開篇導語
Flink將數據sink至Kafka的過程中,在初始化生產者對象FlinkKafkaProducer時通常會採用默認的分區器和序列化器,這樣數據只會發送至指定Topic的某一個分區中。對於存在多分區的Topic我們一般要自定義分區器和序列化器,指定數據發送至不同分區的邏輯。
此篇博客所涉及的組件版本
Flink:1.10.0
Kafka:2.3.0
序列化器
在Kafka生產者將數據寫入至Kafka集羣中時,爲了能夠在網絡中傳輸數據對象,需要先將數據進行序列化處理,對於初學者來說,在初始化生產者對象時,一般都會採用默認的序列化器。
默認的序列化器不會對數據進行任何操作,也不會生成key。如果我們需要指定數據的key或者在數據發送前進行一些定製化的操作,那麼我們就需要自定義序列化器,並且在初始化生產者對象時指定我們自己的序列化器。
分區器
對於Kakfa中一個topic存在多個分區的情況下,我們怎麼知道發送的數據會被分配到哪個分區呢,這時候就要通過分區器來進行區分。
在Kafka中,主要有以下四種數據分區策略
第一種分區策略:給定了分區號,直接將數據發送到指定的分區裏面去
第二種分區策略:沒有給定分區號,給定數據的key值,通過key取hashCode進行分區
第三種分區策略:既沒有給定分區號,也沒有給定key值,直接輪循進行分區
第四種分區策略:自定義分區
分區器就是以上分區策略的代碼實現。
Flink中的Kafka序列化器
源碼解讀
在之前的Flink版中中,自定義Kafka序列化器都是實現KeyedSerializationSchema接口,看一下它的源碼:
//表示當前接口已經不推薦使用
@Deprecated
@PublicEvolving
//當前接口實現時需要指定生產者所要傳輸的對象類型
public interface KeyedSerializationSchema<T> extends Serializable {
//根據傳入的對象生成key並序列化爲字節數組
//如果沒有生成key,這個方法可返回null。
byte[] serializeKey(T element);
//根據傳入的對象生成value並序列化爲字節數組
byte[] serializeValue(T element);
//根據傳入的對象指定需要發送的Topic
//此方法可以返回null,因爲在初始化生產者對象的時候就已經指定了Topic。
String getTargetTopic(T element);
}
以上接口存在三個方法,但是每個輸入的參數都是一樣的,代碼複用性低。
於是現在的Flink版本一般推薦實現KafkaSerializationSchema接口來實現序列化器,看一下它的源碼:
//當前接口實現時需要指定生產者所要傳輸的對象類型
@PublicEvolving
public interface KafkaSerializationSchema<T> extends Serializable {
/**
* Serializes given element and returns it as a {@link ProducerRecord}.
*
* @param element element to be serialized
* @param timestamp timestamp (can be null)
* @return Kafka {@link ProducerRecord}
*/
ProducerRecord<byte[], byte[]> serialize(T element, @Nullable Long timestamp);
}
可以看到當前接口只需要實現serialize方法,根據傳入的對象構造並返回一個ProducerRecord<byte[], byte[]>對象。
我們來看一下ProducerRecord類的源碼:
package org.apache.kafka.clients.producer;
import java.util.Objects;
import org.apache.kafka.common.header.Header;
import org.apache.kafka.common.header.Headers;
import org.apache.kafka.common.header.internals.RecordHeaders;
public class ProducerRecord<K, V> {
private final String topic;
private final Integer partition;
private final Headers headers;
private final K key;
private final V value;
private final Long timestamp;
public ProducerRecord(String topic, Integer partition, Long timestamp, K key, V value, Iterable<Header> headers) {
if (topic == null) {
throw new IllegalArgumentException("Topic cannot be null.");
} else if (timestamp != null && timestamp < 0L) {
throw new IllegalArgumentException(String.format("Invalid timestamp: %d. Timestamp should always be non-negative or null.", timestamp));
} else if (partition != null && partition < 0) {
throw new IllegalArgumentException(String.format("Invalid partition: %d. Partition number should always be non-negative or null.", partition));
} else {
this.topic = topic;
this.partition = partition;
this.key = key;
this.value = value;
this.timestamp = timestamp;
this.headers = new RecordHeaders(headers);
}
}
public ProducerRecord(String topic, Integer partition, Long timestamp, K key, V value) {
this(topic, partition, timestamp, key, value, (Iterable)null);
}
public ProducerRecord(String topic, Integer partition, K key, V value, Iterable<Header> headers) {
this(topic, partition, (Long)null, key, value, headers);
}
public ProducerRecord(String topic, Integer partition, K key, V value) {
this(topic, partition, (Long)null, key, value, (Iterable)null);
}
public ProducerRecord(String topic, K key, V value) {
this(topic, (Integer)null, (Long)null, key, value, (Iterable)null);
}
public ProducerRecord(String topic, V value) {
this(topic, (Integer)null, (Long)null, (Object)null, value, (Iterable)null);
}
public String topic() {
return this.topic;
}
public Headers headers() {
return this.headers;
}
public K key() {
return this.key;
}
public V value() {
return this.value;
}
public Long timestamp() {
return this.timestamp;
}
public Integer partition() {
