camel-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Claus Ibsen <claus.ib...@gmail.com>
Subject Re: [DISCUSS - Camel 2.0 - Internal API reworkings] - Channel and AsyncProcessor
Date Fri, 24 Apr 2009 11:59:09 GMT
On Fri, Apr 24, 2009 at 10:17 AM, Willem Jiang <willem.jiang@gmail.com> wrote:
>
> +1 for introducing Channel into Camel world.
>
> For the AsyncProcessor part, it's is really difficult to understand.
> I think it is more than Call back, it will not block the calling thread
> which will be used in the camel-jhc component[1].
> And I also did an enhancement[2] on the ErrorHandler of
> DeadLetterChannel by leveraging the AsyncProcessor to avoid the blocking
> the calling thread when the DeadLetterChannel waits for another retry.
>
> If the UnitOfWork can address these, I'm happy to give my +1 for
> removing the AsyncProcessor. If not, we need to find out a replacement
> before removing it.
>
> [1] http://cwiki.apache.org/CAMEL/asynchronous-processing.html
> [2] https://issues.apache.org/activemq/browse/CAMEL-1129

Yeah the AsyncProcessor and the callback notion currently implemented
in Camel is flawed, brittle, broken, not used at all from end user
perspective
and it started to spread out into the code base where it should not have.

I really think it should be removed, and if we need it, a new Async
API introduced that resemble the JDK async support (for instance the
Future).

To fix it I really think we need to
1) Remove the existing AsyncProcessor
2) At a later stage add a new API that leverages the JDK API.


About redelivery using different threads
=============================
Note the using another thread for redeliver is only feasable for non
transacted routes.
Transacted routes tend to depend on ThreadLocal and reuse session and
other stuff.

So the redeliver using different thread is not always desirable.
Note: But the redelivery is only supported by DeadLetterChannel that
does not support transacted routes anyway.
So you can say we are safe here.
But the end user does not have any choice to configure what they want.

I do think that end users should have the choice which model they
would like Camel to use for redelivery handling.

And with the Channel we would have another possiblity for redelivery
as we could just "route the message" back to the previous channel
and let it have some "delay time" before its "visible" on the channel
queue for re-processing.

So Willam, it would be much easier to impl. the use different thread
for redelivery when the Channel is more enhanced in Camel.
For instance for InOnly routes where there are no caller waiting we
can safely use others threads for redeliver (if not transacted)
But for InOut we need to add our own barrier that waits until the
UnitOfWork is complete before we can return the response.

Again I cannot stress too much that the camel-core API needs this
cleanup. You will get to this conclusion if you have been working 8h
around the clock in the camel-core code for as long as I have.



