Kafka笔记整理(三):消费形式验证与性能测试
[TOC]
Kafka消费形式验证
前面的《Kafka笔记整理(一)》中有提到消费者的消费形式,说明如下:
1、每个consumer属于一个consumer group,可以指定组id。group.id
2、消费形式:
组内:组内的消费者消费同一份数据;同时只能有一个consumer消费一个Topic中的1个partition;
一个consumer可以消费多个partitions中的消息。所以,对于一个topic,同一个group中推荐不能有多于
partitions个数的consumer同时消费,否则将意味着某些consumer将无法得到消息。
组间:每个消费组消费相同的数据,互不影响。
3、在一个consumer多个线程的情况下,一个线程相当于一个消费者。
例如:partition为3,一个consumer起了3个线程消费,另一个后来的consumer就无法消费。
下面就来验证Kafka的消费形式,不过需要说明的是,在消费者的程序代码中,可以指定消费者的group.id(我们下面将会在配置文件中指定)。
而在使用kafka的shell命令时,其实也是可以指定配置文件来指定消费者的group.id的,如果不指定,那么kafka将会随机生成一个group.id(kafka-console-consumer.sh中的kafka.tools.ConsoleConsumer类,如果没有指定group.id,其策略是随机生成)。
在后面的程序代码中,会使用同一group.id开启4个消费的线程(因为我们创建的topic有3个partition),然后在终端中通过kafka shell来开启另外一个消费者,进而达到验证kafka消费形式的目的。
另外,在测试中使用的topic如下:
$ kafka-topics.sh --describe hadoop --zookeeper uplooking01:2181,uplooking02:2181,uplooking03:2181
Topic:hadoop PartitionCount:3 ReplicationFactor:3 Configs:
Topic: hadoop Partition: 0 Leader: 103 Replicas: 103,101,102 Isr: 103,101,102
Topic: hadoop Partition: 1 Leader: 101 Replicas: 101,102,103 Isr: 101,102,103
Topic: hadoop Partition: 2 Leader: 102 Replicas: 102,103,101 Isr: 102,103,101
即partition为3,副本因为也为3.
程序代码
KafkaProducerOps.java
package com.uplooking.bigdata.kafka.producer;
import com.uplooking.bigdata.kafka.constants.Constants;
import org.apache.kafka.clients.producer.KafkaProducer;
import org.apache.kafka.clients.producer.Producer;
import org.apache.kafka.clients.producer.ProducerRecord;
import java.io.IOException;
import java.io.InputStream;
import java.util.Properties;
import java.util.Random;
/**
* 通过这个KafkaProducerOps向Kafka topic中生产相关的数据
* <p>
* Producer
*/
public class KafkaProducerOps {
public static void main(String[] args) throws IOException {
/**
* 专门加载配置文件
* 配置文件的格式:
* key=value
*
* 在代码中要尽量减少硬编码
* 不要将代码写死,要可配置化
*/
Properties properties = new Properties();
InputStream in = KafkaProducerOps.class.getClassLoader().getResourceAsStream("producer.properties");
properties.load(in);
/**
* 两个泛型参数
* 第一个泛型参数:指的就是kafka中一条记录key的类型
* 第二个泛型参数:指的就是kafka中一条记录value的类型
*/
String[] girls = new String[]{"姚慧莹", "刘向前", "周 新", "杨柳"};
Producer<String, String> producer = new KafkaProducer<String, String>(properties);
Random random = new Random();
int start = 1;
for (int i = start; i <= start + 20; i++) {
String topic = properties.getProperty(Constants.KAFKA_PRODUCER_TOPIC);
String key = i + "";
String value = "今天的<--" + girls[random.nextInt(girls.length)] + "-->很美很美哦~";
ProducerRecord<String, String> producerRecord =
new ProducerRecord<String, String>(topic, key, value);
producer.send(producerRecord);
}
producer.close();
}
}
KafkaConsumerOps.java
package com.uplooking.bigdata.kafka.consumer;
import org.apache.kafka.clients.consumer.Consumer;
import org.apache.kafka.clients.consumer.ConsumerRecord;
import org.apache.kafka.clients.consumer.ConsumerRecords;
import org.apache.kafka.clients.consumer.KafkaConsumer;
import java.io.IOException;
import java.util.Arrays;
