///

## Multiple Workers

You can start multiple workers with the `--workers` command line option:

//// tab | `fastapi`

If you use the `fastapi` command:

<div class="termy">

```console
$ <pre> <font color="#4E9A06">fastapi</font> run --workers 4 <u style="text-decoration-style:single">main.py</u>

		workers = workers[:len(workers)-1]
		worker <- expiryOp(nil)
		es.stats.workers.Add(-1)
	}
	// Atomically replace workers.
	es.workers.Store(&workers)
}

// Worker handles 4 types of expiration tasks.
// 1. Expiry of objects, includes regular and transitioned objects
// 2. Expiry of noncurrent versions due to NewerNoncurrentVersions

  // TODO(b/136478427): Figure out how to correctly shut the server down.
  worker_server.release();

  // Update the server def with a new set of names (worker instead of
  // localhost).
  tensorflow::ServerDef updated_server_def = GetServerDef("worker", 2);
  serialized = updated_server_def.SerializeAsString();

  updated_server_def.set_task_index(1);
  absl::Status s = tensorflow::GrpcServer::Create(

/// tip

Don't worry if some of these items about **containers**, Docker, or Kubernetes don't make a lot of sense yet.

I'll tell you more about container images, Docker, Kubernetes, etc. in a future chapter: [FastAPI in Containers - Docker](docker.md){.internal-link target=_blank}.

///

## Previous Steps Before Starting

/// note | Technical Details

The Docker image was created when Uvicorn didn't support managing and restarting dead workers, so it was needed to use Gunicorn with Uvicorn, which added quite some complexity, just to have Gunicorn manage and restart the Uvicorn worker processes.

  };

  tensorflow::mutex execution_mutex_;
  ExecutionState execution_state_ TF_GUARDED_BY(execution_mutex_) =
      ExecutionState::kIdle;
  // Tells the worker thread that there is new work.
  tensorflow::condition_variable start_execute_;
  // The worker thread notifies that work has finished.
  tensorflow::condition_variable finished_execute_;
  // Notifies a StartExecute that the previous Join has finished.

      int64_t init_timeout_in_ms, int retries,
      bool clear_existing_contexts = false) = 0;

  // Initializes context for the local worker and no contexts will be created
  // for remote workers. Currently this only works for resetting context.
  // TODO(b/289445025): Consider removing this when we find a proper fix.
  virtual absl::Status InitializeLocalOnlyContext(const ServerDef& server_def,

		// 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 := h.workers.Load()
			if nWorkers < h.maxWorkers {
				if time.Since(h.lastStarted).Milliseconds() > 10 {

    )

    failureConditions {
        // JavaExecDebugIntegrationTest.debug session fails without debugger might cause JVM crash
        // Some soak tests produce OOM exceptions
        // There are also random worker crashes for some tests.
        // We have test-retry to handle the crash in tests
        javaCrash = false
    }
})

private fun determineFlakyTestStrategy(stage: Stage): String {

	if (checkOld > 0 && len(p.workers) != checkOld) || n == len(p.workers) || n < 1 {
		// Either already satisfied or worker count changed while we waited for the lock.
		return
	}
	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]