Hi Markus,
do I read correctly that you are suggesting replacing Source/Sink by
ByteBuffer? How is that reactive and non-blocking?
Please correct me if I'm mistaken, but ByteBuffer doesn't offer any
callback functionality, all it has is "remaining". We often don't know
how long the input/output is, so we'll just poll "remaining()" method?
Or do you suggest that the "#filter(Context, ByteBuffer)" would be
invoked multiple times?
Anyway, I don't really see how this proposal covers "reactive" usecase.
If you could elaborate little more, maybe including code snippets, it
could help to see what you mean.
Thanks and regards,
Pavel
On 16/03/2017 00:08, Markus KARG wrote:
> I'd love to keep things simple w.r.t. NIO filters. The sole problem with blocking filters is that they act upon InputStream/OutputStream obtained from the context. Hence I would say the following is satisfying:
>
> * Adding a new method "public ByteBuffer filter(Context ctx, ByteBuffer src) default { return src; }" to all four Filter interfaces. Thanks to the default method, this keeps backwards compatibility. This method allows filters to optionally support NIO push model. The runtime engine simply has to forward each package it gets into that method and forward the package it gets back from it. Calling context.abortWith() will stop the chain.
>
> * Marker annotation @NIO at a filter class tells the runtime engine it MUST throw UnsupportedOperation for getEntityStream() / setEntityStream() on the filter contexts to effectively prevent such a filter class to PULL from a potential blocking stream. At the same time, the runtime engine knows that it MUST PUSH each block of information through the filter class's filter(ByteBuffer) method instead of the original JAX-RS 2 filter(ctx) method.
>
> This should be fairly easy to implement, as it is at-least intrusive. At deployment time the runtime engine can inspect the filter for existence of @NIO and set up either the blocking or the NIO invocation handler. Existing custom filters can simply get enhanced by overwriting the new filter(ctx, src) variant and adding @NIO. They even would be executable on JAX-RS 2.0 engines, as their interface is backwards compatible.
>
> BTW, as a vendor of several custom complex filters I strongly support the need to have ALL functionality of blocking filters in NIO filters.
>
> -Markus
>
> -----Original Message-----
> From: Santiago Pericas-Geertsen [mailto:santiago.pericasgeertsen_at_oracle.com]
> Sent: Mittwoch, 15. März 2017 21:29
> To: jsr370-experts_at_jax-rs-spec.java.net
> Subject: Re: [jax-rs-spec users] NIO API review / request for feedback
>
>
>> On Mar 15, 2017, at 11:37 AM, Sergey Beryozkin <sberyozkin_at_talend.com> wrote:
>>
>> I guess if we have new NIO filters then we'd need to decide how the existing filters, example, Logging ones, would work alongside, asking users to write the filters which do the generic processing of input/output streams should still work even if the service code is optimized to do NIO ?
> This is where the mental struggle is, I think. Looking at request entity processing in particular, the NIO model is essentially a *push* model (there’s a publisher for the entity) whereas the existing streams are essentially a *pull* model (we read from a stream).
>
> The Servlet spec worked around this by extending the streams to support a listeners (Read/WriteListener). If we want to support Flows, as we have been trying, can we really mix things? Is there such a thing as a hybrid pipeline?
>
> — Santiago
>
>> Thanks, Sergey
>> On 15/03/17 15:32, Sergey Beryozkin wrote:
>>> Hi Santiago
>>> On 15/03/17 13:06, Santiago Pericas-Geertsen wrote:
>>>> Hi Sergey,
>>>>
>>>>>>> Please see
>>>>>>> https://github.com/pavelbucek/jax-rs/blob/nio/examples/src/main/j
>>>>>>> ava/jaxrs/examples/nio/ServerSideProcessing.java
>>>>>>> There are several methods added on Filters and interceptors, which add the similar functionality as are added for resource methods and Body Readers/Writers.
