On 2/6/2013 8:47 AM, Sergey Beryozkin wrote:
> On 06/02/13 13:01, Bill Burke wrote:
>>
>>
>> On 2/5/2013 6:37 PM, Marek Potociar wrote:
>>>
>>> On Feb 5, 2013, at 8:59 PM, Bill Burke <bburke_at_redhat.com> wrote:
>>>
>>>>
>>>>
>>>> On 2/5/2013 1:53 PM, Marek Potociar wrote:
>>>>> I was under impression that leakage is implementation detail. We did
>>>>> not have to specify anything for sync case, why should we specify it
>>>>> for async case. There's nothing special about async case in this
>>>>> regard.
>>>>>
>>>>
>>>> In he sync case, the Servlet container is wrapping the JAX-RS
>>>> invocation. So, exceptions can be propagated and handled by the
>>>> servlet container if need be. In the async case, a non-servlet thread
>>>> is handling the response. So, cleanup, exception handling, etc. needs
>>>> to either be handled by the jax-rs layer, or user code. IMO, cleanup
>>>> and exception handling should be done in the JAX-RS layer and should
>>>> be as deterministic.
>>>
>>> Servlet is not the only platform JAX-RS runs on. Anyway, I think that
>>> propagating the exception to the IO container is deterministic enough.
>>> Whether or not this happens on the original thread is IMO orthogonal.
>>> All IO containers that support async externalise all the necessary
>>> connection-related context in one way or another and always provide a
>>> facility for propagating exceptions back to this context. I have yet
>>> to see a container that does not do that.
>>>
>>>>
>>>>> That said, I'm fine with adding a short sentence to the
>>>>> AR.resume()/cancel() javadoc stating that "A successful invocation
>>>>> of this method eventually completes the suspended request
>>>>> processing." or similar. FWIW, I do not want to write anything like
>>>>> "when this method completes, the request processing is complete",
>>>>> because I do not want to place restrictions on implementations. E.g.
>>>>> implementors may decide to complete the response processing
>>>>> asynchronously, which means that when the method returns the
>>>>> connection does not have to be necessarily closed yet. Which is also
>>>>> why I think it's better to not say anything...
>>>>>
>>>>
>>>> Yeah, that's ok. If you want to force synchronicity, then you can
>>>> just do it in the callback layer. It just wasn't very clear in the
>>>> javadoc what was supposed to happen with unmapped exceptions. In the
>>>> sync layer, the unmapped exception is rethrown. You just can't do
>>>> that in the async layer, IMO.
>>>
>>> Yes, that's the main difference. While in sync case the propagation is
>>> achieved by re-throwing and automatic rollback of the call stack, in
>>> async case this propagation needs to happen in some other way. Yet
>>> again, I have not seen any async-enabled IO container that would not
>>> expose such facility.
>>>
>>
>> Again, I think you are incorrect. If you do not have explicit language
>> then the user does not know if they have to wrap resume() calls in
>> try/catch blocks and call cancel. The current language of the javadoc
>> makes it seem like unmapped exceptions are thrown from resume() and that
>> the user needs to handle this. Which is why I started this thread in the
>> first place...
>>
>> IMO, at least the javadoc should change from:
>>
>> "The processing of the data by JAX-RS framework follows the same path as
>> * it would for the response data returned synchronously by a JAX-RS
>> resource
>> * method."
>>
>> To:
>>
>> "By executing this method, the request is guaranteed to be completed is
>> some form or another. The processing of the data by the JAX-RS framework
>> follows the same path as it would for the response data returned
>> synchronously by a JAX-RS resource except that unmapped exceptions are
>> not propagated. Depending on the JAX-RS implementation, unmapped
>> exceptions will result in an error status being sent to the client
>> and/or the connection being closed."
>>
>> I think the above gives jax-rs implementations enough flexibility, but
>> makes it clear to the user how to interact with resume()/cancel().
>>
> While I can not say I always agree with Marek, but the current version
> seems quite precise to me. Even in the sync case, the unmapped exception
> will result in 500 being sent (by the container, on the servlet thread),
> so why distinguish how this 500 may end up being reported by the
> container in the async case ?
>
Maybe clear to you, but not clear to me. The resume() javadoc states
that execution works the same as the sync case. If you read spec
section 3.3.4 on Exceptions, bullet #4 states that unmapped and
unhandled exceptions are wrapped and allowed to propagate to the
container. There *IS* no container to propagate to in the async case.
So, this leads me to believe that you have to have try/catch blocks
around resume() calls.
I think unmapped exceptions should be handled by resume() and cancel()
by either sending an error code back to the client or closing the
connection. This should be explicitly stated, or you'll have users like
me misinterpreting the javadoc and having try/catch blocks around our
resume() methods.
> Moreover, in the async case, one can capture the unmapped exception with
> CompletionCallback and do something else with it; though I've no idea at
> the moment what should happen if CompletionCallback rethrows it :-)
>
That is something I wasn't confused about. resume() should always
return true unless the request is not in a suspended state. It should
return true even if there was an error condition. If the user wants to
know of error conditions, they need to implement a CompletionCallback.
--
Bill Burke
JBoss, a division of Red Hat
http://bill.burkecentral.com