If JIoEndpoint can't get a worker thread immediately, it'll enter the object lock queue of workers. Once a worker is released the endpoint will do a workers.notify(), which will wake up an arbitrary thread.

In effect it has a FIRO queue and not a FIFO queue.

Given a SLA where each request must be completed <3s (which is the current performance requirement) the deviation of response times is such that forfeits this requirement.

The endpoint must therefore use a fair mechanism to dish out worker threads.