camel-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Claus Ibsen <claus.ib...@gmail.com>
Subject Re: Aggregator lock
Date Wed, 18 Sep 2013 15:41:07 GMT
And you just need to set executorService and refer to a thread pool to
be shared / or a thread pool profile as template for settings.
http://camel.apache.org/threading-model.html

That is one option to configure which is not verbose or error prone!

On Wed, Sep 18, 2013 at 5:21 PM, Claus Ibsen <claus.ibsen@gmail.com> wrote:
> It works as designed.
>
> If you want aggregator to use concurrent threads for downstream then
> you need to configure that.
>
> On Wed, Sep 18, 2013 at 5:15 PM, Baris Acar <baris@acar.org.uk> wrote:
>> Hi Claus,
>>
>> Thanks for your reply! I've tried using parallelProcessing and it comes with a few
drawbacks as I've mentioned already. We're going with it as a workaround but I'm interested
to know whether you consider the issue I've reported to be a bug.
>>
>> Do you believe that it is intentional/expected that by default the AggregateProcessor
*holds a mutual exclusion lock* across all downstream processing, by default? It's really
very unexpected to me, and the docs you link to make no mention of acquiring a lock over other
code unrelated to the aggregation. As a user of a framework, one needs to know if a framework
is going to acquire a mutual exclusion lock over my code, in order to reason about the parallelism.
>>
>> Importantly - are there any other processors which acquire a lock over all downstream
processing?
>>
>> Barış
>>
>>
>> On 18 Sep 2013, at 11:54, Claus Ibsen <claus.ibsen@gmail.com> wrote:
>>
>>> Hi
>>>
>>> See the parallelProcessing / executorService option on the aggregator
>>> http://camel.apache.org/aggregator2
>>>
>>> On Wed, Sep 18, 2013 at 2:49 AM, Baris Acar <baris@acar.org.uk> wrote:
>>>> Hi,
>>>>
>>>> I'm seeing some surprising behaviour with my camel route, and was hoping
>>>> someone in this group could help, as my trawl through the docs and Camel
In
>>>> Action book have not found the answers I'm looking for. Apologies if this
>>>> question has been clearly answered elsewhere :-/
>>>>
>>>> I have a route that looks a little like the following:
>>>>
>>>>    from("seda:foo?concurrentConsumers=2")
>>>>      .aggregate(header("myId"), myAggregationStrategy).completionSize(5)
>>>>        .log("Sending out ${body} after a short pause...")
>>>>        .delay(3000) // simulate a lengthy process
>>>>        .log("Sending out ${body} imminently!")
>>>>        .to(...) // other downstream processing
>>>>
>>>> Note that I'm using a SEDA with two *concurrent* consumers. I expected that
>>>> once a SEDA consumer thread has picked up a message that completes an
>>>> aggregation, that downstream processing will continue on that consumer
>>>> thread, whilst other such downstream processing for another 'completed
>>>> aggregation' message may be happening in parallel on the other SEDA
>>>> consumer thread.
>>>>
>>>> What I'm finding instead is that whilst all of the work downstream of
>>>> aggregate() does occur across the two consumer threads, it is serialised;
>>>> no two threads execute the processors at the same time. This becomes quite
>>>> noticeable if this downstream work is lengthy. I've uploaded a sample to
>>>> https://github.com/bacar/aggregator-lock, which you can run with mvn test
>>>> -Dtest=AggregateLock. It started from a sample from the CIA book.
>>>>
>>>> For example, you can see the whilst the second "Sending... after a short
>>>> pause" does occur on a separate thread (#2), it does not start until after
>>>> thread #1 has completed, despite the 3s delay():
>>>>
>>>> 2013-09-18 00:45:15,693 [el-1) thread #1 - Threads] INFO  route1 - Sending
>>>> out aggregated [1:0, 1:1, 1:2, 1:3, 1:4] after a short pause...
>>>> 2013-09-18 00:45:18,695 [el-1) thread #1 - Threads] INFO  route1 - Sending
