Valentin, I understand your point and it makes sense to me. To summarize discussion I'll do the following. Leave timeout, use LT for spammed stack traces and file separate ticket for serialization issue in calculations as suggested Semyon.
On Wed, Nov 16, 2016 at 1:11 AM, Valentin Kulichenko < [email protected]> wrote: > Dmitry, > > I absolutely agree with this and I'm not saying that timeout solves any of > these problems. However, it provides a bit of safety for such cases (client > continues working instead of hanging forever) and also allows not too wait > for the service too long which can be also useful sometimes (e.g. service > is redeployed and initialization takes a lot of time for some reason). It > makes perfect sense to have it and should not be difficult to add. Am I > missing something? > > -Val > > On Tue, Nov 15, 2016 at 1:12 PM, Dmitry Karachentsev < > [email protected]> wrote: > > > Valentin, as I understand, this situation is abnormal, and if user thread > > hangs on getting service proxy, it's worth to check logs for failure > > instead of making workaround. > > > > > > On 15.11.2016 19:47, Valentin Kulichenko wrote: > > > >> I would still add a timeout there. In my view, it makes sense to have > such > >> option. Currently user thread can block indefinitely or loop in > >> while(true) > >> forever. > >> > >> -Val > >> > >> On Tue, Nov 15, 2016 at 7:10 AM, Dmitriy Karachentsev < > >> [email protected]> wrote: > >> > >> Perfect, thanks! > >>> > >>> On Tue, Nov 15, 2016 at 5:56 PM, Vladimir Ozerov <[email protected] > > > >>> wrote: > >>> > >>> To avoid log pollution we usually use LT class (alias for > >>>> > >>> GridLogThrottle). > >>> > >>>> Please check if it can help you. > >>>> > >>>> On Tue, Nov 15, 2016 at 5:48 PM, Dmitriy Karachentsev < > >>>> [email protected]> wrote: > >>>> > >>>> Vladimir, thanks for your reply! > >>>>> > >>>>> What you suggest definitely makes sense and it looks more reasonable > >>>>> solution. > >>>>> But there remains other thing. The second issue, that this solution > >>>>> > >>>> solves, > >>>> > >>>>> is prevent log pollution with GridServiceNotFoundException. The > reason > >>>>> > >>>> why > >>>> > >>>>> it happens is because GridServiceProxy#invokeMethod() designed to > >>>>> > >>>> catch > >>> > >>>> it > >>>> > >>>>> and ClusterTopologyCheckedException and retry over and over again > >>>>> > >>>> unless > >>> > >>>> service is become available, but on the same time there are tons of > >>>>> > >>>> stack > >>> > >>>> traces printed on remote node. > >>>>> > >>>>> If we want to avoid changing API, probably we need to add option to > >>>>> > >>>> mute > >>> > >>>> that exceptions in GridJobWorker. > >>>>> > >>>>> What do you think? > >>>>> > >>>>> On Tue, Nov 15, 2016 at 5:10 PM, Vladimir Ozerov < > [email protected] > >>>>> wrote: > >>>>> > >>>>> Also we implemented the same thing for platforms some time ago. In > >>>>>> > >>>>> short, > >>>> > >>>>> job result processing was implemented as follows (pseudocode): > >>>>>> > >>>>>> // Execute. > >>>>>> Object res; > >>>>>> > >>>>>> try { > >>>>>> res = job.run(); > >>>>>> } > >>>>>> catch (Exception e) { > >>>>>> res = e > >>>>>> } > >>>>>> > >>>>>> // Serialize result. > >>>>>> try { > >>>>>> SERIALIZE(res); > >>>>>> } > >>>>>> catch (Exception e) { > >>>>>> try{ > >>>>>> // Serialize serialization error. > >>>>>> SERIALIZE(new IgniteException("Failed to serialize > result.", > >>>>>> > >>>>> e)); > >>>> > >>>>> } > >>>>>> catch (Exception e) { > >>>>>> // Cannot serialize serialization error, so pass only > string > >>>>>> > >>>>> to > >>> > >>>> exception. > >>>>>> SERIALIZE(new IgniteException("Failed to serialize result: > " > >>>>>> > >>>>> + > >>> > >>>> e.getMessage()); > >>>>>> } > >>>>>> } > >>>>>> > >>>>>> On Tue, Nov 15, 2016 at 5:05 PM, Vladimir Ozerov < > >>>>>> > >>>>> [email protected] > >>> > >>>> wrote: > >>>>>> > >>>>>> Dmitriy, > >>>>>>> > >>>>>>> Shouldn't we report serialization problem properly instead? We > >>>>>>> > >>>>>> already > >>>> > >>>>> had > >>>>>> > >>>>>>> a problem when node hanged during job execution in case it was > >>>>>>> > >>>>>> impossible > >>>>> > >>>>>> to deserialize the job on receiver side. It was resolved properly - > >>>>>>> > >>>>>> we > >>>> > >>>>> caught exception on receiver side and reported it back sender side. > >>>>>>> > >>>>>>> I believe we must do the same for services. Otherwise we may end up > >>>>>>> > >>>>>> in > >>>> > >>>>> messy API which doesn't resolve original problem. > >>>>>>> > >>>>>>> Vladimir. > >>>>>>> > >>>>>>> On Tue, Nov 15, 2016 at 4:38 PM, Dmitriy Karachentsev < > >>>>>>> [email protected]> wrote: > >>>>>>> > >>>>>>> Hi Igniters! > >>>>>>>> > >>>>>>>> I'd like to modify our public API and add > >>>>>>>> > >>>>>>> IgniteServices.serviceProxy() > >>>>> > >>>>>> method with timeout argument as part of the task > >>>>>>>> https://issues.apache.org/jira/browse/IGNITE-3862 > >>>>>>>> > >>>>>>>> In short, without timeout, in case of serialization error, or so, > >>>>>>>> > >>>>>>> service > >>>>>> > >>>>>>> acquirement may hang and infinitely log errors. > >>>>>>>> > >>>>>>>> Do you have any concerns about this change? > >>>>>>>> > >>>>>>>> Thanks! > >>>>>>>> Dmitry. > >>>>>>>> > >>>>>>>> > >>>>>>> > > >