>>>>>> The pattern used in filters and interceptors feels a bit strange,
>>>>>> and perhaps not inline with the rest of the API. In particular,
>>>>>>
>>>>>> static class RequestInterceptor implements ReaderInterceptor {
>>>>>> @Override
>>>>>> public Object aroundReadFrom(ReaderInterceptorContext
>>>>>> context) throws IOException, WebApplicationException {
>>>>>>
>>>>>> // we might need an Executor/ExecutorService when
>>>>>> creating a processor
>>>>>>
>>>>>> context.addProcessor(PROCESSOR);
>>>>>> return context.proceed();
>>>>>> }
>>>>>> }
>>>>>>
>>>>>> is not really filtering, but instead it is constructing a “flow” pipeline by adding one processor. I think this is because, even though we defined NIO readers/writers, we are trying to re-use the existing non-NIO filters and interceptors.
>>>>> Can that (custom) PROCESSOR be added implicitly so that the user do
>>>>> not have to write the boilerplate filters
>>>> Possibly, but I’m not sure how would that fit with the rest of the API.
>>>>
>>>>>> If I need a NIO filter, why not define one directly (a type extending processor) and associate it with the NIO resource method as we do with normal filters? I see challenges doing this too, but it seems more natural to me.
>>>>> Can you clarify please ? Having Nio filters in addition to all of the existing filters/interceptors is indeed can become 'noisy’.
>>>> NIO processing is very different from normal request processing. It just a feels a bit odd to call one of the existing filters or interceptors only for the purpose of “setting up” a flow pipeline (i.e., adding a processor). I think a filter/interceptor will be for NIO or it won’t.
>>> I have to admit that now that I think about it I don't really understand why NIO filters are needed ? We have new NIOReader/Writer which are good.
>>> What would dedicated NIO filters do in reality but adding the custom
>>> processors ? Sorry, I thought I'd clarify before continuing with
>>> either supporting or not supporting the introduction of NIP specific
>>> filters :-)
>>>
>>>>> Would it make sense to have an annotation like @Nio(MyProcessor.class) which would tell the runtime it would need to add an interceptor which would exactly what Pavel prototyped above ?
>>>> Would that be on the resource method? We can, but that is not the way we associate filters/interceptors today: we define them globally or use name binding to associate them with resource methods.
>>> I thought a bit more, @NIO annotation introducing is not a good idea, Pavel mentioned that of course the custom Processors would likely need the executors and other customizations, so it won't help much.
>>>
>>> Thanks, Sergey
>>>
>>>> — Santiago
>>>>
>>>>>>> On 10/03/2017 18:16, Markus KARG wrote:
>>>>>>>> Pavel,
>>>>>>>> maybe I missed it, but is it planned to also provide a non-blocking replacement for the InputStream / OutputStream methods of Filters and other streaming components?
>>>>>>>> Thanks
>>>>>>>> -Markus
>>>>>>>> From: Pavel Bucek [mailto:pavel.bucek_at_oracle.com]
>>>>>>>> Sent: Dienstag, 7. März 2017 15:26
>>>>>>>> To: jsr370-experts_at_jax-rs-spec.java.net
>>>>>>>> Subject: NIO API review / request for feedback Dear EG members,
>>>>>>>>
>>>>>>>> please allow me to share the direction of what we are thinking about JAX-RS NIO support.
>>>>>>>>
>>>>>>>> As stated before, SSE wasn't the only place were Flow APIs should be utilized - NIO is another area where it can be utilized quite heavily. And that was mostly the main idea - to see where it does make sense to use that.
>>>>>>>>
>>>>>>>> Similarly to SSE, there is a plan to minimize / extract the Flow to different classes, but there should always be a clear path how to convert an instance into Flow.* (or org.reactivestreams.*) interfaces. It is not done yet, to keep things as clear as possible.
