[
https://issues.apache.org/jira/browse/BEAM-4432?focusedWorklogId=115937&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-115937
]
ASF GitHub Bot logged work on BEAM-4432:
----------------------------------------
Author: ASF GitHub Bot
Created on: 26/Jun/18 12:37
Start Date: 26/Jun/18 12:37
Worklog Time Spent: 10m
Work Description: iemejia commented on a change in pull request #5519:
[BEAM-4432] Adding Sources to produce Synthetic output for Batch pipelines
URL: https://github.com/apache/beam/pull/5519#discussion_r198120428
##########
File path:
sdks/java/io/common/src/main/java/org/apache/beam/sdk/io/common/synthetic/SyntheticBoundedInput.java
##########
@@ -0,0 +1,452 @@
+/*
+ * 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.sdk.io.common.synthetic;
+
+import static com.google.common.base.Preconditions.checkArgument;
+import static com.google.common.base.Preconditions.checkNotNull;
+import static org.apache.commons.math3.stat.StatUtils.sum;
+
+import com.fasterxml.jackson.annotation.JsonProperty;
+import com.fasterxml.jackson.databind.annotation.JsonDeserialize;
+import com.google.common.base.MoreObjects;
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.List;
+import java.util.NoSuchElementException;
+import java.util.Random;
+import org.apache.beam.sdk.coders.ByteArrayCoder;
+import org.apache.beam.sdk.coders.Coder;
+import org.apache.beam.sdk.coders.KvCoder;
+import org.apache.beam.sdk.io.OffsetBasedSource;
+import org.apache.beam.sdk.io.Read;
+import org.apache.beam.sdk.options.PipelineOptions;
+import org.apache.beam.sdk.values.KV;
+import org.apache.beam.sdk.values.PCollection;
+import org.apache.commons.math3.distribution.ConstantRealDistribution;
+import org.joda.time.Duration;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+/**
+ * This {@link SyntheticBoundedInput} class provides a parameterizable batch
custom source that is
+ * deterministic.
+ *
+ * <p>The {@link SyntheticBoundedSource} generates a {@link PCollection} of
+ * {@code KV<byte[], byte[]>}. A fraction of the generated records {@code
KV<byte[], byte[]>} are
+ * associated with "hot" keys, which are uniformly distributed over a fixed
number of hot keys.
+ * The remaining generated records are associated with "random" keys.
+ * Each record will be slowed down by a certain sleep time generated based on
the specified sleep
+ * time distribution when the {@link SyntheticSourceReader} reads each record.
+ * The record {@code KV<byte[], byte[]>} is generated deterministically based
on the record's
+ * position in the source, which enables repeatable execution for debugging.
+ * The SyntheticBoundedInput configurable parameters are defined in {@link
+ * SyntheticBoundedInput.SourceOptions}.
+ *
+ * <p>To read a {@link PCollection} of {@code KV<byte[], byte[]>} from {@link
+ * SyntheticBoundedInput}, use {@link SyntheticBoundedInput#readFrom} to
construct the synthetic
+ * source with synthetic source options.
+ * See {@link SyntheticBoundedInput.SourceOptions} for how to construct an
instance.
+ * An example is below:
+ * <pre> {@code
+ * Pipeline p = ...;
+ * SyntheticBoundedInput.SourceOptions sso = ...;
+ *
+ * // Construct the synthetic input with synthetic source options.
+ * PCollection<KV<byte[], byte[]>> input =
p.apply(SyntheticBoundedInput.readFrom(sso));
+ * } </pre>
+ */
+public class SyntheticBoundedInput {
+ /**
+ * Read from the synthetic source options.
+ */
+ public static Read.Bounded<KV<byte[], byte[]>> readFrom(SourceOptions
options) {
+ checkNotNull(options, "Input synthetic source options should not be
null.");
+ options.validate();
+ return Read.from(new SyntheticBoundedSource(options));
+ }
+
+ /**
+ * A {@link SyntheticBoundedSource} that reads {@code KV<byte[], byte[]>}.
