Re: [Engine-devel] Announcing a proof of concept REST API for VDSM

On Wed, Nov 30, 2011 at 09:14:16AM +0200, Itamar Heim wrote: > On 11/29/2011 11:36 PM, Adam Litke wrote: >> On Tue, Nov 29, 2011 at 12:54:44PM -0800, Chris Wright wrote: >>> * Adam Litke (agl(a)us.ibm.com) wrote: >>>> On Tue, Nov 29, 2011 at 11:39:08AM -0800, Chris Wright wrote: >>>>> * Adam Litke (agl(a)us.ibm.com) wrote: >>>>>> >>>>>> https://github.com/aglitke/vdsm-rest/ >>>>>> >>>>>> Today I am releasing a proof of concept implementation of a REST API for vdsm. >>>>>> You can find the code on github. My goal is to eventually replace the current >>>>>> xmlrpc interface with a REST API. Once completed, ovirt-engine could switch to >>>>>> this new API. The major advantages to making this change are: 1) VDSM will gain >>>>>> a structured API that conceptually, structurally, and functionally aligns with >>>>>> the ovirt-engine REST API, 2) this new API can be made public, thus providing an >>>>>> entry point for direct virtualization management@the node level. >>>>> >>>>> Adam, this looks like a nice PoC. I didn't see how API versioning is >>>>> handled. Any VDSM developers willing to review this work? >>>> >>>> Thanks for taking a look. I am not handling versioning yet. I think we can add >>>> a version field to the root API object. As for compatibility, we'll just have >>>> to decide on an API backwards-compat support policy. Would this be enough to >>>> handle versioning issues? We shouldn't need anything like capabilities because >>>> the API is discoverable. >>> >>> Right, that seems sensible. >>> >>> Did you find cases where RPC to REST resource mapping was difficult? >> >> I haven't yet fully implemented the current vdsm API but I suspect that certain >> calls (like the ones you mention below) will require some extensions to what I >> have available currently. The main missing piece is probably events and a nice >> polling API. Another big piece of work will be to rebase onto the newly >> redesigned storage model. >> >>> I could see something like migrate() plus migrateStatus() and >>> migrateCancel() needing to add some kind of operational state that to the >>> resource. And something like monitorCommand() which has both a possible >>> side-effect and some freefrom response... >> >> Hopefully monitorCommand will not be too bad, since vdsm should be asking >> libvirt for the VM details when they are needed. Of course we'll need to be >> testing to make sure we aren't keeping state around. Also, I would expect >> monitorCommand to 'taint' the VM in the eyes of the vdsm API (as it does for >> libvirt). >> >>>>> <snip> >>>>>> ovirt-engine wants to subscribe to asynchronous events. REST APIs do not >>>>>> typically support async events and instead rely on polling of resources. I am >>>>>> investigating what options are available for supporting async events via REST. >>>>> >>>>> I think typical is either polling or long polling. If it's a single >>>>> resource, then perhaps long polling would be fine (won't be a large >>>>> number of connections tied up if it's only a single resource). >>>> >>>> Not sure if this is what you are referring to, but I was thinking we could do a >>>> batch-polling mechanism where an API user passes in a list of task UUIDs and/or >>>> event URIs. The server would respond with the status of these resources in one >>>> response. I have some ideas on how we could wire up asynchronous events on the >>>> server side to reduce the amount of actual work that such a batch-polling >>>> operation would require. >>> >>> Oh, I just meant this: >>> >>> Polling (GET /event + 404 loop) >>> Long polling (GET + block ... can chew up a thread connection) >> >> Yep. And we can talk later about building an API for efficient, repeated >> polling. I wonder if the ovirt-engine guys could weigh in as to whether a REST > > cc-ing engine-devel... > >> interface with event polling would be acceptable to them. It is critical that >> we settle on an API that can become _the_ first-class vehicle for interacting >> with vdsm. > > i have two points for consideration around this: > 1. as the api between ovirt-engine and vdsm, I had a preference for the > bus like nature of QMF, as it would allow multiple ovirt-engine to load > balance handling of messages from the queue, and multiple consumers for > some messages (say, history service picking up the stats in parallel to > engine picking them, rather than copying them from engine). I tend to agree, using a bus like QMF between ovirt-engine and vdsm is an inherantly more scalable network architecture, since it avoids the need to have a direct point-to-point connection between the engine and every single node, instead you can build a resilient grid by stategically deploying QPid brokers. As compared to REST, it is also much better at coping with async events, since you can have a push, rather than pull, model which VDSM just puts events onto the bus as they're generated, to be lazily consumed by any remote nodes when desired. NB, I don't mean to imply that there should not *also* be a REST API for the node level. A REST API has the very compelling property that it is trivially accessible from anything with HTTP client support, which is basically everything in existance today. > 2. as node level api, i think a lightweight ovirt-engine managing a > single node and exposing the exact same REST API and behavior of the > multi-node ovirt engine would be easier to cosnume for someone that > wants to interact with a single node same way they would interact with > ovirt-engine. If the REST API is well specified, you wouldn't need to necessarily have a lightweight ovirt-engine deployed at the node level, you could just have VDSM implement the same REST specification natively, as is impl in the engine. Or perhaps provide a some shered code that can be plugged into both VDSM& the enegine to facilitate provision of the same REST interface.