第5 章：hdfs 客户端HA 连接NameNodeProxiesClient

分享未结  0  847

李延 LV6 2022-03-24

悬赏：20积分

# 1、HA 高可用配置

在连接HA的namenode时，我们需要添加如下配置，这样才可以连接通过集群名称连接。

![](//cdn.hiboot.cn/be8e051db68d404386c71225be4c2620.png)

同时fs.defaultFS也要配置为集群的名称才可以。那么这又是如何被加载的。我们分析一下。

# 2、 HA客户端代理创建

我们在前面分析知道namenode最后是ClientNamenodeProtocolTranslatorPB通过RPC向服务端发送请求的，而在DFSClient 创建时，有一个HA的判断，我们回顾一下

```java
public static ProxyAndInfo<ClientProtocol> createProxyWithClientProtocol(
    Configuration conf, URI nameNodeUri, AtomicBoolean fallbackToSimpleAuth)
    throws IOException {
  //判断是否有ha 代理累
  AbstractNNFailoverProxyProvider<ClientProtocol> failoverProxyProvider =
      createFailoverProxyProvider(conf, nameNodeUri, ClientProtocol.class,
          true, fallbackToSimpleAuth);
	//没有代理直接创建连接
  if (failoverProxyProvider == null) {
    //获取连接地址
    InetSocketAddress nnAddr = DFSUtilClient.getNNAddress(nameNodeUri);
    Text dtService = SecurityUtil.buildTokenService(nnAddr);
    //创建连接
    ClientProtocol proxy = createNonHAProxyWithClientProtocol(nnAddr, conf,
        UserGroupInformation.getCurrentUser(), true, fallbackToSimpleAuth);
    return new ProxyAndInfo<>(proxy, dtService, nnAddr);
  } else {
    return createHAProxy(conf, nameNodeUri, ClientProtocol.class,
        failoverProxyProvider);
  }
}
```

这里我们看到首先它去获取了createFailoverProxyProvider 对象，而createFailoverProxyProvider方法如下：

```java
protected static <T> AbstractNNFailoverProxyProvider<T> createFailoverProxyProvider(
    Configuration conf, URI nameNodeUri, Class<T> xface, boolean checkPort,
    AtomicBoolean fallbackToSimpleAuth, HAProxyFactory<T> proxyFactory)
    throws IOException {
  Class<FailoverProxyProvider<T>> failoverProxyProviderClass = null;
  AbstractNNFailoverProxyProvider<T> providerNN;
  try {
    // Obtain the class of the proxy provider
    //在配置文件中获取到dfs.client.failover.proxy.provider.[nameservice ID] 配置的class
    failoverProxyProviderClass = getFailoverProxyProviderClass(conf,
        nameNodeUri);
    if (failoverProxyProviderClass == null) {
      return null;
    }
    // Create a proxy provider instance.
    //通过反射获取创建实例
    Constructor<FailoverProxyProvider<T>> ctor = failoverProxyProviderClass
        .getConstructor(Configuration.class, URI.class,
            Class.class, HAProxyFactory.class);
    FailoverProxyProvider<T> provider = ctor.newInstance(conf, nameNodeUri,
        xface, proxyFactory);

// If the proxy provider is of an old implementation, wrap it.
    if (!(provider instanceof AbstractNNFailoverProxyProvider)) {
      providerNN = new WrappedFailoverProxyProvider<>(provider);
    } else {
      providerNN = (AbstractNNFailoverProxyProvider<T>)provider;
    }
  } catch (Exception e) {
    final String message = "Couldn't create proxy provider " +
        failoverProxyProviderClass;
    LOG.debug(message, e);
    if (e.getCause() instanceof IOException) {
      throw (IOException) e.getCause();
    } else {
      throw new IOException(message, e);
    }
  }

// Check the port in the URI, if it is logical.
  if (checkPort && providerNN.useLogicalURI()) {
    int port = nameNodeUri.getPort();
    if (port > 0 &&
        port != HdfsClientConfigKeys.DFS_NAMENODE_RPC_PORT_DEFAULT) {
      // Throwing here without any cleanup is fine since we have not
      // actually created the underlying proxies yet.
      throw new IOException("Port " + port + " specified in URI "
          + nameNodeUri + " but host '" + nameNodeUri.getHost()
          + "' is a logical (HA) namenode"
          + " and does not use port information.");
    }
  }
  providerNN.setFallbackToSimpleAuth(fallbackToSimpleAuth);
  return providerNN;
}
```

这里我们看到是一个通过配置创建对于实例的过程。我们回头再看createHAProxy 方法

```java
public static <T> ProxyAndInfo<T> createHAProxy(
    Configuration conf, URI nameNodeUri, Class<T> xface,
    AbstractNNFailoverProxyProvider<T> failoverProxyProvider) {
  Preconditions.checkNotNull(failoverProxyProvider);
  // HA case
  DfsClientConf config = new DfsClientConf(conf);
  T proxy = (T) RetryProxy.create(xface, failoverProxyProvider,
      RetryPolicies.failoverOnNetworkException(
          RetryPolicies.TRY_ONCE_THEN_FAIL, config.getMaxFailoverAttempts(),
          config.getMaxRetryAttempts(), config.getFailoverSleepBaseMillis(),
          config.getFailoverSleepMaxMillis()));

Text dtService;
  if (failoverProxyProvider.useLogicalURI()) {
    dtService = HAUtilClient.buildTokenServiceForLogicalUri(nameNodeUri,
        HdfsConstants.HDFS_URI_SCHEME);
  } else {
    dtService = SecurityUtil.buildTokenService(
        DFSUtilClient.getNNAddress(nameNodeUri));
  }
  return new ProxyAndInfo<>(proxy, dtService,
      DFSUtilClient.getNNAddressCheckLogical(conf, nameNodeUri));
}

public static <T> Object create(Class<T> iface,
      FailoverProxyProvider<T> proxyProvider, RetryPolicy retryPolicy) {
    return Proxy.newProxyInstance(
        proxyProvider.getInterface().getClassLoader(),
        new Class<?>[] { iface },
        new RetryInvocationHandler<T>(proxyProvider, retryPolicy)
        );
  }
```