+ */
+ public static class SyntheticBoundedSource extends
OffsetBasedSource<KV<byte[], byte[]>> {
+ private static final long serialVersionUID = 0;
+ private static final Logger LOG =
LoggerFactory.getLogger(SyntheticBoundedSource.class);
+
+ private final SourceOptions sourceOptions;
+
+ public SyntheticBoundedSource(SourceOptions sourceOptions) {
+ this(0, sourceOptions.numRecords, sourceOptions);
+ }
+
+ public SyntheticBoundedSource(long startOffset, long endOffset,
SourceOptions sourceOptions) {
+ super(startOffset, endOffset, 1);
+ this.sourceOptions = sourceOptions;
+ LOG.debug("Constructing {}", toString());
+ }
+
+ @Override
+ public Coder<KV<byte[], byte[]>> getDefaultOutputCoder() {
+ return KvCoder.of(ByteArrayCoder.of(), ByteArrayCoder.of());
+ }
+
+ @Override
+ // TODO: test cases where the source size could not be estimated (i.e.,
return 0).
+ // TODO: test cases where the key size and value size might differ from
record to record.
+ // The key size and value size might have their own distributions.
+ public long getBytesPerOffset() {
+ return sourceOptions.bytesPerRecord >= 0
+ ? sourceOptions.bytesPerRecord
+ : sourceOptions.keySizeBytes + sourceOptions.valueSizeBytes;
+ }
+
+ @Override
+ public void validate() {
+ super.validate();
+ sourceOptions.validate();
+ }
+
+ @Override
+ public String toString() {
+ return MoreObjects.toStringHelper(this)
+ .add("options", sourceOptions)
+ .add("indexRange", "[" + getStartOffset() + ", " + getEndOffset() +
")")
+ .toString();
+ }
+
+ @Override
+ public final SyntheticBoundedSource createSourceForSubrange(long start,
long end) {
+ checkArgument(
+ start >= getStartOffset(),
+ "Start offset value " + start
+ + " of the subrange cannot be smaller than the start offset
value " + getStartOffset()
+ + " of the parent source");
+ checkArgument(
+ end <= getEndOffset(),
+ "End offset value " + end + " of the subrange cannot be larger than
the end offset value "
+ + getEndOffset() + " of the parent source");
+
+ return new SyntheticBoundedSource(start, end, sourceOptions);
+ }
+
+ @Override
+ public long getMaxEndOffset(PipelineOptions options) {
+ return getEndOffset();
+ }
+
+ @Override
+ public SyntheticSourceReader createReader(PipelineOptions pipelineOptions)
throws IOException {
+ return new SyntheticSourceReader(this);
+ }
+
+ @Override
+ public List<SyntheticBoundedSource> split(
+ long desiredBundleSizeBytes, PipelineOptions options) throws Exception
{
+ List<SyntheticBoundedSource> res = new ArrayList<>();
+
+ // Choose number of bundles either based on explicit parameter,
+ // or based on size and hints.
+ int desiredNumBundles =
+ (sourceOptions.forceNumInitialBundles == null)
+ ? ((int) Math.ceil(1.0 * getEstimatedSizeBytes(options) /
desiredBundleSizeBytes))
+ : sourceOptions.forceNumInitialBundles;
+
+ // Generate relative bundle sizes using the given distribution.
+ double[] relativeSizes = new double[desiredNumBundles];
+ for (int i = 0; i < relativeSizes.length; ++i) {
+ relativeSizes[i] =
+ sourceOptions.bundleSizeDistribution.sample(
+ sourceOptions.hashFunction.hashInt(i).asLong());
+ }
+ double s = sum(relativeSizes);
+
+ // Generate offset ranges proportional to the relative sizes.
+ long[] sizes = new long[relativeSizes.length];
+ long startOffset = getStartOffset();
+ double sizeSoFar = 0;
+ for (int i = 0; i < relativeSizes.length; ++i) {
+ sizeSoFar += relativeSizes[i];
+ long endOffset =
+ (i == relativeSizes.length - 1)
+ ? getEndOffset()
+ : (long) (getStartOffset() + sizeSoFar * (getEndOffset() -
getStartOffset()) / s);
+ if (startOffset != endOffset) {
+ res.add(createSourceForSubrange(startOffset, endOffset));
+ }
+ sizes[i] = endOffset - startOffset;
+ startOffset = endOffset;
+ }
+ LOG.info("Split into {} bundles of sizes: {}", sizes.length,
Arrays.toString(sizes));
+ return res;
+ }
+ }
+
+ /**
+ * Shape of the progress reporting curve as a function of the current offset
+ * in the {@link SyntheticBoundedSource}.
+ */
+ public enum ProgressShape {
+ /**
+ * Reported progress grows linearly from 0 to 1.
