什么是Hyperloglog?
- 一个在大数据量下统计基数的算法, 占用内存小, 误差小, 但是会损失一定精度(Kylin中需要高精度可以用bitmap)。
作为数据人, 我们为何要了解它?
- 它与我们的部分实际业务是有关联的, 理解原理能更好的做好工作。
- 应用了Hyperloglog算法的框架:
- Redis
- Apache Kylin
- 应用了Hyperloglog算法的框架:
理解方式
-
有两种理解方式
-
在理想状态下, 将一对数据hash至[0, 1], 每两点间距离d相等, 则这堆数据的基数即为 1/d。
- 但实际情况通常都不能如愿, 只能用分桶取kmax的方式不断逼近该基数值(积分?)。
- 分桶将数据分为m组, 每组取第k个位置的值, 所有组中得到最大的kmax, (k - 1)/kmax 即为得到估计的基数。
-
以抛硬币的方式理解
-
以抛硬币出现一次反面为一次过程, 记录为1, 若抛硬币为正面则记录为0。
-
当实验次数k很大时, 硬币不出现反面的概率基本为0。
-
转换到基数的思想是: 可以用第一个1出现前0的个数n来统计基数。
-
当基数大致为2n+1时, 硬币的概率统计可以为:
[frac{1}{2}*1+frac{1}{4}*2+frac{1}{8}*3 ...... ]
-
-
算法伪代码
- 流程概括:
- hash成32位的值， 并获取最左位置为1所对应的数
- 初始化m个登记表, m∈[24, 216]
- 计算出每组最大的首零位
- 计算基数预估值并根据预估值大小做调整
Hyperloglog的开源Java实现
/*
* Copyright (C) 2012 Clearspring Technologies, Inc.
*
* Licensed 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 com.clearspring.analytics.stream.cardinality;
import java.io.ByteArrayInputStream;
import java.io.ByteArrayOutputStream;
import java.io.DataInput;
import java.io.DataInputStream;
import java.io.DataOutput;
import java.io.DataOutputStream;
import java.io.Externalizable;
import java.io.IOException;
import java.io.ObjectInput;
import java.io.ObjectInputStream;
import java.io.ObjectOutput;
import java.io.Serializable;
import com.clearspring.analytics.hash.MurmurHash;
import com.clearspring.analytics.util.Bits;
import com.clearspring.analytics.util.IBuilder;
/**
* Java implementation of HyperLogLog (HLL) algorithm from this paper:
* <p/>
* http://algo.inria.fr/flajolet/Publications/FlFuGaMe07.pdf
* <p/>
* HLL is an improved version of LogLog that is capable of estimating
* the cardinality of a set with accuracy = 1.04/sqrt(m) where
* m = 2^b. So we can control accuracy vs space usage by increasing
* or decreasing b.
* 准确度: a = 1.04/sqrt(m), m = 2^b, 可以通过增加或减少b参数来控制精度和占用空间
* <p/>
* The main benefit of using HLL over LL is that it only requires 64%
* of the space that LL does to get the same accuracy.
* Hyperloglog算法最大的优势是它只需要常规loglog算法的64%空间就能维持与其相等的精度
* <p/>
* This implementation implements a single counter. If a large (millions)
* number of counters are required you may want to refer to:
* 此实现仅实现了单个计数器, 如果需要千百万数量的计数器, 请参考以下链接:
* <p/>
* http://dsiutils.di.unimi.it/
* <p/>
* It has a more complex implementation of HLL that supports multiple counters
* in a single object, drastically reducing the java overhead from creating
* a large number of objects.
* 它有更复杂的支持单对象中有多个计数器的Hyperloglog实现, 大幅度减少了java创建大量对象的开销
* <p/>
* This implementation leveraged a javascript implementation that Yammer has
* been working on:
* 该实现对Yammer所做的js实现有一定影响
* <p/>
* https://github.com/yammer/probablyjs
* <p>
* Note that this implementation does not include the long range correction function
* defined in the original paper. Empirical evidence shows that the correction
* function causes more harm than good.
