Java 8 Stream with batch processing

Question

I have a large file that contains a list of items.

I would like to create a batch of items, make an HTTP request with this batch (all of the items are needed as par

Answer 1

We had a similar problem to solve. We wanted to take a stream that was larger than system memory (iterating through all objects in a database) and randomise the order as best as possible - we thought it would be ok to buffer 10,000 items and randomise them.

The target was a function which took in a stream.

Of the solutions proposed here, there seem to be a range of options:

• Use various non-java 8 additional libraries
• Start with something that's not a stream - e.g. a random access list
• Have a stream which can be split easily in a spliterator

Our instinct was originally to use a custom collector, but this meant dropping out of streaming. The custom collector solution above is very good and we nearly used it.

Here's a solution which cheats by using the fact that Streams can give you an Iterator which you can use as an escape hatch to let you do something extra that streams don't support. The Iterator is converted back to a stream using another bit of Java 8 StreamSupport sorcery.

/**
 * An iterator which returns batches of items taken from another iterator
 */
public class BatchingIterator<T> implements Iterator<List<T>> {
    /**
     * Given a stream, convert it to a stream of batches no greater than the
     * batchSize.
     * @param originalStream to convert
     * @param batchSize maximum size of a batch
     * @param <T> type of items in the stream
     * @return a stream of batches taken sequentially from the original stream
     */
    public static <T> Stream<List<T>> batchedStreamOf(Stream<T> originalStream, int batchSize) {
        return asStream(new BatchingIterator<>(originalStream.iterator(), batchSize));
    }

    private static <T> Stream<T> asStream(Iterator<T> iterator) {
        return StreamSupport.stream(
            Spliterators.spliteratorUnknownSize(iterator,ORDERED),
            false);
    }

    private int batchSize;
    private List<T> currentBatch;
    private Iterator<T> sourceIterator;

    public BatchingIterator(Iterator<T> sourceIterator, int batchSize) {
        this.batchSize = batchSize;
        this.sourceIterator = sourceIterator;
    }

    @Override
    public boolean hasNext() {
        prepareNextBatch();
        return currentBatch!=null && !currentBatch.isEmpty();
    }

    @Override
    public List<T> next() {
        return currentBatch;
    }

    private void prepareNextBatch() {
        currentBatch = new ArrayList<>(batchSize);
        while (sourceIterator.hasNext() && currentBatch.size() < batchSize) {
            currentBatch.add(sourceIterator.next());
        }
    }
}

A simple example of using this would look like this:

@Test
public void getsBatches() {
    BatchingIterator.batchedStreamOf(Stream.of("A","B","C","D","E","F"), 3)
        .forEach(System.out::println);
}

The above prints

[A, B, C]
[D, E, F]

For our use case, we wanted to shuffle the batches and then keep them as a stream - it looked like this:

@Test
public void howScramblingCouldBeDone() {
    BatchingIterator.batchedStreamOf(Stream.of("A","B","C","D","E","F"), 3)
        // the lambda in the map expression sucks a bit because Collections.shuffle acts on the list, rather than returning a shuffled one
        .map(list -> {
            Collections.shuffle(list); return list; })
        .flatMap(List::stream)
        .forEach(System.out::println);
}

This outputs something like (it's randomised, so different every time)

A
C
B
E
D
F

The secret sauce here is that there's always a stream, so you can either operate on a stream of batches, or do something to each batch and then flatMap it back to a stream. Even better, all of the above only runs as the final forEach or collect or other terminating expressions PULL the data through the stream.

It turns out that iterator is a special type of terminating operation on a stream and does not cause the whole stream to run and come into memory! Thanks to the Java 8 guys for a brilliant design!

Answer 2

It could be easily done using Reactor:

Flux.fromStream(fileReader.lines().onClose(() -> safeClose(fileReader)))
            .map(line -> someProcessingOfSingleLine(line))
            .buffer(BUFFER_SIZE)
            .subscribe(apiService::makeHttpRequest);

Answer 3

Pure Java 8 solution:

We can create a custom collector to do this elegantly, which takes in a batch size and a Consumer to process each batch:

import java.util.ArrayList;
import java.util.Collections;
import java.util.List;
import java.util.Set;
import java.util.function.*;
import java.util.stream.Collector;

import static java.util.Objects.requireNonNull;


/**
 * Collects elements in the stream and calls the supplied batch processor
 * after the configured batch size is reached.
 *
 * In case of a parallel stream, the batch processor may be called with
 * elements less than the batch size.
 *
 * The elements are not kept in memory, and the final result will be an
 * empty list.
 *
 * @param <T> Type of the elements being collected
 */
class BatchCollector<T> implements Collector<T, List<T>, List<T>> {

    private final int batchSize;
    private final Consumer<List<T>> batchProcessor;


