@jamezp FYI 🙏

I've run locally the jackson2 provider integration tests (w debugging), to verify the number of copies are reduced; it passes and looks good 👍 @jamezp

The failures are not related. This is something I need to get fixed.

@franz1981 Is this something we should do more generally for the org.jboss.resteasy.spi.AsyncMessageBodyWriter?

In general looking at that, writing everything to memory before writing it to the async stream like it could be a potential issue for large objects.

In Quarkus I have implemented a thing called AppendBuffer in the vertx output stream, for Restleasy Reactive, which:

• allow appending encoded chunks in batches
• once the buffer capacity (sum of all chunks) is exhausted, it flush it to http as chunked response, and keep on filling it with the rest of the stream

In this way you both have batchy flush and some predictable memory usage.

This specific Jackson writer here looks to be just in memory, I can tell; which as you said, could be dangerous under heavy load due to memory footprint: I've indeed found this improvement because due to the existing implementation, it was so memory hungry to appear in our profiling data in the performance lab

Awesome, thank you. I'd be tempted to put this also in org.jboss.resteasy.core.messagebody.AsyncBufferedMessageBodyWriter

Good point @jamezp yep, in theory should work similarly -although for a generic use case I would create a stream which append chunks of byte[] instead of reallocating (by copying and enlarging) the same byte[] over and over.
This would be able to skip allocating a copy, if there's a single chunk, and be more friendly if the stream is "long" and made of many small writes.
But to be fair, even this one should work as good as the existing one (saving synchronized to happen).

Perfect. I may copy it over and look at the other chunking separately.

Also, thank for you this. I really want to learn more about performance testing, but it seems like an art form that there either isn't a lot of information on or I'm not able to find it :)

Reach out to me and my team and I can help with it @jamezp no worries :)