import java.util.Collection;
import java.util.Properties;
import java.util.concurrent.*;
/**
* 从kafka topic中消费数据
*/
public class KafkaConsumerOps {
public static void main(String[] args) throws IOException {
//线程池
ScheduledExecutorService service = Executors.newScheduledThreadPool(4);
System.out.println("外部开始时间:" + System.currentTimeMillis());
for (int i =0; i < 4; i++){
ScheduledFuture<?> schedule = service.schedule(
new ConsumerThread(),
5L,
TimeUnit.SECONDS);
}
}
}
class ConsumerThread implements Runnable {
public void run() {
System.out.println("线程ID:" + Thread.currentThread().getId() + "线程开始时间:" + System.currentTimeMillis());
/**
* 两个泛型参数
* 第一个泛型参数:指的就是kafka中一条记录key的类型
* 第二个泛型参数:指的就是kafka中一条记录value的类型
*/
Properties properties = new Properties();
try {
properties.load(KafkaConsumerOps.class.getClassLoader().getResourceAsStream("consumer.properties"));
} catch (IOException e) {
e.printStackTrace();
}
Consumer<String, String> consumer = new KafkaConsumer<String, String>(properties);
Collection<String> topics = Arrays.asList("hadoop");
//消费者订阅topic
consumer.subscribe(topics);
ConsumerRecords<String, String> consumerRecords = null;
while (true) {
//接下来就要从topic中拉取数据
consumerRecords = consumer.poll(1000);
//遍历每一条记录
for (ConsumerRecord consumerRecord : consumerRecords) {
long offset = consumerRecord.offset();
Object key = consumerRecord.key();
Object value = consumerRecord.value();
int partition = consumerRecord.partition();
System.out.println("CurrentThreadID: " + Thread.currentThread().getId() + "\toffset: " + offset + "\tpartition: " + partition + "\tkey: " + key + "\tvalue: " + value);
}
}
}
}
MyKafkaPartitioner.java
package com.uplooking.bigdata.kafka.partitioner;
import org.apache.kafka.clients.producer.Partitioner;
import org.apache.kafka.common.Cluster;
import java.util.Map;
import java.util.Random;
/**
* 创建自定义的分区,根据数据的key来进行划分
* <p>
* 可以根据key或者value的hashCode
* 还可以根据自己业务上的定义将数据分散在不同的分区中
* 需求:
* 根据用户输入的key的hashCode值和partition个数求模
*/
public class MyKafkaPartitioner implements Partitioner {
public void configure(Map<String, ?> configs) {
}
/**
* 根据给定的数据设置相关的分区
*
* @param topic 主题名称
* @param key key
* @param keyBytes 序列化之后的key
* @param value value
* @param valueBytes 序列化之后的value
* @param cluster 当前集群的元数据信息
*/
public int partition(String topic, Object key, byte[] keyBytes, Object value, byte[] valueBytes, Cluster cluster) {
Integer partitionNums = cluster.partitionCountForTopic(topic);
int targetPartition = -1;
if (key == null || keyBytes == null) {
targetPartition = new Random().nextInt(10000) % partitionNums;
} else {
int hashCode = key.hashCode();
targetPartition = hashCode % partitionNums;
System.out.println("key: " + key + ", value: " + value + ", hashCode: " + hashCode + ", partition: " + targetPartition);
}
return targetPartition;
}
public void close() {
}
}
Constants.java
package com.uplooking.bigdata.kafka.constants;
public interface Constants {
/**
* 生产的key对应的常量
*/
String KAFKA_PRODUCER_TOPIC = "producer.topic";