    /**
     * Constructs the batch collector
     *
     * @param batchSize the batch size after which the batchProcessor should be called
     * @param batchProcessor the batch processor which accepts batches of records to process
     */
    BatchCollector(int batchSize, Consumer<List<T>> batchProcessor) {
        batchProcessor = requireNonNull(batchProcessor);

        this.batchSize = batchSize;
        this.batchProcessor = batchProcessor;
    }

    public Supplier<List<T>> supplier() {
        return ArrayList::new;
    }

    public BiConsumer<List<T>, T> accumulator() {
        return (ts, t) -> {
            ts.add(t);
            if (ts.size() >= batchSize) {
                batchProcessor.accept(ts);
                ts.clear();
            }
        };
    }

    public BinaryOperator<List<T>> combiner() {
        return (ts, ots) -> {
            // process each parallel list without checking for batch size
            // avoids adding all elements of one to another
            // can be modified if a strict batching mode is required
            batchProcessor.accept(ts);
            batchProcessor.accept(ots);
            return Collections.emptyList();
        };
    }

    public Function<List<T>, List<T>> finisher() {
        return ts -> {
            batchProcessor.accept(ts);
            return Collections.emptyList();
        };
    }

    public Set<Characteristics> characteristics() {
        return Collections.emptySet();
    }
}

Optionally then create a helper utility class:

import java.util.List;
import java.util.function.Consumer;
import java.util.stream.Collector;

public class StreamUtils {

    /**
     * Creates a new batch collector
     * @param batchSize the batch size after which the batchProcessor should be called
     * @param batchProcessor the batch processor which accepts batches of records to process
     * @param <T> the type of elements being processed
     * @return a batch collector instance
     */
    public static <T> Collector<T, List<T>, List<T>> batchCollector(int batchSize, Consumer<List<T>> batchProcessor) {
        return new BatchCollector<T>(batchSize, batchProcessor);
    }
}

Example usage:

List<Integer> input = Arrays.asList(1, 2, 3, 4, 5, 6, 7, 8, 9, 10);
List<Integer> output = new ArrayList<>();

int batchSize = 3;
Consumer<List<Integer>> batchProcessor = xs -> output.addAll(xs);

input.stream()
     .collect(StreamUtils.batchCollector(batchSize, batchProcessor));

I've posted my code on GitHub as well, if anyone wants to take a look:

Link to Github

Answer 4

I wrote a custom Spliterator for scenarios like this. It will fill lists of a given size from the input Stream. The advantage of this approach is that it will perform lazy processing, and it will work with other stream functions.

public static <T> Stream<List<T>> batches(Stream<T> stream, int batchSize) {
    return batchSize <= 0
        ? Stream.of(stream.collect(Collectors.toList()))
        : StreamSupport.stream(new BatchSpliterator<>(stream.spliterator(), batchSize), stream.isParallel());
}

private static class BatchSpliterator<E> implements Spliterator<List<E>> {

    private final Spliterator<E> base;
    private final int batchSize;

    public BatchSpliterator(Spliterator<E> base, int batchSize) {
        this.base = base;
        this.batchSize = batchSize;
    }

    @Override
    public boolean tryAdvance(Consumer<? super List<E>> action) {
        final List<E> batch = new ArrayList<>(batchSize);
        for (int i=0; i < batchSize && base.tryAdvance(batch::add); i++)
            ;
        if (batch.isEmpty())
            return false;
        action.accept(batch);
        return true;
    }

    @Override
    public Spliterator<List<E>> trySplit() {
        if (base.estimateSize() <= batchSize)
            return null;
        final Spliterator<E> splitBase = this.base.trySplit();
        return splitBase == null ? null
                : new BatchSpliterator<>(splitBase, batchSize);
    }

    @Override
    public long estimateSize() {
        final double baseSize = base.estimateSize();
        return baseSize == 0 ? 0
                : (long) Math.ceil(baseSize / (double) batchSize);
    }

    @Override
    public int characteristics() {
        return base.characteristics();
    }

}

Answer 5

You can use apache.commons :

ListUtils.partition(ListOfLines, 500).stream()
                .map(partition -> processBatch(partition)
                .collect(Collectors.toList());

The partitioning part is done un-lazily but after the list is partitioned you get the benefits of working with streams (e.g. use parallel streams, add filters, etc.). Other answers suggested more elaborate solutions but sometimes readability and maintainability are more important (and sometimes they are not :-) )

Answer 6

Note! This solution reads the whole file before running the forEach.

You could do it with jOOλ, a library that extends Java 8 streams for single-threaded, sequential stream use-cases:

Seq.seq(lazyFileStream)              // Seq<String>
   .zipWithIndex()                   // Seq<Tuple2<String, Long>>
   .groupBy(tuple -> tuple.v2 / 500) // Map<Long, List<String>>
   .forEach((index, batch) -> {
       process(batch);
   });

Behind the scenes, zipWithIndex() is just:

static <T> Seq<Tuple2<T, Long>> zipWithIndex(Stream<T> stream) {
    final Iterator<T> it = stream.iterator();

    class ZipWithIndex implements Iterator<Tuple2<T, Long>> {
        long index;

        @Override
        public boolean hasNext() {
            return it.hasNext();
        }

        @Override
        public Tuple2<T, Long> next() {
            return tuple(it.next(), index++);
        }
    }

    return seq(new ZipWithIndex());
}

... whereas groupBy() is API convenience for:

default <K> Map<K, List<T>> groupBy(Function<? super T, ? extends K> classifier) {
    return collect(Collectors.groupingBy(classifier));
}

(Disclaimer: I work for the company behind jOOλ)