return this.partition;
}
public String toString() {
String headers = this.headers == null ? "null" : this.headers.toString();
String key = this.key == null ? "null" : this.key.toString();
String value = this.value == null ? "null" : this.value.toString();
String timestamp = this.timestamp == null ? "null" : this.timestamp.toString();
return "ProducerRecord(topic=" + this.topic + ", partition=" + this.partition + ", headers=" + headers + ", key=" + key + ", value=" + value + ", timestamp=" + timestamp + ")";
}
public boolean equals(Object o) {
if (this == o) {
return true;
} else if (!(o instanceof ProducerRecord)) {
return false;
} else {
ProducerRecord<?, ?> that = (ProducerRecord)o;
return Objects.equals(this.key, that.key) && Objects.equals(this.partition, that.partition) && Objects.equals(this.topic, that.topic) && Objects.equals(this.headers, that.headers) && Objects.equals(this.value, that.value) && Objects.equals(this.timestamp, that.timestamp);
}
}
public int hashCode() {
int result = this.topic != null ? this.topic.hashCode() : 0;
result = 31 * result + (this.partition != null ? this.partition.hashCode() : 0);
result = 31 * result + (this.headers != null ? this.headers.hashCode() : 0);
result = 31 * result + (this.key != null ? this.key.hashCode() : 0);
result = 31 * result + (this.value != null ? this.value.hashCode() : 0);
result = 31 * result + (this.timestamp != null ? this.timestamp.hashCode() : 0);
return result;
}
}
可以看到ProducerRecord在初始化時需要以下參數:
String topic;//分區名稱,不可以爲空
Integer partition;//當前記錄需要寫入的分區值,可以爲空
Headers headers;//kafka頭信息,可以爲空
K key;//當前記錄的key,可以爲空
V value;//當前記錄的實際value,不可以爲空
Long timestamp;//指定生產者創建當前記錄的時間戳,可以爲空
在ProducerRecord的多個重構的構造函數中,參數最少的一個只需要傳入topic名稱和value即可。其他值都可爲空。
public ProducerRecord(String topic, V value) {
this(topic, (Integer)null, (Long)null, (Object)null, value, (Iterable)null);
}
自定義序列化器示例
基於上述知識,我們可以通過實現KafkaSerializationSchema自定義序列化器,以下是一個最簡單的自定義序列化器
package lenrnflink;
import org.apache.flink.streaming.api.datastream.DataStream;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.connectors.kafka.FlinkKafkaProducer;
import org.apache.flink.streaming.connectors.kafka.KafkaSerializationSchema;
import org.apache.kafka.clients.producer.ProducerRecord;
import javax.annotation.Nullable;
import java.util.Properties;
public class Test {
//實現KafkaSerializationSchema接口來自定義序列化器,傳入的參數設定爲String類型。
public static class MySerializationSchema implements KafkaSerializationSchema<String> {
@Override
public ProducerRecord<byte[], byte[]> serialize(String element, @Nullable Long timestamp) {
//返回一條記錄,指定topic爲test,分區爲0,key爲null,value爲傳入對象轉化而成的字節數組。
return new ProducerRecord<>("test",0,null,element.getBytes());
}
}
public static void main(String[] args) throws Exception {
StreamExecutionEnvironment environment = StreamExecutionEnvironment.getExecutionEnvironment();
Properties properties = new Properties();
properties.setProperty("bootstrap.servers", "cdh1:9092,cdh2:9092,cdh3:9092,cdh4:9092,cdh5:9092");
//初始化Flink-Kafka生產者,並將自定義序列化器作爲參數傳遞。
FlinkKafkaProducer<String> kafkaSink = new FlinkKafkaProducer<String>("test",new MySerializationSchema(),properties,FlinkKafkaProducer.Semantic.AT_LEAST_ONCE);
DataStream<String> dataStream = environment.readTextFile("E:/test.txt","GB2312");
dataStream.addSink(kafkaSink);
environment.execute("WordCount");
}
}
以上只是一個最簡單的自定義序列化器,用戶在自己編寫的時候,可以結合業務邏輯,根據傳遞過來的對象,在序列化器中指定其topic,partition,key和value等。
Flink中的Kafka分區器
源碼解讀
在Flink中,自定義Kafka分區器需要繼承FlinkKafkaPartitioner抽象類,看一下源碼:
@PublicEvolving
public abstract class FlinkKafkaPartitioner<T> implements Serializable {
private static final long serialVersionUID = -9086719227828020494L;
//當前方法接收上游傳過來的並行實例ID和並行實例總數。
//不是抽象方法,可以不用在子類實現。
public void open(int parallelInstanceId, int parallelInstances) {
// overwrite this method if needed.