* 需要注意的是, 此实现没有包含原先paper中的长跨度修正函数。实验表明修正函数的负面影响大于正面影响。
* </p>
* <p/>
* <p>
* Users have different motivations to use different types of hashing functions.
* 使用者有不同的动机来使用不同的哈希函数,
* Rather than try to keep up with all available hash functions and to remove
* the concern of causing future binary incompatibilities this class allows clients
* to offer the value in hashed int or long form.
* 是设法保留所有哈希函数并移除所有会导致将来的
* 二进制不兼容性比如该类允许客户端提供hashed int 或者 hashed long形式的参数。
* This way clients are free to change their hash function on their own time line.
* 此方式下客户端可以随意在它们的时间线上改变它们的哈希函数。
* We recommend using Google's Guava Murmur3_128 implementation as it provides good
* performance and speed when high precision is required.
* 我们推荐使用Google的Guava Murmur3_128实现, 因为它在高精度要求下提供了优秀的性能和速
* 度。
* In our tests the 32bit MurmurHash function included in this project is faster and
* produces better results than the 32 bit murmur3 implementation google provides.
* 在我们的测试中此项目中的32bit MurmurHash 函数 相比Google提供的 32 bit murmur3实现 更
* 快且产生了更好的结果。
* </p>
*/
public class HyperLogLog implements ICardinality, Serializable {
// 注册集
private final RegisterSet registerSet;
private final int log2m;
private final double alphaMM;
/**
* Create a new HyperLogLog instance using the specified standard deviation.
* 通过使用特定的标准差创建一个新的HyperLogLog实例。
*
* rsd是该计数器的相对标准差, 该值越小, 创建计数器就需要更多的空间(精度与空间的取舍)。
* @param rsd - the relative standard deviation for the counter.
* smaller values create counters that require more space.
*/
public HyperLogLog(double rsd) {
this(log2m(rsd));
}
private static int log2m(double rsd) {
return (int) (Math.log((1.106 / rsd) * (1.106 / rsd)) / Math.log(2));
}
private static double rsd(int log2m) {
return 1.106 / Math.sqrt(Math.exp(log2m * Math.log(2)));
}
private static double logBase(double exponent, double base) {
return Math.log(exponent) / Math.log(base);
}
private static int accuracyToLog2m(double accuracy) {
return Math.toIntExact(2 * Math.round(logBase(1.04 / (1 - accuracy), 2)));
}
private static void validateLog2m(int log2m) {
if (log2m < 0 || log2m > 30) {
throw new IllegalArgumentException("log2m argument is "
+ log2m + " and is outside the range [0, 30]");
}
}
/**
* Create a new HyperLogLog instance. The log2m parameter defines the accuracy
* of the counter.
* 创建一个新的Hyperloglog实例, log2m参数定义了计数器的准确度(log2m越大越准确)
* The larger the log2m the better the accuracy.<p/>
* accuracy = 1 - 1.04/sqrt(2^log2m)
*
* @param log2m - the number of bits to use as the basis for the HLL instance
* log2m: 被用作HyperLogLog实例基类的比特数
*/
public HyperLogLog(int log2m) {
this(log2m, new RegisterSet(1 << log2m));
}
/**
* Creates a new HyperLogLog instance using the given registers.
* 用所给的注册集创建一个新的HyperLogLog实例(已过时)。
* Used for unmarshalling a serialized
* instance and for merging multiple counters together.