>>>> out aggregated [1:0, 1:1, 1:2, 1:3, 1:4] imminently!
>>>> 2013-09-18 00:45:18,696 [el-1) thread #2 - Threads] INFO  route1 - Sending
>>>> out aggregated [0:0, 0:1, 0:2, 0:3, 0:4] after a short pause...
>>>> 2013-09-18 00:45:21,698 [el-1) thread #2 - Threads] INFO  route1 - Sending
>>>> out aggregated [0:0, 0:1, 0:2, 0:3, 0:4] imminently!
>>>>
>>>> Is this behaviour expected? I found it _very_ surprising. Did I miss
>>>> something in the docs that describes this behaviour? If the behaviour is
>>>> expected, I am happy to try adding some info to the documentation if
>>>> someone can explain the intent behind it.
>>>>
>>>> I'm not terribly familiar with the code, but I've had a dig around, and it
>>>> looks like the reason for this behaviour is due to the following code
>>>> inside the process() method of
>>>> org.apache.camel.processor.aggregate.AggregateProcessor:
>>>>
>>>>            lock.lock();
>>>>            try {
>>>>                doAggregation(key, copy);
>>>>            } finally {
>>>>                lock.unlock();
>>>>            }
>>>>
>>>> The doAggregation() method performs both the aggregation (i.e., adding the
>>>> new exchange to the repository, checking if the completion criteria have
>>>> been met etc) _and_, if complete, submits the aggregated message to the
>>>> ExecutorService for downstream processing. However, since the default
>>>> executorService is the SynchronousExecutorService, all downstream
>>>> processing occurs synchronously with submission, and consequently, _within_
>>>> the lock above.
>>>>
>>>> Whilst I can see obvious reasons that may make it necessary to perform the
>>>> actual aggregation inside a lock, I do find it quite surprising that the
>>>> downstream processing by default also occurs inside this lock. Are there
>>>> any other processors known to behave in this way, i.e., by taking a lock
>>>> around all downstream processing?
>>>>
>>>> I could potentially work around this issue by dispensing with the SEDA
>>>> concurrentConsumers and using aggregate().parallelProcessing() instead,
>>>> with a suitable executorService() specified, but this introduces a number
>>>> of complications, e.g.:
>>>> - if I repeatedly split() and re-aggregate() (by different criteria), then
>>>> _every time_ I aggregate I have to add
>>>> parallelProcessing()/executorService(); this is verbose and error prone.
>>>> - with repeated aggregates in a route, I need dedicated threads/pools per
>>>> aggregate(), which means way more threads than I really want/need.
>>>> - regardless, I don't get the predictable and simple behaviour I expected
>>>> of 'pick up job from SEDA, aggregate, synchronously process downstream
>>>> jobs' that I'd expected.
>>>>
>>>> Another possible workaround might be the optimistic locking, but I haven't
>>>> had the opportunity to study it yet. It seems unrelated - I think my
>>>> problem is with the very coarse granularity of the pessimistic lock, not
>>>> with whether it's optimistic. Plus, I don't really want my messages to ever
>>>> fail with a 'too many attempts to acquire the optimistic lock' exception,
>>>> and I might have quite high contention).
>>>>
>>>> Many thanks in advance for your help/comments!
>>>>
>>>> Baris.
>>>
>>>
>>>
>>> --
>>> Claus Ibsen
>>> -----------------
>>> Red Hat, Inc.
>>> Email: cibsen@redhat.com
>>> Twitter: davsclaus
>>> Blog: http://davsclaus.com
>>> Author of Camel in Action: http://www.manning.com/ibsen
>
>
>
> --
> Claus Ibsen
> -----------------
> Red Hat, Inc.
> Email: cibsen@redhat.com
> Twitter: davsclaus
> Blog: http://davsclaus.com
> Author of Camel in Action: http://www.manning.com/ibsen



-- 
Claus Ibsen
-----------------
Red Hat, Inc.
Email: cibsen@redhat.com
Twitter: davsclaus
Blog: http://davsclaus.com
Author of Camel in Action: http://www.manning.com/ibsen

Mime
View raw message