activemq-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From rodos77 <eugene.ro...@nexj.com>
Subject Re: prefetchExtension off-by-1 for transacted consumers with prefetchSize > 0
Date Mon, 22 Mar 2010 15:54:47 GMT

Ok, I have to finish some work first but I'll see how much I can get done
after that.  I'm not too familiar with the classes you mentioned but I'll
see what I can figure out.  

In principle, I can wait till 5.4 but I'll see what I can do.

There is no problem with AMQ RA, it's just that we have our own JMS RA that
is capable of integrating with any JMS engine among other features.  We use
this RA in our enterprise stack.



Gary Tully wrote:
> 
> If you are in the thick of it and can propose a patch that would be great.
> A new boolean option on ActiveMQPolicy would be the best way to introduce
> the configuration on a destination basis (set in
> org.apache.activemq.broker.region.policy.PolicyEntry.configure(Broker,
> SystemUsage, QueueSubscription) and if you want more fine grain control,
> an
> option can be added to ConsumerInfo to allow the configuration on a per
> subscription basis, propagating the change though to a connection and
> connection factory. That will require regenerating the wireformat
> marshallers (in activemq-core; mvn -Popenwire-generate generate-sources)
> so
> that the new option is propagated from the client. See how you get on.
> 
> Otherwise, we can try and get this change into the 5.4 release.
> 
> BTW: Is there a problem with the activemq RA that stops you from using it?
> 
> 
> On 22 March 2010 14:56, rodos77 <eugene.rodos@nexj.com> wrote:
> 
>>
>> I would point out that another JIRA that I've opened (
>> https://issues.apache.org/activemq/browse/AMQ-2652 AMQ-2652 ),
>> illustrates
>> a
>> deadlock (hang) which is triggered precisely by the prefetch extension of
>> transacted consumers (see JIRA for details).  So while trying to avoid
>> one
>> potential hang, the use of prefetch extension causes another.
>>
>> However, the solution that you propose to make the prefetch extension
>> configurable and to have it off by default sounds good to me.  How should
>> we
>> proceed to have this put in place?
>>
>>
>>
>> Gary Tully wrote:
>> >
>> > Fair point. From the perspective of the ResourceAdapter where there are
>> > explicit expectations this makes good sense. The prefetch should be
>> > written
>> > in stone.
>> >
>> > From a regular transacted consumers perspective, where the prefetch is
>> > largely hidden, avoiding a hang for large transactions is good because
>> a
>> > hang would be hard to diagnose. I agree it could be considered
>> user/error
>> > but it is difficult to flag it as such.
>> >
>> > It may make sense to make the (extend the prefetch for a
>> > transaction) behavior configurable such that it can be disabled for the
>> RA
>> > or off by default and enabled in the event that large transactions
>> hang.
>> > Thinking more, off by default is probably the best approach as
>> > transactions
>> > that exceed the prefetch should be rare and the RAR behavior would
>> match
>> > expectations.
>> >
>> > On 18 March 2010 22:05, rodos77 <eugene.rodos@nexj.com> wrote:
>> >
>> >>
>> >> I don't really get this.  If you extend the prefetchSize to more than
>> >> what
>> >> it
>> >> was configured to be by the maxMessages parameter in the
>> >> Connection.createConnectionConsumer() call, you are violating the JMS
>> >> spec.
>> >> The maxMessages parameter is defined as "the maximum number of
>> messages
>> >> that
>> >> can be assigned to a server session at one time", so how can you make
>> it
>> >> bigger than what it was set to by the ConnectionConsumer?!
>> >>
>> >> ActiveMQ Resource Adapter has 2 parameters in its ActivationSpec that
>> >> control the maxMessages argument used in the
>> >> Connection.createConnectionConsumer() call: maxSessions and
