Aditya Kousik created KAFKA-21122:
-------------------------------------
Summary: ByteBuffer serialize method to the producer Serializer
Key: KAFKA-21122
URL: https://issues.apache.org/jira/browse/KAFKA-21122
Project: Kafka
Issue Type: Improvement
Components: clients, producer
Reporter: Aditya Kousik
Assignee: Aditya Kousik
The consumer read path avoids a per-record copy for the record value. KIP-863
added a ByteBuffer method to Deserializer so that CompletedFetch#parseRecord
can hand the deserializer the record value straight from the fetch buffer:
{code:java}
default T deserialize(String topic, Headers headers, ByteBuffer data) { ... }
{code}
The default falls back to the {{byte[]}} method, so existing deserializers are
unaffected, and {{ByteBufferDeserializer}} overrides it to skip the copy.
The producer has no matching method. {{Serializer}} only returns {{{}byte[]{}}}:
{code:java}
byte[] serialize(String topic, Headers headers, T data);{code}
So a producer whose record value is already in a ByteBuffer has to copy it into
a byte[] for every record. This is the case for a caller that holds data off
the Java heap or in a reused buffer, for example a Netty ByteBuf exposed
through nioBuffer(), or a service that forwards a payload it does not read.
We should consider adding the missing method so the write side matches the read
side that KIP-863 already added.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