>>>>>>>>
>>>>>>>> One of the bigger concerns is backwards compatibility. There are interfaces, which do work directly with Input/Output stream, which is always a problem for reactive processing, since it is blocking by design. The specification will need to say something like "The runtime is non-blocking as long as the application code doesn't read from a provided InputStream or doesn't set an OutputStream."
>>>>>>>>
>>>>>>>> The motivation for doing this is to allow the JAX-RS apps to be non-blocking and reactive. JAX-RS 2.0 entity handling is designed around reading / producing Input / Output Stream, which is blocking by design. Non-blocking approach should result in higher throughput and better resource utilization of the server. Also, integration and coexistence with modern reactive frameworks should be possible to do without losing the advantage of having that framework (which was almost completely lost when dealing with blocking inputs/outputs).
>>>>>>>>
>>>>>>>> Let's jump into code snippets.
>>>>>>>>
>>>>>>>> Server - EX1 (byte handling):
>>>>>>>>
>>>>>>>> Snippet below shows how to process request entity body return response entity body, using Publisher<ByteBuffer> - no MessageBodyReader/Writer is involved. This can be used as a low-level integration point for other frameworks, which are also reactive and will do the processing, like serializing/deserializing (mapping) to some java type, filtering, etc. Returning a Publisher<ByteBuffer> is a reactive/nio alternative to javax.ws.rs.core.StreamingOutput, consuming an entity using Publisher<ByteBuffer> is a reactive/nio alternative to consuming entity as an InputStream.
>>>>>>>>
>>>>>>>> @POST
>>>>>>>> @Path("/ex1")
>>>>>>>> public Flow.Publisher<ByteBuffer> ex1(Flow.Publisher<ByteBuffer> entity) {
>>>>>>>> Ex1Processor processor = new Ex1Processor();
>>>>>>>> entity.subscribe(processor);
>>>>>>>> return processor;
>>>>>>>> }
>>>>>>>> // ex1 processor
>>>>>>>> public static class Ex1Processor implements Flow.Processor<ByteBuffer, ByteBuffer> {
>>>>>>>> // ...
>>>>>>>> }
>>>>>>>> And there is already an issue, which is not clearly solved.
>>>>>>>>
>>>>>>>> Returning a Publisher instance from the resource method does put some constraints on the Publisher itself - it needs to cache all events which will be emitted prior subscription of the jax-rs implementation Subscriber instance (which is the only way how to get the data from a Publisher).
>>>>>>>>
>>>>>>>> This can be solved by stating that request entity publisher won't produce any events until the resource method is invoked and the implementation Subscriber subscribed to the returned response entity publisher. Or the resource method can return Consumer<Flow.Subscriber<ByteBuffer>>, which would effectively grant control of the implementation subscription process. Any comments or suggestions welcomed [ref Q1].
>>>>>>>>
>>>>>>>> Server - EX2 (consuming stream of pojos):
>>>>>>>>
>>>>>>>> The next example should be slightly more straightforward. It shows how to process a Publisher of custom type, in this case called "POJO".
>>>>>>>>
>>>>>>>> The only limitation or rule here would be that the subscriber for the request entity must be subscribed before the resource method "ex2" invocation ends.
>>>>>>>>
>>>>>>>> New interface is introduced here - NioBodyReader. It has exactly the same responsibility as good old MessageBodyReader, but without using a blocking OutputStream to write the entity. Note that the "core" type is the Publisher<ByteBuffer>, which is in this case mapped (or converted) into Publisher of POJOs.
>>>>>>>>
>>>>>>>> @POST
>>>>>>>> @Path("/ex2")
>>>>>>>> @Consumes(MediaType.APPLICATION_JSON)
>>>>>>>> public void ex2(Flow.Publisher<POJO> entity,
>>>>>>>> @Suspended AsyncResponse response) {
>>>>>>>> // TODO: introduce a helper or modify AsyncResponse to support this pattern directly?