+ */
+ LINEAR,
+ /**
+ * Reported progress decreases linearly from 0.9 to 0.1.
+ */
+ LINEAR_REGRESSING,
+ }
+
+ /**
+ * Synthetic bounded source options.
+ * These options are all JSON, see documentations of individual fields for
details.
+ * {@code SourceOptions} uses jackson annotations
+ * which PipelineOptionsFactory can use to parse and construct an instance.
+ */
+ public static class SourceOptions extends SyntheticOptions {
+ private static final long serialVersionUID = 0;
+
+ /** Total number of generated records. */
+ @JsonProperty public long numRecords;
+
+ /**
+ * Only records whose index is a multiple of this will be split points.
+ * 0 means the source is not dynamically splittable (but is perfectly
statically splittable).
+ * In that case it also doesn't report progress at all.
+ */
+ @JsonProperty public long splitPointFrequencyRecords = 1;
+
+ /**
+ * Distribution for generating initial split bundles.
+ *
+ * <p>When splitting into "desiredBundleSizeBytes", we'll compute the
desired number of bundles
+ * N, then sample this many numbers from this distribution, normalize
their sum to 1, and use
+ * that as the boundaries of generated bundles.
+ *
+ * <p>The Zipf distribution is expected to be particularly useful here.
+ *
+ * <p>E.g., empirically, with 100 bundles, the Zipf distribution with a
parameter of 3.5 will
+ * generate bundles where the largest is about 3x-10x larger than the
median; with a parameter
+ * of 3.0 this ratio will be about 5x-50x; with 2.5, 5x-100x (i.e. 1
bundle can be as large as
+ * all others combined).
+ *
+ * <p>Python code that can be run to experiment with the implications of
various parameters:
+ * <pre>
+ * from numpy.random import zipf
+ * from numpy import median
+ *
+ * def spread(alpha, n):
+ * x = zipf(alpha, n)
+ * return max(x)/median(x)
+ *
+ * def margins(x): return (int(percentile(x, 5)), int(percentile(x, 95)))
+ *
+ * # For every combination of alpha and n, perform 100 experiments and
measure
+ * # how big is the spread. Return the 5% and 95% quantiles of that
experiment.
+ * [(alpha, n, margins([zipf_spread(alpha, n) for i in range(1,100)]))
+ * for alpha in [2.5, 3, 3.5, 4]
+ * for n in [100, 1000, 10000]]
+ * </pre>
+ */
+ @JsonDeserialize(using = SamplerDeserializer.class)
+ public Sampler bundleSizeDistribution = fromRealDistribution(new
ConstantRealDistribution(1));
+
+ /**
+ * If specified, this source will split into exactly this many bundles
regardless of
+ * the hints provided by the service.
+ */
+ @JsonProperty public Integer forceNumInitialBundles;
+
+ /**
+ * See {@link ProgressShape}.
+ */
+ @JsonProperty public ProgressShape progressShape = ProgressShape.LINEAR;
+
+ /**
+ * The distribution for the delay when reading from synthetic source
starts. This delay is
+ * independent of the per-record delay and uses the same types of
distributions as
+ * {@link #delayDistribution}.
+ */
+ @JsonDeserialize(using = SamplerDeserializer.class)
+ public Sampler initializeDelayDistribution = fromRealDistribution(
+ new ConstantRealDistribution(0));
+
+ /**
+ * Generates a random delay value for the synthetic source initialization
using the distribution
+ * defined by {@link #initializeDelayDistribution}.
+ */
+ public Duration nextInitializeDelay(long seed) {
+ return Duration.millis((long) initializeDelayDistribution.sample(seed));
+ }
+
+ @Override
+ public void validate() {
+ super.validate();
+ checkArgument(
+ numRecords >= 0, "numRecords should be a non-negative number, but
found %s.", numRecords);
+ checkNotNull(bundleSizeDistribution, "bundleSizeDistribution");
+ checkArgument(
+ forceNumInitialBundles == null || forceNumInitialBundles > 0,
+ "forceNumInitialBundles, if specified, must be positive, but found
%s",
+ forceNumInitialBundles);
+ checkArgument(
+ splitPointFrequencyRecords >= 0,
+ "splitPointFrequencyRecords must be non-negative, but found %s",
+ splitPointFrequencyRecords);
+ }
+
+ public Record genRecord(long position) {
+ // This method is supposed to generate random records deterministically,
+ // so that results can be reproduced by running the same scenario a
second time.
