//  Worker0:
  //  cluster { job { name: "client" tasks { key: 0 value: "localhost:14526" } }
  //            job { name: "worker" tasks { key: 0 value: "localhost:14522" } }
  //          } job_name: "worker" protocol: "grpc"
  //

  // Create `worker_1` using single host server def `worker_1_server_def`.
  tensorflow::ServerDef worker_1_server_def =
      CreateSingleHostServerDef(cluster_server_def, 1);

      "workloadName": "ratings-v1",
      "workloadType": "deployment",
      "canonicalName": "ratings",
      "canonicalRevision": "v1",
      "network": "network1",
      "node": "ambient-worker2",
      "nativeTunnel": true,
      "status": "Healthy",
      "clusterId": "Kubernetes"
    },
    "/10.244.2.58": {
      "workloadIps": [
        "10.244.2.58"
      ],
      "protocol": "HBONE",

Here are some possible combinations and strategies:

* **Gunicorn** managing **Uvicorn workers**
    * Gunicorn would be the **process manager** listening on the **IP** and **port**, the replication would be by having **multiple Uvicorn worker processes**.
* **Uvicorn** managing **Uvicorn workers**

  int port = tensorflow::testing::PickUnusedPortOrDie();
  // 3. Replace worker1 with a new worker in server_def_0 and server_def_1.
  ReplaceTaskInServerDef(&server_def_0, /*task_index=*/1, "localhost", port);
  ReplaceTaskInServerDef(&server_def_1, /*task_index=*/1, "localhost", port);
  // 4. Start a new task to replace worker1.
  server_def_0.set_task_index(1);
  worker_server1.release();
  ASSERT_TRUE(tensorflow::GrpcServer::Create(

	for len(p.workers) < n {
		input := make(chan ReplicationWorkerOperation, 10000)
		p.workers = append(p.workers, input)

		go p.AddWorker(input, &p.activeWorkers)
	}
	for len(p.workers) > n {
		worker := p.workers[len(p.workers)-1]
		p.workers = p.workers[:len(p.workers)-1]
		xioutil.SafeClose(worker)
	}
}

// ResizeWorkerPriority sets replication failed workers pool size

  server_def.set_task_index(1);
  std::unique_ptr<tensorflow::GrpcServer> worker_server1;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(
                  server_def, tensorflow::Env::Default(), &worker_server1)
                  .ok());
  ASSERT_TRUE(worker_server1->Start().ok());

  server_def.set_task_index(2);
  std::unique_ptr<tensorflow::GrpcServer> worker_server2;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(

So, in this case, you **would not** want to have a process manager like Gunicorn with Uvicorn workers, or Uvicorn using its own Uvicorn workers. You would want to have just a **single Uvicorn process** per container (but probably multiple containers).

	retry:
		// If the channel reaches above half capacity
		// we spawn more workers. The workers spawned
		// from this main worker routine will exit
		// once the channel drops below half capacity
		// and when it's been at least 30 seconds since
		// we launched a new worker.
		if mainWorker && len(h.logCh) > cap(h.logCh)/2 {
			nWorkers := atomic.LoadInt64(&h.workers)
			if nWorkers < h.maxWorkers {

	sync.RWMutex                 // mutex for Cache
	mostRecentStatsMu sync.Mutex // mutex for mostRecentStats

	wlock sync.RWMutex // mutex for active workers

	movingAvgTicker *time.Ticker // Ticker for calculating moving averages
	wTimer          *time.Ticker // ticker for calculating active workers
	qTimer          *time.Ticker // ticker for calculating queue stats
}

func (r *ReplicationStats) trackEWMA() {
	for {
		select {

  /**
   * This counter prevents an ABA issue where a thread may successfully schedule the worker, the
   * worker runs and exhausts the queue, another thread enqueues a task and fails to schedule the
   * worker, and then the first thread's call to delegate.execute() returns. Without this counter,
   * it would observe the QUEUING state and set it to QUEUED, and the worker would never be
   * scheduled again for future submissions.
   */
  @GuardedBy("queue")