>> >> maxMessagesPerSessions.  The maxMessages argument is set to their
>> >> product.
>> >> This makes sense.  If a user configures ActiveMQ to have x maxSessions
>> >> and
>> >> y
>> >> maxMessagesPerSessions, maxMessages (and consequentially the
>> >> prefetchSize)
>> >> should be set to x*y.  But why should this ever be extended? 
>> Shouldn't
>> >> it
>> >> be expected that only a maximum of x*y messages be delivered at the
>> same
>> >> time?  If a user wants to process z <= x*y messages in the same tx,
>> that
>> >> should not pose any problem as the RA will have x ServerSessions each
>> >> capable of processing y messages at the same time.  But if the user
>> wants
>> >> to
>> >> process z > x*y messages at the same time after explicitly configuring
>> >> ActiveMQ to only deliver a maximum of x*y messages at the same time,
>> this
>> >> is
>> >> a user error/problem.  ActiveMQ should not be violating the JMS Spec
>> to
>> >> accommodate this.  What is the point of having those properties in the
>> >> 1st
>> >> place if you are going to extend them and deliver as many messages as
>> >> available?
>> >>
>> >>
>> >>
>> >> Gary Tully wrote:
>> >> >
>> >> > The prefetch extension in the transacted case is necessary for the
>> case
>> >> > where a transacted consumer wants to consumer more than prefetch
>> >> messages
>> >> > in
>> >> > a single transaction. Without the extension, the broker would see a
>> >> full
>> >> > prefetch and not dispatch any more and a receive could block.
>> >> >
>> >> > On 15 March 2010 21:31, rodos77 <eugene.rodos@nexj.com> wrote:
>> >> >
>> >> >>
>> >> >> JIRA AMQ-2651 opened.
>> >> >>
>> >> >> I would be happy to provide a patch but as I stated in my original
>> >> post,
>> >> >> I
>> >> >> don't fully understand the need for prefetchExtension in the case
>> of
>> a
>> >> >> transacted, asynchronous (prefetchSize > 0) consumer.  If somebody
>> >> could
>> >> >> explain it, that would perhaps give me the required knowledge to
>> come
>> >> up
>> >> >> with a patch.
>> >> >>
>> >> >> The way I see it now, I would get rid of the prefetchExtension
(in
>> the
>> >> >> above
>> >> >> mentioned case) all together, but again I'm probably missing some
>> >> piece
>> >> >> of
>> >> >> information and I don't want to cause any side effects.  Barring
>> that,
>> >> I
>> >> >> think there is an off-by-1 error in its calculation but I need
this
>> >> >> confirmed in order to create a proper patch.
>> >> >> --
>> >> >> View this message in context:
>> >> >>
>> >>
>> http://old.nabble.com/prefetchExtension-off-by-1-for-transacted-consumers-with-prefetchSize-%3E-0-tp27866123p27910673.html
>> >> >> Sent from the ActiveMQ - User mailing list archive at Nabble.com.
>> >> >>
>> >> >>
>> >> >
>> >> >
>> >> > --
>> >> > http://blog.garytully.com
>> >> >
>> >> > Open Source Integration
>> >> > http://fusesource.com
>> >> >
>> >> >
>> >>
>> >> --
>> >> View this message in context:
>> >>
>> http://old.nabble.com/prefetchExtension-off-by-1-for-transacted-consumers-with-prefetchSize-%3E-0-tp27866123p27950855.html
>> >> Sent from the ActiveMQ - User mailing list archive at Nabble.com.
>> >>
>> >>
>> >
>> >
>> > --
>> > http://blog.garytully.com
>> >
>> > Open Source Integration
>> > http://fusesource.com
>> >
>> >
>>
>> --
>> View this message in context:
>> http://old.nabble.com/prefetchExtension-off-by-1-for-transacted-consumers-with-prefetchSize-%3E-0-tp27866123p27987505.html
>> Sent from the ActiveMQ - User mailing list archive at Nabble.com.
>>
>>
> 
> 
> -- 
> http://blog.garytully.com
> 
> Open Source Integration
> http://fusesource.com
> 
> 

-- 
View this message in context: http://old.nabble.com/prefetchExtension-off-by-1-for-transacted-consumers-with-prefetchSize-%3E-0-tp27866123p27988380.html
Sent from the ActiveMQ - User mailing list archive at Nabble.com.


Mime
View raw message