On 07/17/2012 01:19 AM, Itamar Heim wrote:
On 07/09/2012 09:52 PM, Saggi Mizrahi wrote:
>
>
> ----- Original Message -----
>> From: "Itamar Heim" <iheim(a)redhat.com>
>> To: "Saggi Mizrahi" <smizrahi(a)redhat.com>
>> Cc: "Adam Litke" <agl(a)us.ibm.com>,
vdsm-devel(a)lists.fedorahosted.org
>> Sent: Monday, July 9, 2012 11:03:43 AM
>> Subject: Re: [vdsm] [RFC] An alternative way to provide a supported
>> interface -- libvdsm
>>
>> On 07/09/2012 05:56 PM, Saggi Mizrahi wrote:
>>> I don't think AMQP is a good low level supported protocol as it's a
>>> very complex protocol to set up and support.
>>> Also brokers are known to have their differences in standard
>>> implementation which means supporting them all is a mess.
>>>
>>> It looks like the most accepted route is the libvirt route of
>>> having a c library abstracting away client server communication
>>> and having more advanced consumers build protocol specific bridges
>>> that may have different support standards.
>>>
>>> On a more personal note, I think brokerless messaging is the way to
>>> go in ovirt because, unlike traditional clustering, worker nodes
>>> are not interchangeable so direct communication is the way to go,
>>> rendering brokers pretty much useless.
>>
>> but brokerless doesn't let multiple consumers which a bus provides?
> All consumers can connect to the host and *some* events can be
> broadcasted to all connected clients.
>
> The real question is weather you want to depend on AMQP's routing \
> message storing
> Also, if you find it preferable to have a centralized host (single
> point of failure) to get all events from all hosts for the price of
> some clients (I assume read only clients) not needing to know the
> locations of all worker nodes.
> But IMHO we already have something like that, it's called the
> ovirt-engine, and it could send aggregated events about the cluster
> (maybe with some extra enginy data).
>
> The question is what does mandating a broker gives us something that
> an "AMQP bridge" wouldn't.
> The only thing I can think of is vdsm can assume unmoderated vdsm to
> vdsm communication bypassing the engine.
> This means that VDSM can have some clustered behavior that requires no
> engine intervention.
> Further more, the engine can send a request and let the nodes decide
> who is performing the operation among themselves.
>
> Essentially:
>
> [ engine ] [ engine ]
> | | VS |
> [vdsm][vdsm] [ broker ]
> | |
> [vdsm][vdsm]
>
> *All links are two way links
>
> This has dire consequences on API usability and supportability. So we
> need to converge on that.
>
> There needs to be a good reason why the aforementioned logic code
> can't sit on a another ovirt specific entity (lets call it
> ovirt-dynamo) that uses VDSM's supported API but it's own APIs (or
> more likely messaging algorithms) are unsupported.
>
> [ engine ]
> | | |
> | [ broker ] |
> | | | |
> [vdsm]-[dynamo] : [dynamo]-[vdsm]
> Host A : Host B
>
> *All links are two way links
1. we have engine today 'in the path' to the history db. but it makes no
sense for engine to be aware of each statistic we want to keep in the
history db.
same would be for an event/stats correlation service.
they don't need to depend on each other for availability/redundancy.
2. we are already looking at quantum integration, which is doing engine
to nodes communication via amqp.
3. with somewhat of a forward looking - moving some scheduling logic
"down to vdsm" will probably mean we'll want one of the nodes to listen
to statistics and state from the other nodes.
to all of these, setting up a bus which allows multiple peer listeners
seems more robust
I'm still against developing a C level binding for amqp and rest support
over a codebase which is in python.
rest and amqp allow for both local and remote bindings in any language.
C bindings should/could be a parallel implementation, but they seem like
an unneeded overhead and complexity in the middle of the codebase.