>
>
> Willem
>
>
> Claus Ibsen wrote:
>> Hi Camel riders
>>
>> As you know we are working on Camel 2.0 and we decided to do a 2nd
>> milestone release.
>> This gives us more time to start on some of the ideas and internal
>> refactorings I wanted
>> to do for Camel 2.1.
>>
>> After having dived really deed in the camel core codebase for the last
>> 6+ months or longer, I
>> feel that we need to cease the moment and do a more extensive house
>> cleaning before the Camel 2.0
>> release. The house cleaning is only internal and will not affect end
>> users of Camel.
>>
>> Here are 2 issues I would like to address in the foreseeable future.
>>
>>
>> 1) Channel
>> ==========
>> To introduce a Channel between each node in the route path. The
>> channel is a composite processor
>> that is responsible for routing the exchange to the next node in the path.
>>
>> We already do this but there is no visible notion of a Channel. The
>> "Channel" exists today as
>> a series of interceptors and an error handler. Today this is wrapped
>> at route build time
>> and then we have a static list of processors the exchanges is routed
>> through a runtime.
>>
>> So what I am working on is to composite this series of interceptors
>> and error handler into a
>> DefaultChannel class that has the logic to do needed work before an
>> Exchange is routed to the next
>> node.
>>
>> I have already discussed this a bit with Gert and James. James agreed
>> that it a was time for introducing
>> the Channel into Camel.
>>
>> So I have started experimenting with this and got it working locally.
>> The Channel is in this first
>> stage just wrapping the existing series of interceptors and error
>> handler and thus the route logic
>> is not affect.
>>
>> In the future the Channel will benefit in many ways as we can at
>> runtime add routing logic for instance
>> - enabling/disabling tracing (and other interceptors) at runtime
>> without reloading routes
>> - adding or removing interceptors at runtime without reloading routes
>> - adding or removing JMX performance metrics at runtime without reloading routes
>> - we could use it to blocking routing
>> - enhanced tooling support
>> - potential you could persist exchanges at the channel
>> - later add true async processing with the help of the channel
>> - easier to traverse the runtime route graph as we can traverse the
>> Channel, where we can have next/prev or the like methods
>> - and much more we can imagine
>>
>>
>>
>> 2) AsyncProcessor
>> =================
>> I propose to remove the AsyncProcessor all together. I had a chat with
>> James about it and it was an experiment by Hiram back in early 2007.
>> Basically he had not worked on the code since, and its not really put
>> into good use. But sadly over time the AsyncProcessor have spread
>> itself into other core parts of Camel.
>>
>> The basic idea was to attach a callback to a route so you could do
>> some "commit" work after the exchange is finished.
>> This idea is actually superseded by the UnitOfWork (introduced by
>> James) that is better for this kind of work. And we have schduled an
>> overhaul for UnitOfWork
>> in Camel 2.1 to allow more of our components to take advantage of this
>> and also expose DSLs for end users to attach their custom processors
>> or route.
>>
>> The code in Camel core that is affected by the AsyncProcessor is much
>> more complex than regular Processor. In fact there are some areas
>> where its not
>> used correctly and causes unforseen side effects that only surfaces in
>> some complex or rare unit tests. This issues is more apparent lately
>> as more
>> and more camel processors supports the async processor directly.
>>
>> The original code by Hiram only leverages the async callback in the
>> file component as it was part of his initial code. No other areas
>> benefits from this.
>> The code is basically making it bloat and complex inside Camel itself.
>>
>> End users do not use this. I do not recall a single question on it in
>> the user forum.
>>
>> So I am in a big +1 to get it out of the codebase.
>>
>> And before you say well how do we do async routing then? Well the seda
>> component is still there to truely spawn a new thread to route an
>> exchange.
>> And we have the UnitOfWork where you are supposed to register your
>> callback. We might make the UnitOfWork a bit easier to register
>> callbacks for
>> only commit or failure depending on what you want. If its a single
>> methods its also easier for dynamic languages to have nice DSL support
>> for it.
>> Well that is just me rambling now..... And the UnitOfWork was James
>> envision for doing async work after the exchange is ended.
>>
>>
>> Summary
>> =======
>>
>> 1) Channel
>> The first cut of this is already done on my local laptop and it will
>> just work as is today. But its bases the foundation for more exotic
>> and advanced
>> stuff in Camel 2.x series.
>>
>> Already in progress.
>>
>> Any thoughts?
>>
>>
>> 2) AsyncProcessor
>> To reduce the complexity for maintaining the Camel codebase. Just as
>> Gert if he was a bit puzzled how to do his StreamCache that uses the
>> Async stuff.
>> When we get it out of there the codebase is much easier to maintain
>> and we have a greater chance to attract potential new committers.
>>
>> I vote +1 for this.
>> Any thoughts?
>>
>>
>>
>>
>
>



-- 
Claus Ibsen
Apache Camel Committer

Open Source Integration: http://fusesource.com
Blog: http://davsclaus.blogspot.com/
Twitter: http://twitter.com/davsclaus
Apache Camel Reference Card:
http://refcardz.dzone.com/refcardz/enterprise-integration

Mime
View raw message