}
producer.properties
############################# Producer Basics #############################
# list of brokers used for bootstrapping knowledge about the rest of the cluster
# format: host1:port1,host2:port2 ...
bootstrap.servers=uplooking01:9092,uplooking02:9092,uplooking03:9092
# specify the compression codec for all data generated: none, gzip, snappy, lz4
compression.type=none
# name of the partitioner class for partitioning events; default partition spreads data randomly
partitioner.class=com.uplooking.bigdata.kafka.partitioner.MyKafkaPartitioner
# the maximum amount of time the client will wait for the response of a request
#request.timeout.ms=
# how long `KafkaProducer.send` and `KafkaProducer.partitionsFor` will block for
#max.block.ms=
# the producer will wait for up to the given delay to allow other records to be sent so that the sends can be batched together
#linger.ms=
# the maximum size of a request in bytes
#max.request.size=
# the default batch size in bytes when batching multiple records sent to a partition
#batch.size=
# the total bytes of memory the producer can use to buffer records waiting to be sent to the server
#buffer.memory=
#####设置自定义的topic
producer.topic=hadoop
key.serializer=org.apache.kafka.common.serialization.StringSerializer
value.serializer=org.apache.kafka.common.serialization.StringSerializer
consumer.properties
# Zookeeper connection string
# comma separated host:port pairs, each corresponding to a zk
# server. e.g. "127.0.0.1:3000,127.0.0.1:3001,127.0.0.1:3002"
zookeeper.connect= uplooking01:2181,uplooking02:2181,uplooking03:2181
bootstrap.servers=uplooking01:9092,uplooking02:9092,uplooking03:9092
# timeout in ms for connecting to zookeeper
zookeeper.connection.timeout.ms=6000
#consumer group id
group.id=test-consumer-group
#consumer timeout
#consumer.timeout.ms=5000
key.deserializer=org.apache.kafka.common.serialization.StringDeserializer
value.deserializer=org.apache.kafka.common.serialization.StringDeserializer
pom.xml
主要是kafka-clients的依赖:
<dependencies>
<!--kafka-->
<dependency>
<groupId>org.apache.kafka</groupId>
<artifactId>kafka-clients</artifactId>
<version>0.10.0.1</version>
</dependency>
</dependencies>
测试
先在终端启动一个消费者,注意由于没有指定配置文件,所以其group.id是随机生成的:
$ kafka-console-consumer.sh --topic hadoop --zookeeper uplooking01:2181,uplooking02:2181,uplooking03:2181
接下来分别执行消费者的代码和生产者的代码,然后观察各个终端的输出。
生产者程序的终端输出如下:
key: 1, value: 今天的<--刘向前-->很美很美哦~, hashCode: 49, partition: 1
key: 2, value: 今天的<--刘向前-->很美很美哦~, hashCode: 50, partition: 2
key: 3, value: 今天的<--刘向前-->很美很美哦~, hashCode: 51, partition: 0
key: 4, value: 今天的<--杨柳-->很美很美哦~, hashCode: 52, partition: 1
key: 5, value: 今天的<--姚慧莹-->很美很美哦~, hashCode: 53, partition: 2
key: 6, value: 今天的<--姚慧莹-->很美很美哦~, hashCode: 54, partition: 0
key: 7, value: 今天的<--杨柳-->很美很美哦~, hashCode: 55, partition: 1
key: 8, value: 今天的<--刘向前-->很美很美哦~, hashCode: 56, partition: 2
key: 9, value: 今天的<--姚慧莹-->很美很美哦~, hashCode: 57, partition: 0
key: 10, value: 今天的<--杨柳-->很美很美哦~, hashCode: 1567, partition: 1
key: 11, value: 今天的<--姚慧莹-->很美很美哦~, hashCode: 1568, partition: 2
key: 12, value: 今天的<--周 新-->很美很美哦~, hashCode: 1569, partition: 0
key: 13, value: 今天的<--姚慧莹-->很美很美哦~, hashCode: 1570, partition: 1
key: 14, value: 今天的<--姚慧莹-->很美很美哦~, hashCode: 1571, partition: 2
key: 15, value: 今天的<--刘向前-->很美很美哦~, hashCode: 1572, partition: 0
key: 16, value: 今天的<--刘向前-->很美很美哦~, hashCode: 1573, partition: 1
