How to correctly read Flux and convert it to a single inputStream

Question

I'm using WebClient and custom BodyExtractorclass for my spring-boot application

WebClient webLCient = WebClient.create();
webClient.get()
   .uri(url, params)
   .accept(MediaType.APPLICATION.XML)
   .exchange()
   .flatMap(response -> {
     return response.body(new BodyExtractor());
   })

BodyExtractor.java

@Override
public Mono<T> extract(ClientHttpResponse response, BodyExtractor.Context context) {
  Flux<DataBuffer> body = response.getBody();
  body.map(dataBuffer -> {
    try {
      JaxBContext jc = JaxBContext.newInstance(SomeClass.class);
      Unmarshaller unmarshaller = jc.createUnmarshaller();

      return (T) unmarshaller.unmarshal(dataBuffer.asInputStream())
    } catch(Exception e){
       return null;
    }
  }).next();
}

Above code works with small payload but not on a large payload, I think it's because I'm only reading a single flux value with next and I'm not sure how to combine and read all dataBuffer.

I'm new to reactor, so I don't know a lot of tricks with flux/mono.

Answer 1

This is really not as complicated as other answers imply.

The only way to stream the data without buffering it all in memory is to use a pipe, as @jin-kwon suggested. However, it can be done very simply by using Spring's BodyExtractors and DataBufferUtils utility classes.

Example:

private InputStream readAsInputStream(String url) throws IOException {
    PipedOutputStream osPipe = new PipedOutputStream();
    PipedInputStream isPipe = new PipedInputStream(osPipe);

    ClientResponse response = webClient.get().uri(url)
        .accept(MediaType.APPLICATION.XML)
        .exchange()
        .block();
    final int statusCode = response.rawStatusCode();
    // check HTTP status code, can throw exception if needed
    // ....

    Flux<DataBuffer> body = response.body(BodyExtractors.toDataBuffers())
        .doOnError(t -> {
            log.error("Error reading body.", t);
            // close pipe to force InputStream to error,
            // otherwise the returned InputStream will hang forever if an error occurs
            try(isPipe) {
              //no-op
            } catch (IOException ioe) {
                log.error("Error closing streams", ioe);
            }
        })
        .doFinally(s -> {
            try(osPipe) {
              //no-op
            } catch (IOException ioe) {
                log.error("Error closing streams", ioe);
            }
        });

    DataBufferUtils.write(body, osPipe)
        .subscribe(DataBufferUtils.releaseConsumer());

    return isPipe;
}

If you don't care about checking the response code or throwing an exception for a failure status code, you can skip the block() call and intermediate ClientResponse variable by using

flatMap(r -> r.body(BodyExtractors.toDataBuffers()))

instead.

Answer 2

A slightly modified version of Bk Santiago's answer makes use of reduce() instead of collect(). Very similar, but doesn't require an extra class:

Java:

body.reduce(new InputStream() {
    public int read() { return -1; }
  }, (s: InputStream, d: DataBuffer) -> new SequenceInputStream(s, d.asInputStream())
).flatMap(inputStream -> /* do something with single InputStream */

Or Kotlin:

body.reduce(object : InputStream() {
  override fun read() = -1
}) { s: InputStream, d -> SequenceInputStream(s, d.asInputStream()) }
  .flatMap { inputStream -> /* do something with single InputStream */ }

Benefit of this approach over using collect() is simply you don't need to have a different class to gather things up.

I created a new empty InputStream(), but if that syntax is confusing, you can also replace it with ByteArrayInputStream("".toByteArray()) instead to create an empty ByteArrayInputStream as your initial value instead.

Answer 3

Here comes another variant from other answers. And it's still not memory-friendly.

static Mono<InputStream> asStream(WebClient.ResponseSpec response) {
    return response.bodyToFlux(DataBuffer.class)
        .map(b -> b.asInputStream(true))
        .reduce(SequenceInputStream::new);
}

static void doSome(WebClient.ResponseSpec response) {
    asStream(response)
        .doOnNext(stream -> {
            // do some with stream
            // close the stream!!!
        })
        .block();
}

Answer 4

I was able to make it work by using Flux#collect and SequenceInputStream

@Override
public Mono<T> extract(ClientHttpResponse response, BodyExtractor.Context context) {
  Flux<DataBuffer> body = response.getBody();
  return body.collect(InputStreamCollector::new, (t, dataBuffer)-> t.collectInputStream(dataBuffer.asInputStream))
    .map(inputStream -> {
      try {
        JaxBContext jc = JaxBContext.newInstance(SomeClass.class);
        Unmarshaller unmarshaller = jc.createUnmarshaller();

        return (T) unmarshaller.unmarshal(inputStream);
      } catch(Exception e){
        return null;
      }
  }).next();
}

InputStreamCollector.java

public class InputStreamCollector {
  private InputStream is;

  public void collectInputStream(InputStream is) {
    if (this.is == null) this.is = is;
    this.is = new SequenceInputStream(this.is, is);
  }

  public InputStream getInputStream() {
    return this.is;
  }
}

Answer 5

There's a much cleaner way to do this using the underlying reactor-netty HttpClient directly, instead of using WebClient. The composition hierarchy is like this:

WebClient -uses-> HttpClient -uses-> TcpClient

Easier to show code than explain:

HttpClient.create()
    .get()
    .responseContent() // ByteBufFlux
    .aggregate() // ByteBufMono
    .asInputStream() // Mono<InputStream>
    .block() // We got an InputStream, yay!

However, as I've pointed out already, using InputStream is a blocking operation, that defeats the purpose of using a non-blocking HTTP client, not to mention aggregating the whole response. See this for a Java NIO vs. IO comparison.

Answer 6

You can use pipes.

static <R> Mono<R> pipeAndApply(
        final Publisher<DataBuffer> source, final Executor executor,
        final Function<? super ReadableByteChannel, ? extends R> function) {
    return using(Pipe::open,
                 p -> {
                     executor.execute(() -> write(source, p.sink())
                             .doFinally(s -> {
                                 try {
                                     p.sink().close();
                                 } catch (final IOException ioe) {
                                     log.error("failed to close pipe.sink", ioe);
                                     throw new RuntimeException(ioe);
                                 }
                             })
                             .subscribe(releaseConsumer()));
                     return just(function.apply(p.source()));
                 },
                 p -> {
                     try {
                         p.source().close();
                     } catch (final IOException ioe) {
                         log.error("failed to close pipe.source", ioe);
                         throw new RuntimeException(ioe);
                     }
                 });
}

Or using CompletableFuture,

static <R> Mono<R> pipeAndApply(
        final Publisher<DataBuffer> source,
        final Function<? super ReadableByteChannel, ? extends R> function) {
    return using(Pipe::open,
                 p -> fromFuture(supplyAsync(() -> function.apply(p.source())))
                         .doFirst(() -> write(source, p.sink())
                                 .doFinally(s -> {
                                     try {
                                         p.sink().close();
                                     } catch (final IOException ioe) {
                                         log.error("failed to close pipe.sink", ioe);
                                         throw new RuntimeException(ioe);
                                     }
                                 })
                                 .subscribe(releaseConsumer())),
                 p -> {
                     try {
                         p.source().close();
                     } catch (final IOException ioe) {
                         log.error("failed to close pipe.source", ioe);
                         throw new RuntimeException(ioe);
                     }
                 });
}