+ // We need to initiate a Random object for each position to make the
record deterministic
+ // because liquid sharding could split the Source at any position.
+ // And we also need a seed to initiate a Random object. The mapping from
the position to
+ // the seed should be fixed. Using the position as seed to feed Random
objects will cause the
+ // generated values to not be random enough because the position values
are
+ // close to each other. To make seeds fed into the Random objects
unrelated,
+ // we use a hashing function to map the position to its corresponding
hashcode,
+ // and use the hashcode as a seed to feed into the Random object.
+ long hashCodeOfPosition = hashFunction.hashLong(position).asLong();
+ return new Record(genKvPair(hashCodeOfPosition),
nextDelay(hashCodeOfPosition));
+ }
+
+ /**
+ * Record generated by {@link #genRecord}.
+ */
+ public static class Record {
+ public final KV<byte[], byte[]> kv;
+ public final Duration sleepMsec;
+
+ public Record(KV<byte[], byte[]> kv, long sleepMsec) {
+ this.kv = kv;
+ this.sleepMsec = new Duration(sleepMsec);
+ }
+ }
+ }
+
+ /**
+ * A reader over the {@link PCollection} of {@code KV<byte[], byte[]>}
+ * from the synthetic source.
+ *
+ * <p>The random but deterministic record at position "i" in the range [A, B)
+ * is generated by using {@link SourceOptions#genRecord}.
+ * Reading each record sleeps according to the sleep time distribution
+ * in {@code SyntheticOptions}.
+ */
+ private static class SyntheticSourceReader
+ extends OffsetBasedSource.OffsetBasedReader<KV<byte[], byte[]>> {
+ private final long splitPointFrequencyRecords;
+
+ private KV<byte[], byte[]> currentKvPair;
+ private long index;
+ private boolean isAtSplitPoint;
+
+ public SyntheticSourceReader(SyntheticBoundedSource source) {
+ super(source);
+ this.currentKvPair = null;
+ this.splitPointFrequencyRecords =
source.sourceOptions.splitPointFrequencyRecords;
+ }
+
+ @Override
+ public synchronized SyntheticBoundedSource getCurrentSource() {
+ return (SyntheticBoundedSource) super.getCurrentSource();
+ }
+
+ @Override
+ protected long getCurrentOffset() throws IllegalStateException {
+ return index;
+ }
+
+ @Override
+ public KV<byte[], byte[]> getCurrent() throws NoSuchElementException {
+ if (currentKvPair == null) {
+ throw new NoSuchElementException(
+ "The current element is unavailable because either the reader is "
+ + "at the beginning of the input and start() or advance()
wasn't called, "
+ + "or the last start() or advance() returned false.");
+ }
+ return currentKvPair;
+ }
+
+ @Override
+ public boolean allowsDynamicSplitting() {
+ return splitPointFrequencyRecords > 0;
+ }
+
+ @Override
+ protected final boolean startImpl() throws IOException {
+ this.index = getCurrentSource().getStartOffset();
+ if (splitPointFrequencyRecords > 0) {
+ while (index % splitPointFrequencyRecords != 0) {
+ ++index;
+ }
+ }
+
+ SourceOptions options = getCurrentSource().sourceOptions;
+ SyntheticUtils.delay(options.nextInitializeDelay(this.index),
Review comment:
Is the value of the return of this function needed for something? At first
glance I thought that delays were for event time but I have the impression I
misintepreted this. So maybe worth documenting on `Record`, Nexmark uses
something like `TimestampedValue<Record>` to model event time. Wondering if it
makes sense to add also the implementation of `getCurrentTimestamp()`. Which
provides value even for BoundedSources.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
Issue Time Tracking
-------------------
Worklog Id: (was: 115937)
Time Spent: 2h 40m (was: 2.5h)
> Performance tests need a way to generate Synthetic data
> -------------------------------------------------------
>
> Key: BEAM-4432
> URL: https://issues.apache.org/jira/browse/BEAM-4432
> Project: Beam
> Issue Type: Improvement
> Components: testing
> Reporter: Pablo Estrada
> Assignee: Pablo Estrada
> Priority: Minor
> Time Spent: 2h 40m
> Remaining Estimate: 0h
>
> GenerateSequence fal.lls short in this regard, as we may want to generate
> data in custom distributions, or with specific repeatability requirements /
> and hardcoded delays for autoscaling.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