* 用于解组一个序列化过的实例以及合并多个计数器
*
* @param registerSet - the initial values for the register set
* 注册集的初始值
*/
@Deprecated
public HyperLogLog(int log2m, RegisterSet registerSet) {
validateLog2m(log2m);
this.registerSet = registerSet;
this.log2m = log2m;
int m = 1 << this.log2m;
alphaMM = getAlphaMM(log2m, m);
}
@Override
public boolean offerHashed(long hashedValue) {
// j becomes the binary address determined by the first b log2m of x
// j成为了由第一个b(即log2m)所决定的地址, >>> 无符号右移, 若hashedValue为正则高位补0, 若为负责, 则右移后补0. 等价于:
/*
if(hashedValue == 0){
j = 0
} else if(hashValue > 0){
j = hashedValue >> (Long.SIZE - log2m) = hashedValue/2^(Long.SIZE - log2m)
} else {
j = -hashedValue >> (Long.SIZE - log2m) = -hashedValue/2^(Long.SIZE - log2m)
}
*/
// j will be between 0 and 2^log2m j会在0~2^log2m之间
// 比较j位置的桶内的数值与传入的值r, 比较当前值和新值, 如果新值大就更新
final int j = (int) (hashedValue >>> (Long.SIZE - log2m));
final int r = Long.numberOfLeadingZeros((hashedValue << this.log2m) | (1 << (this.log2m - 1)) + 1) + 1;
return registerSet.updateIfGreater(j, r);
}
@Override
public boolean offerHashed(int hashedValue) {
// j becomes the binary address determined by the first b log2m of x
// j will be between 0 and 2^log2m
final int j = hashedValue >>> (Integer.SIZE - log2m);
final int r = Integer.numberOfLeadingZeros((hashedValue << this.log2m) | (1 << (this.log2m - 1)) + 1) + 1;
return registerSet.updateIfGreater(j, r);
}
@Override
public boolean offer(Object o) {
final int x = MurmurHash.hash(o);
return offerHashed(x);
}
@Override
public long cardinality() {
double registerSum = 0;
int count = registerSet.count;
double zeros = 0.0;
for (int j = 0; j < registerSet.count; j++) {
int val = registerSet.get(j);
registerSum += 1.0 / (1 << val);
if (val == 0) {
zeros++;
}
}
double estimate = alphaMM * (1 / registerSum);
if (estimate <= (5.0 / 2.0) * count) {
// Small Range Estimate 小范围的预估
return Math.round(linearCounting(count, zeros));
} else {
return Math.round(estimate);
}
}
@Override
public int sizeof() {
return registerSet.size * 4;
}
@Override
public byte[] getBytes() throws IOException {
ByteArrayOutputStream baos = new ByteArrayOutputStream();
DataOutput dos = new DataOutputStream(baos);
writeBytes(dos);
baos.close();
return baos.toByteArray();
}
private void writeBytes(DataOutput serializedByteStream) throws IOException {
serializedByteStream.writeInt(log2m);
serializedByteStream.writeInt(registerSet.size * 4);
for (int x : registerSet.readOnlyBits()) {
serializedByteStream.writeInt(x);
}
}
/**
* Add all the elements of the other set to this set.
* 将所有其他结合的元素放入此集合
* <p/>
* This operation does not imply a loss of precision.
* 此操作不会产生精度的损失
*
* @param other A compatible Hyperloglog instance (same log2m)
* 另一个可兼容的HyperLogLog实例(相同的 log2m)
* @throws CardinalityMergeException if other is not compatible
*/
public void addAll(HyperLogLog other) throws CardinalityMergeException {
if (this.sizeof() != other.sizeof()) {
throw new HyperLogLogMergeException("Cannot merge estimators of different sizes");
}
registerSet.merge(other.registerSet);
}
@Override
public ICardinality merge(ICardinality... estimators) throws CardinalityMergeException {
HyperLogLog merged = new HyperLogLog(log2m, new RegisterSet(this.registerSet.count));
merged.addAll(this);
if (estimators == null) {
return merged;
}
for (ICardinality estimator : estimators) {
if (!(estimator instanceof HyperLogLog)) {
throw new HyperLogLogMergeException("Cannot merge estimators of different class");
}
HyperLogLog hll = (HyperLogLog) estimator;
merged.addAll(hll);
}
return merged;
}
private Object writeReplace() {
return new SerializationHolder(this);
}
/**
* This class exists to support Externalizable semantics for
* HyperLogLog objects without having to expose a public
* constructor, public write/read methods, or pretend final
* fields aren't final.
