.
I understand your point -- certainly we want to use existing RPC mechanisms for libvirt and node management, not maintain our own. However, given a libvirt-qpid daemon on the node that handles RPC over QMF (for example), is there not some value in having libvirt expose a consistent API for the operations people want to do on a host regardless of whether they have directly to do with managing a virtual machine or not? I will note that when I presented the large client with the option of QMF talking to multiple agents on the node but exposing (effectively) a single API and a single connection, they seemed much happier. So perhaps the right way to attack this is with the ovirt-qpid daemon we are currently working on. Daniel V., any further thoughts on this? --Hugh