qpid-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From CLIVE <cl...@ckjltd.co.uk>
Subject Re: Message API - Real world usage issue
Date Tue, 11 Feb 2014 19:34:53 GMT
Ted,

Thanks for the response and your comments.

I have had to handle the case of multiple Receivers attaching to the 
same queue on several occasions; primarily because the customer has felt 
that it was easier to handle one queue with multiple bindings (up to 
100), rather than having a hundred queues with single bindings; message 
order was also a contributing factor.

The point of the post was just to raise it as a possible issue for 
future improvement..

I carried out a straw poll of 10 developers today at work. I gave them 
the two examples previously described and asked what they would expect 
to happen for the case where multiple Receivers were created for the 
same queue. They all expected the correct Receiver to be returned from 
the nextReceiver method, not the undeterministic behaviour that they 
would see.

I wouldn't have thought that it would take that much code/effort to add 
some additional functionality in the messaging API Implementation to 
support the behavior that, it would appear, most developers would expect 
to see. If I find some time I will take a look and see how it could be done.

Clive


On 10/02/2014 22:17, Ted Ross wrote:
> Clive,
>
> What you are observing is what I expect:  In the second scenario where 
> you use the same queue for each of the three receivers, the receiver 
> that receives any particular message will be non-deterministic.
>
> This is because the binding key is applied between the exchange and 
> the queue (i.e. it is used to determine which queue(s) the message 
> should be enqueued on).  Multiple receivers on a queue will receive 
> messages from the queue in an undetermined order, but no message shall 
> be delivered to more than one receiver.  In the second case, all of 
> the messages are placed on the same queue in the order in which they 
> arrive.  The queue acts as a buffer between the routing rule that 
> matched the message and the receiver that provided the routing rule.
>
> It would be simpler to do the following:
>
>   Rxer 1 - "amq.topic/bill; {link: {x-declare: {auto-delete:true}}}"
>   Rxer 2 - "amq.topic/ben; {link: {x-declare: {auto-delete:true}}}"
>   Rxer 3 - "amq.topic/tim; {link: {x-declare: {auto-delete:true}}}"
>
> This will give you the determinism you want.  This will cause the 
> creation of a temporary queue for each receiver that will receive the 
> messages that match the topic key (following the slash in the address).
>
> -Ted
>
>
> On 02/10/2014 04:39 PM, CLIVE wrote:
>> Fraser,
>>
>> Thanks for the response. The real problem is that the behavior of a 
>> Receiver is different depending on the multiplicity of the binding 
>> strategy used. If you use a single queue with a single binding then 
>> messages will get delivered to the required receiver. If you use 
>> multiple Receivers bound to the same queue, the Receiver called by 
>> the messaging API, when delivering a message to your application, may 
>> not be the one that you think!!
>>
>> So if I create three Receivers in the same application, with the 
>> following bindings (note unique queue names)
>>
>>    Rxer 1 - "queue1; {create: receiver, node: 
>> {x-declare:{auto-delete:true}, x-bindings: [{exchange: 'amq.topic', 
>> queue: 'queue1', key: 'bill'}]}}"
>>    Rxer 2 - "queue2; {create: receiver, node: 
>> {x-declare:{auto-delete:true}, x-bindings: [{exchange: 'amq.topic', 
>> queue: 'queue1', key: 'ben'}]}}"
>>    Rxer 3 - "queue3; {create: receiver, node: 
>> {x-declare:{auto-delete:true}, x-bindings: [{exchange: 'amq.topic', 
>> queue: 'queue1', key: 'tim'}]}}"
>>
>> And then send a message on the amq.topic exchange with a subject of 
>> 'tim'. Then Rxer3 will get returned by the 'nextReceiver' method on 
>> the associated Session object.
>>
>> But if I change the bindings so they related to the same queue
>>
>>    Rxer 1 - "queue1; {create: receiver, node: 
>> {x-declare:{auto-delete:true}, x-bindings: [{exchange: 'amq.topic', 
>> queue: 'queue1', key: 'bill'}]}}"
>>    Rxer 2 - "queue1; {create: receiver, node: 
>> {x-declare:{auto-delete:true}, x-bindings: [{exchange: 'amq.topic', 
>> queue: 'queue1', key: 'ben'}]}}"
>>    Rxer 3 - "queue1; {create: receiver, node: 
>> {x-declare:{auto-delete:true}, x-bindings: [{exchange: 'amq.topic', 
>> queue: 'queue1', key: 'tim'}]}}"
>>
>> And send the same message again, Which Receiver would you expect to 
>> get returned from the sessions nextReceiver method?
>>
>> I would expect the same result as in the first example, Rxer 3. But 
>> this does not happen, anyone of the three receivers might get called.
>>
>> This doesn't seem right to me and as a result you have to produce 
>> quite a bit of application level logic to handle this scenario; 
>> especially when your bindings are being passed down to you 
>> dynamically by several client applications.
>>
>> Hope this explains it a bit better than my last attempt.
>>
>> Clive
>>
>>
>> On 07/02/2014 10:03, Fraser Adams wrote:
>>> On 06/02/14 19:07, CLIVE wrote:
>>>> Hi,
>>>>
>>>> [snip]
>>>>
>>>> The first use case requires the dynamic creation of Receivers, but 
