Open Stack

On 12/09/2013 07:15 PM, Russell Bryant wrote:
> On 12/09/2013 12:56 PM, Gordon Sim wrote:
>>> In the case of Nova (and others that followed Nova's messaging
>>> patterns), I firmly believe that for scaling reasons, we need to move
>>> toward it becoming the norm to use peer-to-peer messaging for most
>>> things.  For example, the API and conductor services should be talking
>>> directly to compute nodes instead of through a broker.
>>
>> Is scale the only reason for preferring direct communication? I don't
>> think an intermediary based solution _necessarily_ scales less
>> effectively (providing it is distributed in nature, which for example is
>> one of the central aims of the dispatch router in Qpid).
>>
>> That's not to argue that peer-to-peer shouldn't be used, just trying to
>> understand all the factors.
>
> Scale is the primary one.  If the intermediary based solution is easily
> distributed to handle our scaling needs, that would probably be fine,
> too.  That just hasn't been our experience so far with both RabbitMQ and
> Qpid.

Understood. The Dispatch Router was indeed created from an understanding 
of the limitations and drawbacks of the 'federation' feature of qpidd 
(which was the primary mechanism for scaling beyond one broker) as well 
learning lessons around the difficulties of message replication and storage.

>> One other pattern that can benefit from intermediated message flow is in
>> load balancing. If the processing entities are effectively 'pulling'
>> messages, this can more naturally balance the load according to capacity
>> than when the producer of the workload is trying to determine the best
>> balance.
>
> Yes, that's another factor.  Today, we rely on the message broker's
> behavior to equally distribute messages to a set of consumers.

Sometimes you even _want_ message distribution to be 'unequal', if the
load varies by message or the capacity by consumer. E.g. If one consumer
is particularly slow (or is given a particularly arduous task), it may 
not be optimal for it to receive the same portion of subsequent messages 
as other less heavily loaded or more powerful consumers.

> One example is how Nova components talk to the nova-scheduler service.
> All instances of the nova-scheduler service are reading off a single
> 'scheduler' queue, so messages hit them round-robin.
>
> In the case of the zeromq driver, this logic is embedded in the client.
>   It has to know about all consumers and handles choosing where each
> message goes itself.  See references to the 'matchmaker' code for this.
>
> Honestly, using a distributed more lightweight router like Dispatch
> sounds *much* nicer.
>
>>>   The exception
>>> to that is cases where we use a publish-subscribe model, and a broker
>>> serves that really well.  Notifications and notification consumers
>>> (such as Ceilometer) are the prime example.
>>
>> The 'fanout' RPC cast would perhaps be another?
>
> Good point.
>
> In Nova we have been working to get rid of the usage of this pattern.
> In the latest code the only place it's used AFAIK is in some code we
> expect to mark deprecated (nova-network).

Interesting. Is that because of problems in scaling the messaging 
solution or for other reasons?

[...]
> I'm very interested in diving deeper into how Dispatch would fit into
> the various ways OpenStack is using messaging today.  I'd like to get
> a better handle on how the use of Dispatch as an intermediary would
> scale out for a deployment that consists of 10s of thousands of
> compute nodes, for example.
>
> Is it roughly just that you can have a network of N Dispatch routers
> that route messages from point A to point B, and for notifications we
> would use a traditional message broker (qpidd or rabbitmq) ?

For scaling the basic idea is that not all connections are made to the 
same process and therefore not all messages need to travel through a 
single intermediary process.

So for N different routers, each have a portion of the total number of 
publishers and consumers connected to them. Though client can 
communicate even if they are not connected to the same router, each 
router only needs to handle the messages sent by the publishers directly 
attached, or sent to the consumer directly attached. It never needs to 
see messages between publishers and consumer that are not directly attached.

To address your example, the 10s of thousands of compute nodes would be 
spread across N routers. Assuming these were all interconnected, a 
message from the scheduler would only travel through at most two of 
these N routers (the one the scheduler was connected to and the one the 
receiving compute node was connected to). No process needs to be able to 
handle 10s of thousands of connections itself (as contrasted with full 
direct, non-intermediated communication, where the scheduler would need 
to manage connections to each of the compute nodes).

This basic pattern is the same as networks of brokers, but Dispatch 
router has been designed from the start to simply focus on that problem 
(and not deal with all other broker related features, such as 
transactions, durability, specialised queueing etc).

The other difference is that Dispatch Router does not accept 
responsibility for messages, i.e. it does not offer any 
store-and-forward behaviour. Any acknowledgement is end-to-end. This 
avoids it having to replicate messages. On failure they can if needed by 
replayed by the original sender.

The Dispatch Router can work for pub-sub patterns as well, though not 
store and forward directly. In theory, for flows where store-and-forward 
is needed, that can be supplied by an additional service e.g. a more 
traditional broker, which would take responsibility for replaying over 
from the publisher in order that subscribers could if needed have 
message replayed even after the original publisher had exited.

--Gordon.