>>>>>>>> entity.subscribe(
>>>>>>>> // POJO subscriber - consumer
>>>>>>>> new Flow.Subscriber<POJO>() {
>>>>>>>> @Override
>>>>>>>> public void onSubscribe(Flow.Subscription subscription) {
>>>>>>>> // ...
>>>>>>>> }
>>>>>>>> @Override
>>>>>>>> public void onNext(POJO item) {
>>>>>>>> // ...
>>>>>>>> }
>>>>>>>> @Override
>>>>>>>> public void onError(Throwable throwable) {
>>>>>>>> response.resume(throwable);
>>>>>>>> }
>>>>>>>> @Override
>>>>>>>> public void onComplete() {
>>>>>>>> response.resume(Response.ok().build());
>>>>>>>> }
>>>>>>>> }
>>>>>>>> );
>>>>>>>> }
>>>>>>>> @Provider
>>>>>>>> @Consumes(MediaType.APPLICATION_JSON)
>>>>>>>> public static class Ex2NioBodyReader implements NioBodyReader<POJO> {
>>>>>>>> @Override
>>>>>>>> public boolean isReadable(Class<?> type, Type genericType, Annotation[] annotations, MediaType mediaType) {
>>>>>>>> return true;
>>>>>>>> }
>>>>>>>> @Override
>>>>>>>> public Flow.Publisher<POJO>
>>>>>>>> readFrom(Flow.Publisher<ByteBuffer> entity, Class<POJO> type,
>>>>>>>> Type genericType,
>>>>>>>> Annotation[] annotations,
>>>>>>>> MediaType mediaType,
>>>>>>>> MultivaluedMap<String, String> httpHeaders) {
>>>>>>>> Ex2MappingProcessor mappingProcessor = new Ex2MappingProcessor();
>>>>>>>> entity.subscribe(mappingProcessor);
>>>>>>>> return mappingProcessor;
>>>>>>>> }
>>>>>>>> }
>>>>>>>> // mapping Publisher<ByteBuffer> to Publisher<POJO> //
>>>>>>>> ByteBuffers are expected to contain JSON (indicated by @Consumes on the resource method and NioBodyReader).
>>>>>>>> public static class Ex2MappingProcessor implements Flow.Subscriber<ByteBuffer>, Flow.Publisher<POJO> {
>>>>>>>> // ...
>>>>>>>> }
>>>>>>>> Same issue as [Q1] is valid for this as well - same solution will need to be applied for "readFrom" method.
>>>>>>>> Another issue is about what should be passed to "isReadable" method as "type" parameter. I'm not exactly sure whether we can safely obtain generic type of a parameter from the resource method (public void ex2(Flow.Publisher< POJO> entity, ..)). Any comments/suggestions welcomed [ref Q2].
>>>>>>>>
>>>>>>>> Note that using @Suspended shouldn't be enforced here; it should be possible to return a Response directly and still be able to consume the requestentity.
>>>>>>>>
>>>>>>>> Server - EX3 (producing list of POJOs):
>>>>>>>>
>>>>>>>> The last (for now) example shows how we can produce and write POJOs. Resource method doesn't take any parameters and provides a Publisher of POJO objects, which will be converted to JSON in NioBodyWriter. NioBodyReader is a reactive alternative to MessageBodyReader from older version of the specification.
>>>>>>>>
>>>>>>>> @GET
>>>>>>>> @Path("/ex3")
>>>>>>>> @Produces(MediaType.APPLICATION_JSON)
>>>>>>>> public Flow.Publisher<POJO> ex3() {
>>>>>>>> Flow.Publisher<POJO> pojoPublisher = null;
>>>>>>>> // source of the POJO "stream" can be anything - database call, client call to
>>>>>>>> // another service, ...
>>>>>>>> //
>>>>>>>> // DB
>>>>>>>> // .getEmployees(department) // StreamPublisher<EmployeeDbModel> -- reactive stream
>>>>>>>> // .map((Function<EmployeeDbModel, EmployeeToReturn>) employeeDbModel -> {
>>>>>>>> // // ...