key: 17, value: 今天的<--杨柳-->很美很美哦~, hashCode: 1574, partition: 2
key: 18, value: 今天的<--刘向前-->很美很美哦~, hashCode: 1575, partition: 0
key: 19, value: 今天的<--杨柳-->很美很美哦~, hashCode: 1576, partition: 1
key: 20, value: 今天的<--姚慧莹-->很美很美哦~, hashCode: 1598, partition: 2
key: 21, value: 今天的<--杨柳-->很美很美哦~, hashCode: 1599, partition: 0
消费者程序的终端输出如下:
外部开始时间:1521991118178
线程ID:20线程开始时间:1521991123182
线程ID:21线程开始时间:1521991123182
线程ID:23线程开始时间:1521991123182
线程ID:22线程开始时间:1521991123182
CurrentThreadID: 22 offset: 78 partition: 1 key: 1 value: 今天的<--刘向前-->很美很美哦~
CurrentThreadID: 22 offset: 79 partition: 1 key: 4 value: 今天的<--杨柳-->很美很美哦~
CurrentThreadID: 22 offset: 80 partition: 1 key: 7 value: 今天的<--杨柳-->很美很美哦~
CurrentThreadID: 22 offset: 81 partition: 1 key: 10 value: 今天的<--杨柳-->很美很美哦~
CurrentThreadID: 22 offset: 82 partition: 1 key: 13 value: 今天的<--姚慧莹-->很美很美哦~
CurrentThreadID: 23 offset: 81 partition: 0 key: 3 value: 今天的<--刘向前-->很美很美哦~
CurrentThreadID: 23 offset: 82 partition: 0 key: 6 value: 今天的<--姚慧莹-->很美很美哦~
CurrentThreadID: 23 offset: 83 partition: 0 key: 9 value: 今天的<--姚慧莹-->很美很美哦~
CurrentThreadID: 23 offset: 84 partition: 0 key: 12 value: 今天的<--周 新-->很美很美哦~
CurrentThreadID: 23 offset: 85 partition: 0 key: 15 value: 今天的<--刘向前-->很美很美哦~
CurrentThreadID: 23 offset: 86 partition: 0 key: 18 value: 今天的<--刘向前-->很美很美哦~
CurrentThreadID: 22 offset: 83 partition: 1 key: 16 value: 今天的<--刘向前-->很美很美哦~
CurrentThreadID: 23 offset: 87 partition: 0 key: 21 value: 今天的<--杨柳-->很美很美哦~
CurrentThreadID: 21 offset: 78 partition: 2 key: 2 value: 今天的<--刘向前-->很美很美哦~
CurrentThreadID: 22 offset: 84 partition: 1 key: 19 value: 今天的<--杨柳-->很美很美哦~
CurrentThreadID: 21 offset: 79 partition: 2 key: 5 value: 今天的<--姚慧莹-->很美很美哦~
CurrentThreadID: 21 offset: 80 partition: 2 key: 8 value: 今天的<--刘向前-->很美很美哦~
CurrentThreadID: 21 offset: 81 partition: 2 key: 11 value: 今天的<--姚慧莹-->很美很美哦~
CurrentThreadID: 21 offset: 82 partition: 2 key: 14 value: 今天的<--姚慧莹-->很美很美哦~
CurrentThreadID: 21 offset: 83 partition: 2 key: 17 value: 今天的<--杨柳-->很美很美哦~
CurrentThreadID: 21 offset: 84 partition: 2 key: 20 value: 今天的<--姚慧莹-->很美很美哦~
消费者shell的终端输出如下:
$ kafka-console-consumer.sh --topic hadoop --zookeeper uplooking01:2181,uplooking02:2181,uplooking03:2181
今天的<--刘向前-->很美很美哦~
今天的<--姚慧莹-->很美很美哦~
今天的<--刘向前-->很美很美哦~
今天的<--姚慧莹-->很美很美哦~
今天的<--姚慧莹-->很美很美哦~
今天的<--杨柳-->很美很美哦~
今天的<--姚慧莹-->很美很美哦~
今天的<--刘向前-->很美很美哦~
今天的<--姚慧莹-->很美很美哦~
今天的<--姚慧莹-->很美很美哦~
今天的<--周 新-->很美很美哦~
今天的<--刘向前-->很美很美哦~
今天的<--刘向前-->很美很美哦~
今天的<--杨柳-->很美很美哦~
今天的<--刘向前-->很美很美哦~
今天的<--杨柳-->很美很美哦~
今天的<--杨柳-->很美很美哦~
今天的<--杨柳-->很美很美哦~
今天的<--姚慧莹-->很美很美哦~
今天的<--刘向前-->很美很美哦~
今天的<--杨柳-->很美很美哦~
分析
因为使用kafka shell的消费者的group.id是随机生成的,所以其肯定可以消费到topic下partition的消息,这是属于组间的消费。
而由于在消费者的程序代码中,4个线程都是使用同一个group.id的(都是使用consumer.properties这个配置文件),按照理论知识的理解,因为topic hadoop只有3个partition,所以只能有3个线程即3个consumer进行消息的消费,而观察输出,通过线程ID,发现确实只有三个线程消费了topic中的消息,这也验证了kafka组内消息的消费形式。
Kafka性能测试
参考文档:https://cwiki.apache.org/confluence/display/KAFKA/Performance+testing
生产能力测试
在kafka的安装目录的bin里有性能的评估工具bin/kafka-producer-perf-test.sh
,主要输出4项指标,总共发送消息量(以MB为单位),每秒发送消息量(MB/second),发送消息总数,每秒发送消息数(records/second)。
测试如下:
[uplooking@uplooking01 ~]$ kafka-producer-perf-test.sh --topic flume-kafka --num-records 1000000 --producer-props bootstrap.servers=uplooking01:9092,uplooking02:9092,uplooking03:9092 --throughput 10000 --record-size 100
49972 records sent, 9994.4 records/sec (0.95 MB/sec), 3.1 ms avg latency, 258.0 max latency.
