This needs to be investigated, reading the code suggests we might be losing data under certain conditions e.g. when terminating the pipeline. The async processing model here is far too simplistic.
The bundle won't ever know about pending writes and won't block to wait for any such operation. The same way exceptions are thrown into nowhere. Test cases don't capture this as they operate on completed futures only (so exceptions in the callbacks get thrown on the thread of processElement).
Also, this entirely removes backpressure from a stream. When used with a much faster source we will continue to accumulate more and more memory as the number of concurrent pending async operations is not limited.
Spotify's scio contains a JavaAsyncDoFn that illustrates how it can be done.