qianye1001 opened a new issue, #10488:
URL: https://github.com/apache/rocketmq/issues/10488

   ### Before Creating the Enhancement Request
   
   - [x] I have confirmed that this should be classified as an enhancement 
rather than a bug/feature.
   
   ### Summary
   
   In the Proxy receive-message path, `GrpcConverter.buildMessage()` uses 
`ByteString.copyFrom(messageExt.getBody())` to set the message body in the gRPC 
response. `ByteString.copyFrom()` allocates a new `byte[]` and copies the 
entire body content via `System.arraycopy`, even though the source `byte[]` is 
a standalone allocation that is never mutated after this point. This can be 
replaced with `UnsafeByteOperations.unsafeWrap()` to wrap the existing `byte[]` 
by reference (zero-copy).
   
   ### Motivation
   
   On the Proxy receive path (pop message from Broker → gRPC response to 
client), the message body undergoes the following copy chain:
   
   | Step | Location | Operation | Copy? |
   |------|----------|-----------|-------|
   | 1 | `RemotingCommand.decode()` | `new byte[bodyLength]` + 
`byteBuffer.readBytes(bodyData)` — Netty ByteBuf → heap byte[] | Necessary 
(ByteBuf released after decode) |
   | 2 | `MessageDecoder.decode()` | `new byte[bodyLen]` + 
`byteBuffer.get(body)` — extract per-message body from batch | Necessary (split 
individual messages) |
   | 3 | `GrpcConverter.buildMessage()` | 
`ByteString.copyFrom(messageExt.getBody())` — body byte[] → new byte[] inside 
ByteString | **Unnecessary** |
   | 4 | gRPC/protobuf serialization | ByteString → network buffer | Necessary |
   
   Step 3 is redundant because:
   - The `body` byte[] was freshly allocated in `MessageDecoder.decode()` (line 
508), sized exactly to `bodyLen`. It is not a slice of a larger buffer.
   - After `GrpcConverter.buildMessage()`, no code path mutates 
`messageExt.getBody()`. The `MessageExt` object is only used for reading 
receipt handles (in auto-renew) and is then eligible for GC.
   - `ByteString` is immutable by contract; `UnsafeByteOperations.unsafeWrap()` 
simply wraps the byte[] by reference into a `LiteralByteString` without 
copying, which is safe as long as the caller guarantees the byte[] won't be 
mutated — which holds here.
   
   At high throughput (e.g., 100k msgs/s with 4KB body), this eliminates 
~400MB/s of unnecessary heap allocation and `System.arraycopy` on the receive 
path.
   
   ### Describe the Solution You'd Like
   
   In 
`proxy/src/main/java/org/apache/rocketmq/proxy/grpc/v2/common/GrpcConverter.java`,
 change line 105:
   
   ```java
   // Before:
   .setBody(ByteString.copyFrom(messageExt.getBody()))
   
   // After:
   .setBody(UnsafeByteOperations.unsafeWrap(messageExt.getBody()))
   ```
   
   Add the import:
   ```java
   import com.google.protobuf.UnsafeByteOperations;
   ```
   
   `UnsafeByteOperations` is a public API in protobuf-java (available since 
3.x), already on the classpath. The change is a single line with no behavioral 
difference — the ByteString content is identical, only the internal copy is 
skipped.
   
   ### Describe Alternatives You've Considered
   
   - **`ByteString.wrap(byte[])`**: This is a package-private method in 
protobuf-java and cannot be called from outside `com.google.protobuf`. 
`UnsafeByteOperations.unsafeWrap()` is the public equivalent that delegates to 
it.
   - **Keep `copyFrom` and accept the overhead**: At scale the unnecessary copy 
is measurable in both CPU (arraycopy) and GC pressure (short-lived byte[] 
allocations).
   
   ### Additional Context
   
   - Protobuf version: 3.20.1 (as used by RocketMQ).
   - `UnsafeByteOperations.unsafeWrap(byte[])` internally calls 
`ByteString.wrap(byte[])`, which constructs a `LiteralByteString` with a direct 
reference to the passed byte[] — confirmed by bytecode inspection.
   - This is the receive (consumer) path only. The send (producer) path has a 
symmetric issue (`ByteString.toByteArray()` in 
`SendMessageActivity.buildMessage()`), which could be addressed in a separate 
issue.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to