blob: d0f46ea807c2936e547a15bd8139605ab25f0840 [file] [log] [blame]
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.beam.runners.spark.structuredstreaming.translation.batch;
import java.util.ArrayList;
import java.util.Collection;
import java.util.Collections;
import java.util.HashMap;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Set;
import java.util.stream.Collectors;
import org.apache.beam.runners.spark.structuredstreaming.translation.helpers.EncoderHelpers;
import org.apache.beam.sdk.coders.Coder;
import org.apache.beam.sdk.coders.CoderException;
import org.apache.beam.sdk.coders.IterableCoder;
import org.apache.beam.sdk.transforms.Combine;
import org.apache.beam.sdk.transforms.windowing.BoundedWindow;
import org.apache.beam.sdk.transforms.windowing.PaneInfo;
import org.apache.beam.sdk.transforms.windowing.TimestampCombiner;
import org.apache.beam.sdk.transforms.windowing.WindowFn;
import org.apache.beam.sdk.util.CoderUtils;
import org.apache.beam.sdk.util.WindowedValue;
import org.apache.beam.sdk.values.KV;
import org.apache.beam.sdk.values.WindowingStrategy;
import org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.Iterables;
import org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.Lists;
import org.apache.spark.sql.Encoder;
import org.apache.spark.sql.expressions.Aggregator;
import org.joda.time.Instant;
import scala.Tuple2;
/** An {@link Aggregator} for the Spark Batch Runner.
* The accumulator is a {@code Iterable<WindowedValue<AccumT>> because an {@code InputT} can be in multiple windows. So, when accumulating {@code InputT} values, we create one accumulator per input window.
* */
class AggregatorCombiner<K, InputT, AccumT, OutputT, W extends BoundedWindow>
extends Aggregator<
WindowedValue<KV<K, InputT>>,
Iterable<WindowedValue<AccumT>>,
Iterable<WindowedValue<OutputT>>> {
private final Combine.CombineFn<InputT, AccumT, OutputT> combineFn;
private WindowingStrategy<InputT, W> windowingStrategy;
private TimestampCombiner timestampCombiner;
private Coder<AccumT> accumulatorCoder;
private IterableCoder<WindowedValue<AccumT>> bufferEncoder;
private IterableCoder<WindowedValue<OutputT>> outputCoder;
public AggregatorCombiner(
Combine.CombineFn<InputT, AccumT, OutputT> combineFn,
WindowingStrategy<?, ?> windowingStrategy,
Coder<AccumT> accumulatorCoder,
Coder<OutputT> outputCoder) {
this.combineFn = combineFn;
this.windowingStrategy = (WindowingStrategy<InputT, W>) windowingStrategy;
this.timestampCombiner = windowingStrategy.getTimestampCombiner();
this.accumulatorCoder = accumulatorCoder;
this.bufferEncoder =
IterableCoder.of(
WindowedValue.FullWindowedValueCoder.of(
accumulatorCoder, windowingStrategy.getWindowFn().windowCoder()));
this.outputCoder =
IterableCoder.of(
WindowedValue.FullWindowedValueCoder.of(
outputCoder, windowingStrategy.getWindowFn().windowCoder()));
}
@Override
public Iterable<WindowedValue<AccumT>> zero() {
return new ArrayList<>();
}
private Iterable<WindowedValue<AccumT>> createAccumulator(WindowedValue<KV<K, InputT>> inputWv) {
// need to create an accumulator because combineFn can modify its input accumulator.
AccumT accumulator = combineFn.createAccumulator();
AccumT accumT = combineFn.addInput(accumulator, inputWv.getValue().getValue());
return Lists.newArrayList(
WindowedValue.of(accumT, inputWv.getTimestamp(), inputWv.getWindows(), inputWv.getPane()));
}
@Override
public Iterable<WindowedValue<AccumT>> reduce(
Iterable<WindowedValue<AccumT>> accumulators, WindowedValue<KV<K, InputT>> inputWv) {
return merge(accumulators, createAccumulator(inputWv));
}
@Override
public Iterable<WindowedValue<AccumT>> merge(
Iterable<WindowedValue<AccumT>> accumulators1,
Iterable<WindowedValue<AccumT>> accumulators2) {
// merge the windows of all the accumulators
Iterable<WindowedValue<AccumT>> accumulators = Iterables.concat(accumulators1, accumulators2);
Set<W> accumulatorsWindows = collectAccumulatorsWindows(accumulators);
Map<W, W> windowToMergeResult;
try {
windowToMergeResult = mergeWindows(windowingStrategy, accumulatorsWindows);
} catch (Exception e) {
throw new RuntimeException("Unable to merge accumulators windows", e);
}
// group accumulators by their merged window
Map<W, List<Tuple2<AccumT, Instant>>> mergedWindowToAccumulators = new HashMap<>();
for (WindowedValue<AccumT> accumulatorWv : accumulators) {
// Encode a version of the accumulator if it is in multiple windows. The combineFn is able to
// mutate the accumulator instance and this could lead to incorrect results if the same
// instance is merged across multiple windows so we decode a new instance as needed. This
// prevents issues during merging of accumulators.
byte[] encodedAccumT = null;
if (accumulatorWv.getWindows().size() > 1) {
try {
encodedAccumT = CoderUtils.encodeToByteArray(accumulatorCoder, accumulatorWv.getValue());
} catch (CoderException e) {
throw new RuntimeException(
String.format(
"Unable to encode accumulator %s with coder %s.",
accumulatorWv.getValue(), accumulatorCoder),
e);
}
}
for (BoundedWindow accumulatorWindow : accumulatorWv.getWindows()) {
W mergedWindowForAccumulator = windowToMergeResult.get(accumulatorWindow);
mergedWindowForAccumulator =
(mergedWindowForAccumulator == null)
? (W) accumulatorWindow
: mergedWindowForAccumulator;
// Decode a copy of the accumulator when necessary.
