Avoid allocating ByteBuf's on non-SDK threads

[dagnir]

Expected Behavior

We use the default PooledByteBufAllocator in Netty which creates per-thread caches stored in a ThreadLocal. To avoid polluting threads we don't own, we should ensureByteBufs allocated by the Netty client are allocated only on SDK threads.

Current Behavior

It's possible for the Netty client to allocate within a customer's thread, for example, in the onNext of an SDK Subscriber with a subscription to a customer provided Publisher:

aws-sdk-java-v2/http-clients/netty-nio-client/src/main/java/software/amazon/awssdk/http/nio/netty/internal/NettyRequestExecutor.java

Lines 436 to 446 in 1fee753

	public void onNext(ByteBuffer byteBuffer) {
	if (done) {
	return;
	}
	try {
	int newLimit = clampedBufferLimit(byteBuffer.remaining());
	byteBuffer.limit(newLimit);
	ByteBuf buffer = channel.alloc().buffer(byteBuffer.remaining());
	buffer.writeBytes(byteBuffer);
	HttpContent content = new DefaultHttpContent(buffer);

Possible Solution

The easiest solution would be to allocate using io.netty.buffer.Unpooled. Otherwise, we should defer the allocation and run it in the EventLoop.

Steps to Reproduce (for bugs)

Context

Your Environment

• AWS Java SDK version used: 2.5.10
• JDK version used:
• Operating System and version:

[dagnir]

Now that I think about it, using Unpooled#wrappedBuffer(ByteBuffer) will probably be more efficient since there's no actual copying involved.

[zoewangg]

Hmm, I wonder if we should use unpooledBuffer. As per the Netty Best Practices slides, there are some issues with unpooled buffers and the recommended way is to use pooled buffers. Have we taken it into account?

[dagnir]

I'm not sure if that guidance applies here. The NIO ByteBuffer that we wrap gets allocated and deallocated either way since it's provided the the Publisher. In the current code, we allocate a ByteBuf only to copy the NIO ByteBuffer's contents into it, and we just avoid that step in this PR.

[zoewangg]

Oh I see what you mean. Yeah, should be good then. 👍