>>>> before creating a new receiver, I would like to know if I already 
>>>> have a receiver that would match the required binding. This is not 
>>>> possible at the moment because the binding matching algorithms are 
>>>> hidden from public view; they are buried deep inside the Brokers 
>>>> Exchange Implementation code.
>>> You know that you can get the binding information from QMF don't you 
>>> Clive? I guess I'm missing what you're looking for if it's something 
>>> different than that. And I guess to be fair to get the information 
>>> via QMF you'd need a bit of code, but I'd have thought that this 
>>> would be the most appropriate way to get the information.
>>>
>>>
>>> Out of curiosity why do you need to know if you already have a 
>>> receiver that would match the binding?
>>>
>>> One thing that's worth mentioning, I'm suspecting that (like me) 
>>> you've mainly been using AMQP 0.10 - If I'm reading you correctly 
>>> you sound like you are dynamically creating queue nodes and passing 
>>> x-bindings.
>>>
>>> I've been doing that for a few years, but a few weeks back I started 
>>> looking at AMQP 1.0 and that primarily takes a perspective of 
>>> addressing the topic like exchanges and the queues end up being 
>>> subscription queues and all of the stuff that relates to binding and 
>>> the like ends up in the link (not node) config.
>>>
>>> For me at any rate that was quite a different perspective on things 
>>> (I wrote up what I was up to in the "A write up of some AMQP 1.0 
>>> Experiments" post) previously I've been focussing on the queues, so 
>>> I was dynamically creating queue nodes and passing x-bindings in 
>>> AMQP 0.10, but in AMQP 1.0 I've been addressing the exchanges (topic 
>>> type nodes) and using the link to specify what I need. For me it 
>>> took a bit of getting used to because I was so ingrained doing it 
>>> the other way, but I think I'm getting it now.
>>>
>>>
>>>>
>>>> The second use case in question requires a client application to 
>>>> dynamically create multiple receivers for the same queue, but with 
>>>> slightly different binding keys bound to an exchange. When a 
>>>> message from an exchange gets put in the queue and delivered to the 
>>>> client (via a receiver)
>>> I'm not sure if I'm correctly interpreting what you are saying here, 
>>> so you want a client that has a single queue, but each receiver adds 
>>> different binding keys right? You do know that this will result in 
>>> what amounts to an OR condition - both keys will be bound and a 
>>> message will be put on the queue if either match so consumer A of 
>>> the queue would receive messages due to consumer B's key - is that 
>>> what you mean.
>>>
>>> The following AMQP 1.0 consumers will do what you seem to be saying, 
>>> there's a single shared subscription queue called queue1, the first 
>>> consumer binds *.news the second *.weather
>>>
>>> ./drain --connection-options {protocol:amqp1.0} -b localhost -f \
>>> "amq.topic/*.news; {node: {capabilities: [shared]}, link: {name: 
>>> queue1}}"
>>>
>>> ./drain --connection-options {protocol:amqp1.0} -b localhost -f \
>>> "amq.topic/*.weather; {node: {capabilities: [shared]}, link: {name: 
>>> queue1}}"
>>>
>>> qpid-config -r queues gives
>>>
>>> Queue 'queue1'
>>>     bind [queue1] => ''
>>>     bind [*.news] => amq.topic
>>>     bind [*.weather] => amq.topic
>>>
>>>
>>> For AMQP 0.10 the following would create a similar effect (not sure 
>>> if you want auto delete or not, if not remove the x-declare below 
>>> and for the AMQP 1.0 example above add reliability: at-least-once to 
>>> the link Map)
>>>
>>> ./drain -b localhost -f \
>>> "queue1; {create: receiver, node: {x-declare:{auto-delete:True}, 
>>> x-bindings: [{exchange: 'amq.topic', queue: 'queue1', key: 
>>> '*.news'}]}}"
>>>
>>> ./drain -b localhost -f \
>>> "queue1; {create: receiver, node: {x-declare:{auto-delete:True}, 
>>> x-bindings: [{exchange: 'amq.topic', queue: 'queue1', key: 
>>> '*.weather'}]}}"
>>>
>>>
>>> The following also works for AMQP 0.10
>>>
>>> ./drain -b localhost -f \
>>> "queue1; {create: receiver, node: {x-declare:{auto-delete:True}}, 
>>> link: {x-bindings: [{exchange: 'amq.topic', queue: 'queue1', key: 
>>> '*.news'}]}}"
>>>
>>> ./drain -b localhost -f \
>>> "queue1; {create: receiver, node: {x-declare:{auto-delete:True}}, 
>>> link: {x-bindings: [{exchange: 'amq.topic', queue: 'queue1', key: 
>>> '*.weather'}]}}"
>>>
>>> Don't know if this is what you are looking for.
>>>
>>>
>>> Note that in none of the cases above have I worked out how to remove 
>>> a binding other than by removing the queue so if you add the first 
>>> then the second then delete the second both bindings remain in place 
>>> - I did wonder about putting the x-declare/auto delete stuff on the 
>>> link in the second AMQP 0.10 example, but that doesn't seem to 
>>> remove the binding, so I'm not sure if that's possible.
>>>
>>>
>>>> I need to route the message to the correct application level 
>>>> destination(s). To do this I need to undertake a matching operation 
>>>> between the routing key of the message and the binding key(s) of 
>>>> the created receivers; qpid does not deliver the message to the 