AccumT accumT;
if (encodedAccumT != null) {
try {
accumT = CoderUtils.decodeFromByteArray(accumulatorCoder, encodedAccumT);
} catch (CoderException e) {
throw new RuntimeException(
String.format(
"Unable to encode accumulator %s with coder %s.",
accumulatorWv.getValue(), accumulatorCoder),
e);
}
} else {
accumT = accumulatorWv.getValue();
}
// we need only the timestamp and the AccumT, we create a tuple
Tuple2<AccumT, Instant> accumAndInstant =
new Tuple2<>(
accumT,
timestampCombiner.assign(mergedWindowForAccumulator, accumulatorWv.getTimestamp()));
if (mergedWindowToAccumulators.get(mergedWindowForAccumulator) == null) {
mergedWindowToAccumulators.put(
mergedWindowForAccumulator, Lists.newArrayList(accumAndInstant));
} else {
mergedWindowToAccumulators.get(mergedWindowForAccumulator).add(accumAndInstant);
}
}
}
// merge the accumulators for each mergedWindow
List<WindowedValue<AccumT>> result = new ArrayList<>();
for (Map.Entry<W, List<Tuple2<AccumT, Instant>>> entry :
mergedWindowToAccumulators.entrySet()) {
W mergedWindow = entry.getKey();
List<Tuple2<AccumT, Instant>> accumsAndInstantsForMergedWindow = entry.getValue();
// we need to create the first accumulator because combineFn.mergerAccumulators can modify the
// first accumulator
AccumT first = combineFn.createAccumulator();
Iterable<AccumT> accumulatorsToMerge =
Iterables.concat(
Collections.singleton(first),
accumsAndInstantsForMergedWindow.stream()
.map(x -> x._1())
.collect(Collectors.toList()));
result.add(
WindowedValue.of(
combineFn.mergeAccumulators(accumulatorsToMerge),
timestampCombiner.combine(
accumsAndInstantsForMergedWindow.stream()
.map(x -> x._2())
.collect(Collectors.toList())),
mergedWindow,
PaneInfo.NO_FIRING));
}
return result;
}
@Override
public Iterable<WindowedValue<OutputT>> finish(Iterable<WindowedValue<AccumT>> reduction) {
List<WindowedValue<OutputT>> result = new ArrayList<>();
for (WindowedValue<AccumT> windowedValue : reduction) {
result.add(windowedValue.withValue(combineFn.extractOutput(windowedValue.getValue())));
}
return result;
}
@Override
public Encoder<Iterable<WindowedValue<AccumT>>> bufferEncoder() {
return EncoderHelpers.fromBeamCoder(bufferEncoder);
}
@Override
public Encoder<Iterable<WindowedValue<OutputT>>> outputEncoder() {
return EncoderHelpers.fromBeamCoder(outputCoder);
}
private Set<W> collectAccumulatorsWindows(Iterable<WindowedValue<AccumT>> accumulators) {
Set<W> windows = new HashSet<>();
for (WindowedValue<?> accumulator : accumulators) {
for (BoundedWindow untypedWindow : accumulator.getWindows()) {
@SuppressWarnings("unchecked")
W window = (W) untypedWindow;
windows.add(window);
}
}
return windows;
}
private Map<W, W> mergeWindows(WindowingStrategy<InputT, W> windowingStrategy, Set<W> windows)
throws Exception {
WindowFn<InputT, W> windowFn = windowingStrategy.getWindowFn();
if (!windowingStrategy.needsMerge()) {
// Return an empty map, indicating that every window is not merged.
return Collections.emptyMap();
}
Map<W, W> windowToMergeResult = new HashMap<>();
windowFn.mergeWindows(new MergeContextImpl(windowFn, windows, windowToMergeResult));
return windowToMergeResult;
}
private class MergeContextImpl extends WindowFn<InputT, W>.MergeContext {
private Set<W> windows;
private Map<W, W> windowToMergeResult;
MergeContextImpl(WindowFn<InputT, W> windowFn, Set<W> windows, Map<W, W> windowToMergeResult) {
windowFn.super();
this.windows = windows;
this.windowToMergeResult = windowToMergeResult;
}
@Override
public Collection<W> windows() {
return windows;
}
@Override
public void merge(Collection<W> toBeMerged, W mergeResult) throws Exception {
for (W w : toBeMerged) {
windowToMergeResult.put(w, mergeResult);
}
}
}
}