Currently, Ray only supports one worker per process. This is okay with Python workers, because:

1. Workers don't share much resources with each other;
2. Multi-threading is inefficient in Python (CPython, to be more specific), because of GIL.

However, as Java worker has been introduced, this becomes an issue.
Because in a Java process, JVM itself costs significant amount of memory overhead. Using multi-processes means duplicating the overhead.

For the short term, I'll work on supporting one worker per thread.
For the long term, we can even use coroutines to make each worker very light-weight.