* 该类的存在时为了支持Hyperloglog对象的外部化语义并不暴露公有构造器, 公有读写方式, 或
* 者预防最终fields不为final
*
* In short, Externalizable allows you to skip some of the more
* verbose meta-data default Serializable gets you, but still
* includes the class name. In that sense, there is some cost
* to this holder object because it has a longer class name. I
* imagine people who care about optimizing for that have their
* own work-around for long class names in general, or just use
* a custom serialization framework. Therefore we make no attempt
* to optimize that here (eg. by raising this from an inner class
* and giving it an unhelpful name).
* 简短的说Externalizable允许你跳过一些冗长的元数据默认序列化, 但仍包含类名。如此, 维持该长名对象就有一定的开销。此处没有做优化的想法。
*/
private static class SerializationHolder implements Externalizable {
HyperLogLog hyperLogLogHolder;
public SerializationHolder(HyperLogLog hyperLogLogHolder) {
this.hyperLogLogHolder = hyperLogLogHolder;
}
/**
* required for Externalizable
* Externalizable 不需要序列化的时候可以用
*/
public SerializationHolder() {
}
@Override
public void writeExternal(ObjectOutput out) throws IOException {
hyperLogLogHolder.writeBytes(out);
}
@Override
public void readExternal(ObjectInput in) throws IOException, ClassNotFoundException {
hyperLogLogHolder = Builder.build(in);
}
private Object readResolve() {
return hyperLogLogHolder;
}
}
public static class Builder implements IBuilder<ICardinality>, Serializable {
private static final long serialVersionUID = -2567898469253021883L;
private final double rsd;
private transient int log2m;
/**
* Uses the given RSD percentage to determine how many bytes the constructed HyperLogLog will use.
* 使用所给的RSD比例来决定所构造的HyperLogLog会占用多少字节(已过时)
* @deprecated Use {@link #withRsd(double)} instead. This builder's constructors did not match the (already
* themselves ambiguous) constructors of the HyperLogLog class, but there is no way to make them match without
* risking behavior changes downstream.
*/
@Deprecated
public Builder(double rsd) {
this.log2m = log2m(rsd);
validateLog2m(log2m);
this.rsd = rsd;
}
/** This constructor is private to prevent behavior change for ambiguous usages. (Legacy support).
* 此构造器为了以防语意不清的使用, 所以是私有的。
*/
private Builder(int log2m) {
this.log2m = log2m;
validateLog2m(log2m);
this.rsd = rsd(log2m);
}
private void readObject(ObjectInputStream in) throws IOException, ClassNotFoundException {
in.defaultReadObject();
this.log2m = log2m(rsd);
}
@Override
public HyperLogLog build() {
return new HyperLogLog(log2m);
}
@Override
public int sizeof() {
int k = 1 << log2m;
return RegisterSet.getBits(k) * 4;
}
public static Builder withLog2m(int log2m) {
return new Builder(log2m);
}
public static Builder withRsd(double rsd) {
return new Builder(rsd);
}
public static Builder withAccuracy(double accuracy) { return new Builder(accuracyToLog2m(accuracy)); }
public static HyperLogLog build(byte[] bytes) throws IOException {
ByteArrayInputStream bais = new ByteArrayInputStream(bytes);
return build(new DataInputStream(bais));
}
public static HyperLogLog build(DataInput serializedByteStream) throws IOException {
int log2m = serializedByteStream.readInt();
int byteArraySize = serializedByteStream.readInt();
return new HyperLogLog(log2m,
new RegisterSet(1 << log2m, Bits.getBits(serializedByteStream, byteArraySize)));
}
}
@SuppressWarnings("serial")
protected static class HyperLogLogMergeException extends CardinalityMergeException {
public HyperLogLogMergeException(String message) {
super(message);
}
}
protected static double getAlphaMM(final int p, final int m) {
// See the paper.
switch (p) {
case 4:
return 0.673 * m * m;
case 5:
return 0.697 * m * m;
case 6:
return 0.709 * m * m;
default:
return (0.7213 / (1 + 1.079 / m)) * m * m;
}
}
protected static double linearCounting(int m, double V) {
return m * Math.log(m / V);
}
}