}
/**
* @param record the record value
* @param key serialized key of the record
* @param value serialized value of the record
* @param targetTopic target topic for the record
* @param partitions found partitions for the target topic
* @return the id of the target partition
*/
//當前方法爲抽象方法,需要在子類中實現其方法體。指定其具體的分區。
public abstract int partition(T record, byte[] key, byte[] value, String targetTopic, int[] partitions);
}
此類主要爲open方法用於接收上游傳過來的並行實例ID和總數。和partition抽象方法,進行指定分區的具體操作。
Flink中根據此實現了一個默認的分區器FlinkFixedPartitioner,看其源碼:
@PublicEvolving
public class FlinkFixedPartitioner<T> extends FlinkKafkaPartitioner<T> {
private static final long serialVersionUID = -3785320239953858777L;
private int parallelInstanceId;
@Override
public void open(int parallelInstanceId, int parallelInstances) {
Preconditions.checkArgument(parallelInstanceId >= 0, "Id of this subtask cannot be negative.");
Preconditions.checkArgument(parallelInstances > 0, "Number of subtasks must be larger than 0.");
//接收上游傳過來的並行實例ID,並將值傳遞給成員變量,用於分區操作。
this.parallelInstanceId = parallelInstanceId;
}
@Override
public int partition(T record, byte[] key, byte[] value, String targetTopic, int[] partitions) {
Preconditions.checkArgument(
partitions != null && partitions.length > 0,
"Partitions of the target topic is empty.");
//通過Flink並行實例的id去和Kafka分區的數量取餘來決定這個實例的數據寫到哪個Kafka分區
return partitions[parallelInstanceId % partitions.length];
}
@Override
public boolean equals(Object o) {
return this == o || o instanceof FlinkFixedPartitioner;
}
@Override
public int hashCode() {
return FlinkFixedPartitioner.class.hashCode();
}
}
通過源碼可以看出,當前分區器主要是通過Flink並行實例的id和Kafka分區的數量取餘來決定這個實例的數據寫到哪個Kafka分區,並且一個實例只寫Kafka中的一個分區。
這樣做的好處最大限度的利用了Flink和Kafka的可擴展性,提高數據處理效率。
自定義分區器示例
package lenrnflink;
import org.apache.flink.api.common.serialization.SimpleStringSchema;
import org.apache.flink.streaming.api.datastream.DataStream;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.connectors.kafka.FlinkKafkaProducer;
import org.apache.flink.streaming.connectors.kafka.KafkaSerializationSchema;
import org.apache.flink.streaming.connectors.kafka.partitioner.FlinkKafkaPartitioner;
import org.apache.kafka.clients.producer.ProducerRecord;
import javax.annotation.Nullable;
import java.util.Optional;
import java.util.Properties;
public class Test {
//繼承FlinkKafkaPartitioner自定義分區器
public static class KafkaPartitioner extends FlinkKafkaPartitioner<String> {
@Override
//當前分區器的邏輯爲,如果傳入的對象開頭爲1,則分區的值返回1,開頭爲2,則分區的值返回2,其他則返回0。
public int partition(String s, byte[] bytes, byte[] bytes1, String s2, int[] ints) {
if(s.startsWith("1")){
return 1;
}
if(s.startsWith("2")){
return 2;
}
return 0;
}
}
public static void main(String[] args) throws Exception {
StreamExecutionEnvironment environment = StreamExecutionEnvironment.getExecutionEnvironment();
Properties properties = new Properties();
properties.setProperty("bootstrap.servers", "cdh1:9092,cdh2:9092,cdh3:9092,cdh4:9092,cdh5:9092");
//必須通過Optional容器添加自定義分區器。
Optional<FlinkKafkaPartitioner> ps = Optional.of(new KafkaPartitioner());
//初始化Flink-kafka生產者時,指定自定義分區器。
FlinkKafkaProducer kafkaSink = new FlinkKafkaProducer("test",new SimpleStringSchema(), properties, ps);
DataStream<String> dataStream = environment.readTextFile("E:/test.txt","GB2312");
dataStream.addSink(kafkaSink);
environment.execute("WordCount");
}
}
以上爲一個非常簡單的自定義分區器,用戶可根據業務需要制訂一個屬於自己的分區器。
結束語
以下爲FlinkKafkaProducer源碼
/** @deprecated */
@Deprecated