>>>>>>>> // });
>>>>>>>> return pojoPublisher;
>>>>>>>> }
>>>>>>>> @Provider
>>>>>>>> @Produces(MediaType.APPLICATION_JSON)
>>>>>>>> public static class Ex3NioBodyWriter implements NioBodyWriter<POJO> {
>>>>>>>> @Override
>>>>>>>> public boolean isWriteable(Class<?> type, Type genericType, Annotation[] annotations, MediaType mediaType) {
>>>>>>>> return true;
>>>>>>>> }
>>>>>>>> @Override
>>>>>>>> public void writeTo(Flow.Publisher<POJO>
>>>>>>>> entityObjectPublisher, Flow.Subscriber<ByteBuffer> subscriber,
>>>>>>>> Class<?> type,
>>>>>>>> Type genericType,
>>>>>>>> Annotation[] annotations,
>>>>>>>> MediaType mediaType,
>>>>>>>> MultivaluedMap<String, Object> httpHeaders) {
>>>>>>>> // map Publisher<POJO> to Publisher<ByteBuffer> and subscribe Flow.Subscriber<ByteBuffer> to it.
>>>>>>>> }
>>>>>>>> }
>>>>>>>> Resource method is minimalistic, [Q1] applies here as well.
>>>>>>>>
>>>>>>>> The example introduces NioBodyWriter and its isWriteable method does have [Q2], similarly to NioBodyReader. #writeTo doesn't have any issues - [Q1] is mitigated there because the implementation passes a supplier to the implementation - there doesn't need to be anything returned. Something similar might be able to do for NioBodyWriter as well.
>>>>>>>>
>>>>>>>> Comment to writing multiple POJO instances:
>>>>>>>> https://github.com/pavelbucek/jax-rs/blob/bfc5b3d6caecab2f6304f9
>>>>>>>> 2ac7b44a7ad6a5fdff/jaxrs-api/src/main/java/javax/ws/rs/ext/NioBo
>>>>>>>> dyWriter.java#L82
>>>>>>>>
>>>>>>>> Important point to mention is that even when producing multiple instances, the intention here is still to return the single HTTP response.
>>>>>>>>
>>>>>>>> ===
>>>>>>>>
>>>>>>>> We have more, but this email is already too long - I will post more after there is some feedback on the presented concepts and issues. Please let us know if this format is OK or if you'd prefer something else - I guess I could do a screencast, hangout or something similar.
>>>>>>>>
>>>>>>>> Source links:
>>>>>>>>
>>>>>>>> - complete server example:
>>>>>>>> https://github.com/pavelbucek/jax-rs/blob/bfc5b3d6caecab2f6304f9
>>>>>>>> 2ac7b44a7ad6a5fdff/examples/src/main/java/jaxrs/examples/nio/Nio
>>>>>>>> Resource.java
>>>>>>>> - client (to be discussed):
>>>>>>>> https://github.com/pavelbucek/jax-rs/blob/bfc5b3d6caecab2f6304f9
>>>>>>>> 2ac7b44a7ad6a5fdff/examples/src/main/java/jaxrs/examples/nio/Nio
>>>>>>>> Client.java
>>>>>>>> - server side processing (including interceptors):
>>>>>>>> https://github.com/pavelbucek/jax-rs/blob/bfc5b3d6caecab2f6304f9
>>>>>>>> 2ac7b44a7ad6a5fdff/examples/src/main/java/jaxrs/examples/nio/Ser
>>>>>>>> verSideProcessing.java
>>>>>>>>
>>>>>>>> (Not including direct link to individual examples, since we will
>>>>>>>> continue working on them...)
>>>>>>>>
>>>>>>>> Looking forward to your feedback!
>>>>>>>>
>>>>>>>> Thanks and regards,
>>>>>>>> Pavel & Santiago
>>>
>