Status

Current state: Accepted

Discussion thread: https://lists.apache.org/thread/5fmgl51x7h3gsx9mn7xq0n28j700rys0

Voting thread: https://lists.apache.org/thread/jgtzlb4l60hs6t3o67x53w8gonk2ffvo

PR: https://github.com/apache/kafka/pull/12685

JIRA: 

Motivation

Currently, we use Serializer#serialize(String, Headers, T) in KafkaProducer#doSend(ProducerRecord, Callback) to serialize key and value. First, we call Serializer#serialize(String, Headers, T) to convert T into byte[], then use Utils#wrapNullable(byte[]) to convert byte[] into ByteBuffer, and finally write ByteBuffer into MemoryRecordsBuilder through DefaultRecord#writeTo(DataOutputStream, int, long, ByteBuffer, ByteBuffer, Header[]).

Why don't we add a serializeToByteBuffer(String, Headers, T) method to Serializer, and then use Serializer#serializeToByteBuffer(String, Headers, T) in KafkaProducer#doSend(ProducerRecord, Callback)? If T is an instance of ByteBuffer or T is based on ByteBuffer, we would reduce a lot of memory allocation and memory copying.

Additionally, I plan to ultimately replace byte[] with ByteBuffer in Serializer: in 3.x versions, both Serializer#serializeToByteBuffer() and Serializer#serialize() exist. Starting from version 4.0, Serializer#serializeToByteBuffer() will be removed and its implementation will replace Serializer#serialize(), which will return ByteBuffer by default, and Serializer#serialize() will not modify the input ByteBuffer.

Public Interfaces

We propose adding default method Serializer#serializeToByteBuffer(String, T), Serializer#serializeToByteBuffer(String, Headers, T) and Partitioner#partition(String, Object, ByteBuffer, Object, ByteBuffer, Cluster):

Serializer

default ByteBuffer serializeToByteBuffer(String topic, T data) {
    return wrapNullable(serialize(topic, data));
}

default ByteBuffer serializeToByteBuffer(String topic, Headers headers, T data) {
    return wrapNullable(serialize(topic, headers, data));
}

ByteBufferSerializer

/** 
* Note that this method will modify the position and limit of the input ByteBuffer. 
* 
* @param topic   topic associated with data 
* @param data    typed data 
* @return serialized ByteBuffer
*/
@Override
public ByteBuffer serializeToByteBuffer(String topic, ByteBuffer data) {
    if (data == null) {
        return null;
    }

    // Consider that ByteBuffer#wrap(byte[]) return a ByteBuffer that does not need to call flip().
    if (data.position() > 0) {
        data.flip();
    }
    return data;
}

@Override
public ByteBuffer serializeToByteBuffer(String topic, Headers headers, ByteBuffer data) {
    return serializeToByteBuffer(topic, data);
}

Partitioner

default int partition(String topic, Object key, ByteBuffer keyBytes, Object value, ByteBuffer valueBytes, Cluster cluster) {
    return partition(topic, key, toNullableArray(keyBytes), value, toNullableArray(valueBytes), cluster);
}

Proposed Changes

There are the following changes:

• Serializer
  • Adding default method Serializer#serializeToByteBuffer(String, T) and Serializer#serializeToByteBuffer(String, Headers, T);
  • Implement serializeToByteBuffer(String, T) and serializeToByteBuffer(String, Headers, T) for ByteBufferSerializer;
  • Using Serializer#serializeToByteBuffer(String, Headers, T) to serialize key&value in KafkaProducer#doSend(ProducerRecord, Callback).
• Partitioner
  • Adding default method Partitioner#partition(String, Object, ByteBuffer, Object, ByteBuffer, Cluster);
  • Implement partition(String, Object, ByteBuffer, Object, ByteBuffer, Cluster) for RoundRobinPartitioner;
  • Using Partitioner#partition(String, Object, ByteBuffer, Object, ByteBuffer, Cluster) in KafkaProducer#partition(ProducerRecord, ByteBuffer, ByteBuffer, Cluster).
• Utils
  • Adding method murmur2(ByteBuffer data).

Compatibility, Deprecation, and Migration Plan

• Compatibility

  • We just add default methods Serializer#serializeToByteBuffer(String, T), Serializer#serializeToByteBuffer(String, Headers, T) and Partitioner#partition(String, Object, ByteBuffer, Object, ByteBuffer, Cluster) which is compatible with the existing Serializers.

  • The impact on ByteBuffer offsets is consistent after calling ByteBufferSerializer#serialize(String, ByteBuffer) and ByteBufferSerializer#serializeToByteBuffer(String, Headers, T).

Rejected Alternatives

There is no alternative.