>>>> receiver with the most exact binding key match.
>>> I guess than I'm not understanding you here. As far as I'm aware if 
>>> you've got multiple bindings between an exchange and a queue then 
>>> the message will be delivered on to the queue if either binding 
>>> matches, so it behaves like a logical OR. In your scenario if the 
>>> first receiver adds *.news then the second adds *.weather then from 
>>> that point on they will *both* start to receive (*.news OR *.weather)
>>>
>>>
>>>> So basically the receivers, and their bindings, enable the required 
>>>> messages to get delivered to the required client, but I then need 
>>>> to undertake application level routing to route the message to one 
>>>> or more application level classes, based on message routing key/ 
>>>> receiver binding key matches.
>>> So I'm still totally baffled why you want to send them to the same 
>>> queue if you are then demultiplexing at the application level. 
>>> Surely (for example) you'd be better having a news queue for the 
>>> *.news messages and a weather queue for the *.weather messages. If 
>>> you force them down the same queue then you are going to have to do 
>>> application level demultiplexing, which it sounds like you don't 
>>> want to do, but why use a single queue.
>>>
>>> What's actually driving the single queue requirement? That sounds 
>>> like the root of your problems, without knowing the nuance of your 
>>> scenario it feels like your approaching the problem from the wrong 
>>> angle and fighting the middleware rather than letting it work for 
>>> you. I'm sure I've missed something subtle in your use case.
>>>
>>>
>>>>
>>>> Unfortunately in both cases the messaging API does not provide 
>>>> visibility of the bind matching algorithms and so I have to create 
>>>> several utility classes to support this functionality.
>>>>
>>>> Would it be possible to create a Binding.h class in the messaging 
>>>> API to support matching of bindings from all the supported exchange 
>>>> types?
>>>
>>> I'm not actually sure what you are asking for here. Are you asking 
>>> for a client side filtering API?
>>>
>>> As I say I'm having trouble getting under the skin of your use case. 
>>> If I'm reading it correctly it sounds like you are wanting to have a 
>>> single queue but have multiple bindings between an exchange and that 
>>> queue, which will result in messages for both bindings making their 
>>> way on to the queue and then, to get around that, to apply a client 
>>> side filter to deliver the right message to the right receiver - is 
>>> that correct?
>>>
>>> I'm afraid that I'm still not clear why you want to do that on the 
>>> client rather than on the broker??
>>>
>>>
>>>
>>> Other's might have a better view, but I'm not sure that client side 
>>> filtering fits into the qpid::messaging API per se (and binding 
>>> probably wouldn't be a good idea anyway as it's a legacy AMQP 0.10 
>>> concept).
>>>
>>>
>>> One thought moving forward (and I'm far from an expert) might be to 
>>> think in terms of AMQP 1.0, so the Qpid Broker may be viewed as 
>>> essentially an AMQP 1.0 container and it has a whole bunch of 
>>> capabilities, including the ability to filter (the traditional 
>>> bindings plus - really cool - message selectors). The 
>>> qpid::messaging API is about interacting with nodes on a container 
>>> and attaching links with specified properties.
>>>
>>> As it happens though an AMQP 1.0 client application can also be 
>>> thought of as a container, so an interesting thought might be a 
>>> client application containing its own addressable node. In this 
>>> scenario you'd establish all the stuff previously discussed with the 
>>> broker and the consumer client would have all messages delivered to 
>>> the node on the client, you could then (in theory 'cause none of 
>>> this exists) create AMQP links (on the client) to the node (on the 
>>> client) passing filter properties on attachment (such as a selector).
>>>
>>> As I say none of this exists at the moment (except on the broker) 
>>> but it might be interesting to consider if it would be possible to 
>>> modularise things such that some of these fairly general purpose 
>>> AMQP 1.0 "services" could be extracted from the broker and made 
>>> available as a toolkit for creating general purpose AMQP 1.0 
>>> containers.
>>>
>>> As I say I'm no expert and tentatively finding my feet with AMQP 
>>> 1.0, Gordon Sim would be far better placed than I to say whether 
>>> that a) makes sense from an AMQP 1.0 perspective b) how feasible it 
>>> is and c) how likely it is to happen :-)
>>>
>>> Hope I've managed to be at least some help Clive,
>>> Cheers,
>>> Frase
>>>
>>>
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
>>> For additional commands, e-mail: users-help@qpid.apache.org
>>>
>>> .
>>>
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
>> For additional commands, e-mail: users-help@qpid.apache.org
>>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
> For additional commands, e-mail: users-help@qpid.apache.org
>
> .
>


---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
For additional commands, e-mail: users-help@qpid.apache.org


Mime
View raw message