notifyAll() // Because doReadTimeout() may have changed.
      }
    }

    // Only DATA frames are subject to flow-control. Transmit the HEADER frame if the connection
    // flow-control window is fully depleted.
    if (!flushHeaders) {
      withLock {
        flushHeaders = (connection.writeBytesTotal >= connection.writeBytesMaximum)
      }
    }

    connection.writeHeaders(id, outFinished, responseHeaders)

            switch (phase) {
                case "clean":
                    packaged = false;
                    break;
                case "package":
                case "install":
                case "deploy":
                    packaged = true;
                    break;
                default:
                    break;
            }
        }
        return packaged;
    }

    // correct.
    this.state.markReady();
  }

  /**
   * Registers a {@link Listener} to be {@linkplain Executor#execute executed} on the given
   * executor. The listener will not have previous state changes replayed, so it is suggested that
   * listeners are added before any of the managed services are {@linkplain Service#startAsync
   * started}.
   *

  // * We are more responsive to completion than timeouts. This is because parkNanos depends on
  //   system scheduling and as such we could either miss our deadline, or unpark() could be delayed
  //   so that it looks like we timed out even though we didn't. For comparison FutureTask respects
  //   completion preferably and AQS is non-deterministic (depends on where in the queue the waiter

     * returns 0, but if we had a SingletonImmutableSortedSet, this might no
     * longer be the case.)
     *
     * javax.naming.Name and java.util.concurrent.Delayed might work, but
     * they're fairly obscure, we've invented our own interface and class.
     */
    Interface a = new Impl();
    Interface b = new Impl();

```

### 8.2 Network Partition Handling
```java
public class WitnessNetworkPartitionHandler {
    public void handleNetworkPartition() {
        log.warn("Network partition detected - witness notifications may be delayed");
        
        // Switch to more aggressive connection retry
        // Increase heartbeat frequency
        // Consider fallback mechanisms
    }
    
    public void handlePartitionRecovery() {

Но теперь, когда Uvicorn (и команда `fastapi`) поддерживают `--workers`, нет причин использовать базовый Docker-образ вместо сборки своего (кода получается примерно столько же 😅).

///

## Развёртывание образа контейнера { #deploy-the-container-image }

После того как у вас есть образ контейнера (Docker), его можно развёртывать несколькими способами.

Например:

* С **Docker Compose** на одном сервере
* В кластере **Kubernetes**

     * returns 0, but if we had a SingletonImmutableSortedSet, this might no
     * longer be the case.)
     *
     * javax.naming.Name and java.util.concurrent.Delayed might work, but
     * they're fairly obscure, we've invented our own interface and class.
     */
    Interface a = new Impl();
    Interface b = new Impl();

但现在 Uvicorn（以及 `fastapi` 命令）已经支持使用 `--workers`，因此没有理由不自己构建基础镜像（代码量几乎一样 😅）。

///

## 部署容器镜像 { #deploy-the-container-image }

得到容器（Docker）镜像后，有多种方式可以部署。

例如：

* 在单台服务器上使用 **Docker Compose**
* 使用 **Kubernetes** 集群
* 使用 Docker Swarm Mode 集群
* 使用 Nomad 等其他工具
* 使用云服务，接收你的容器镜像并部署

- Kubeadm: add a upgrade health check that deploys a Job ([#81319](https://github.com/kubernetes/kubernetes/pull/81319), [@neolit123](https://github.com/neolit123))