50200 records sent, 10040.0 records/sec (0.96 MB/sec), 2.4 ms avg latency, 141.0 max latency.
50020 records sent, 10004.0 records/sec (0.95 MB/sec), 1.2 ms avg latency, 19.0 max latency.
50010 records sent, 10000.0 records/sec (0.95 MB/sec), 2.3 ms avg latency, 127.0 max latency.
50010 records sent, 10002.0 records/sec (0.95 MB/sec), 1.3 ms avg latency, 24.0 max latency.
50020 records sent, 10004.0 records/sec (0.95 MB/sec), 2.4 ms avg latency, 186.0 max latency.
50010 records sent, 10002.0 records/sec (0.95 MB/sec), 15.1 ms avg latency, 466.0 max latency.
50020 records sent, 10002.0 records/sec (0.95 MB/sec), 11.1 ms avg latency, 405.0 max latency.
50000 records sent, 10000.0 records/sec (0.95 MB/sec), 1.2 ms avg latency, 19.0 max latency.
50030 records sent, 10004.0 records/sec (0.95 MB/sec), 1.2 ms avg latency, 20.0 max latency.
50000 records sent, 10000.0 records/sec (0.95 MB/sec), 1.2 ms avg latency, 30.0 max latency.
50010 records sent, 10002.0 records/sec (0.95 MB/sec), 1.3 ms avg latency, 19.0 max latency.
49990 records sent, 9998.0 records/sec (0.95 MB/sec), 1.4 ms avg latency, 49.0 max latency.
50033 records sent, 10006.6 records/sec (0.95 MB/sec), 37.9 ms avg latency, 617.0 max latency.
50010 records sent, 10002.0 records/sec (0.95 MB/sec), 1.5 ms avg latency, 74.0 max latency.
50007 records sent, 10001.4 records/sec (0.95 MB/sec), 1.3 ms avg latency, 19.0 max latency.
50000 records sent, 10000.0 records/sec (0.95 MB/sec), 1.8 ms avg latency, 132.0 max latency.
50010 records sent, 10002.0 records/sec (0.95 MB/sec), 1.2 ms avg latency, 15.0 max latency.
50020 records sent, 10000.0 records/sec (0.95 MB/sec), 1.9 ms avg latency, 121.0 max latency.
1000000 records sent, 9999.200064 records/sec (0.95 MB/sec), 4.96 ms avg latency, 617.00 ms max latency, 1 ms 50th, 3 ms 95th, 105 ms 99th, 541 ms 99.9th.
参数说明如下:
--num-records 1000000 总共生产的消息数量
--throughput 10000 每秒需要生产的消息数量
--record-size 100 每条消息的大小,单位为字节
消费能力测试
[uplooking@uplooking01 ~]$ kafka-consumer-perf-test.sh --topic flume-kafka --messages 1000000 --broker-list uplooking01:9092,uplooking02:9092,uplooking03:9092 --threads 3 --zookeeper uplooking01:2181,uplooking02:2181,uplooking03:2181
start.time, end.time, data.consumed.in.MB, MB.sec, data.consumed.in.nMsg, nMsg.sec
2018-03-26 05:17:21:185, 2018-03-26 05:17:22:458, 97.3055, 76.4380, 1020325, 801512.1760
上面的测试为需要消费一百万条消息,输出的参数说明如下:
开始时间 结束时间 消费消息总大小 每秒消费大小 消费消息总条数 每秒消费条数
start.time, end.time, data.consumed.in.MB, MB.sec, data.consumed.in.nMsg, nMsg.sec
2018-03-26 05:17:21:185, 2018-03-26 05:17:22:458, 97.3055, 76.4380, 1020325, 801512.1760