public FlinkKafkaProducer(String brokerList, String topicId, SerializationSchema<IN> serializationSchema) {
this(topicId, (KeyedSerializationSchema)(new KeyedSerializationSchemaWrapper(serializationSchema)), getPropertiesFromBrokerList(brokerList), (Optional)Optional.of(new FlinkFixedPartitioner()));
}
/** @deprecated */
@Deprecated
public FlinkKafkaProducer(String topicId, SerializationSchema<IN> serializationSchema, Properties producerConfig) {
this(topicId, (KeyedSerializationSchema)(new KeyedSerializationSchemaWrapper(serializationSchema)), producerConfig, (Optional)Optional.of(new FlinkFixedPartitioner()));
}
/** @deprecated */
@Deprecated
public FlinkKafkaProducer(String topicId, SerializationSchema<IN> serializationSchema, Properties producerConfig, Optional<FlinkKafkaPartitioner<IN>> customPartitioner) {
this(topicId, (KeyedSerializationSchema)(new KeyedSerializationSchemaWrapper(serializationSchema)), producerConfig, (Optional)customPartitioner);
}
/** @deprecated */
@Deprecated
public FlinkKafkaProducer(String brokerList, String topicId, KeyedSerializationSchema<IN> serializationSchema) {
this(topicId, serializationSchema, getPropertiesFromBrokerList(brokerList), Optional.of(new FlinkFixedPartitioner()));
}
/** @deprecated */
@Deprecated
public FlinkKafkaProducer(String topicId, KeyedSerializationSchema<IN> serializationSchema, Properties producerConfig) {
this(topicId, serializationSchema, producerConfig, Optional.of(new FlinkFixedPartitioner()));
}
/** @deprecated */
@Deprecated
public FlinkKafkaProducer(String topicId, KeyedSerializationSchema<IN> serializationSchema, Properties producerConfig, FlinkKafkaProducer.Semantic semantic) {
this(topicId, serializationSchema, producerConfig, Optional.of(new FlinkFixedPartitioner()), semantic, 5);
}
/** @deprecated */
@Deprecated
public FlinkKafkaProducer(String defaultTopicId, KeyedSerializationSchema<IN> serializationSchema, Properties producerConfig, Optional<FlinkKafkaPartitioner<IN>> customPartitioner) {
this(defaultTopicId, serializationSchema, producerConfig, customPartitioner, FlinkKafkaProducer.Semantic.AT_LEAST_ONCE, 5);
}
/** @deprecated */
@Deprecated
public FlinkKafkaProducer(String defaultTopicId, KeyedSerializationSchema<IN> serializationSchema, Properties producerConfig, Optional<FlinkKafkaPartitioner<IN>> customPartitioner, FlinkKafkaProducer.Semantic semantic, int kafkaProducersPoolSize) {
this(defaultTopicId, serializationSchema, (FlinkKafkaPartitioner)customPartitioner.orElse((Object)null), (KafkaSerializationSchema)null, producerConfig, semantic, kafkaProducersPoolSize);
}
public FlinkKafkaProducer(String defaultTopic, KafkaSerializationSchema<IN> serializationSchema, Properties producerConfig, FlinkKafkaProducer.Semantic semantic) {
this(defaultTopic, serializationSchema, producerConfig, semantic, 5);
}
public FlinkKafkaProducer(String defaultTopic, KafkaSerializationSchema<IN> serializationSchema, Properties producerConfig, FlinkKafkaProducer.Semantic semantic, int kafkaProducersPoolSize) {
this(defaultTopic, (KeyedSerializationSchema)null, (FlinkKafkaPartitioner)null, serializationSchema, producerConfig, semantic, kafkaProducersPoolSize);
}
在閱讀Flink中的Kafka生產者源碼FlinkKafkaProducer時發現其多個構造函數,凡是參數中包含FlinkKafkaProducer的都被標記爲了deprecated,說明官方已經不推薦使用自定義分區器來進行數據的分區操作。
只有參數包含KafkaSerializationSchema的兩個構造函數是正常的,說明現在官方推薦使用KafkaSerializationSchema接口來進行序列化的操作。
並且閱讀源碼的過程中可以發現,KafkaSerializationSchema中也有對數據的分區操作。只需要結合KafkaContextAware接口即可實現獲取Flink並行實例ID和數量的功能。