这里我们看到我们刚创建的proxyProvider对象，被传给了RetryInvocationHandler，而RetryInvocationHandler 就是我们最后的代理。它们大概关系为：

- RetryInvocationHandler 支持重试的客户端，当请求异常时，重新生成连接进行请求。
- ConfiguredFailoverProxyProvider 我们配置的代理，HA 代理。为RetryInvocationHandler 不断生成新的RPC连接。

# 3、ConfiguredFailoverProxyProvider

ConfiguredFailoverProxyProvider 父类为： AbstractNNFailoverProxyProvider。 前面我们分析到ConfiguredFailoverProxyProvider 主要功能是不断生成新的RPC代理。

## 3.1 构造函数

```java
public ConfiguredFailoverProxyProvider(Configuration conf, URI uri,
    Class<T> xface, HAProxyFactory<T> factory) {
  this(conf, uri, xface, factory, DFS_NAMENODE_RPC_ADDRESS_KEY);
}

public ConfiguredFailoverProxyProvider(Configuration conf, URI uri,
    Class<T> xface, HAProxyFactory<T> factory, String addressKey) {
  super(conf, uri, xface, factory);
  this.proxies = getProxyAddresses(uri, addressKey);
}
```

这里getProxyAddresses 方法就是加载配置文件中所有namnode的节点

```java
protected List<NNProxyInfo<T>> getProxyAddresses(URI uri, String addressKey) {
  final List<NNProxyInfo<T>> proxies = new ArrayList<NNProxyInfo<T>>();
  Map<String, Map<String, InetSocketAddress>> map =
      DFSUtilClient.getAddresses(conf, null, addressKey);
  //更加uri的host加载对应的名称
  Map<String, InetSocketAddress> addressesInNN = map.get(uri.getHost());

if (addressesInNN == null || addressesInNN.size() == 0) {
    throw new RuntimeException("Could not find any configured addresses " +
        "for URI " + uri);
  }
	// 将加载的名称转换为list
  Collection<InetSocketAddress> addressesOfNns = addressesInNN.values();
  try {
    addressesOfNns = getResolvedHostsIfNecessary(addressesOfNns, uri);
  } catch (IOException e) {
    throw new RuntimeException(e);
  }
  for (InetSocketAddress address : addressesOfNns) {
    proxies.add(new NNProxyInfo<T>(address));
  }
  // Randomize the list to prevent all clients pointing to the same one
  boolean randomized = getRandomOrder(conf, uri);
  if (randomized) {
    Collections.shuffle(proxies);
  }

// The client may have a delegation token set for the logical
  // URI of the cluster. Clone this token to apply to each of the
  // underlying IPC addresses so that the IPC code can find it.
  HAUtilClient.cloneDelegationTokenForLogicalUri(ugi, uri, addressesOfNns);
  return proxies;
}
```

## 3.2 getProxy

```java
@Override
public synchronized ProxyInfo<T> getProxy() {
  NNProxyInfo<T> current = proxies.get(currentProxyIndex);
  return createProxyIfNeeded(current);
}
```

这里我们看到它通过currentProxyIndex 来确定使用哪个url，而createProxyIfNeeded 在父类中实现，与单机情况一样的过程，如下：

```java
protected NNProxyInfo<T> createProxyIfNeeded(NNProxyInfo<T> pi) {
  if (pi.proxy == null) {
    assert pi.getAddress() != null : "Proxy address is null";
    try {
      pi.proxy = factory.createProxy(conf,
          pi.getAddress(), xface, ugi, false, getFallbackToSimpleAuth());
    } catch (IOException ioe) {
      LOG.error("{} Failed to create RPC proxy to NameNode at {}",
          this.getClass().getSimpleName(), pi.address, ioe);
      throw new RuntimeException(ioe);
    }
  }
  return pi;
}
```

那么我们需要关注的是currentProxyIndex是如何变化的

## 3.3 incrementProxyIndex

在incrementProxyIndex 方法中，我们可以看到currentProxyIndex 的切换，它就是一个轮询的方式。

```
synchronized void incrementProxyIndex() {
  currentProxyIndex = (currentProxyIndex + 1) % proxies.size();
}
```

# 4 RetryInvocationHandler

前面我们分析了ConfiguredFailoverProxyProvider ，这里我们再看对应方法在RetryInvocationHandler时何时被调用的

首先它时jdk的动态代理，在我们使用namenode方法时，invoke方法就会被调用

```java
public Object invoke(Object proxy, Method method, Object[] args)
    throws Throwable {
  final boolean isRpc = isRpcInvocation(proxyDescriptor.getProxy());
  final int callId = isRpc? Client.nextCallId(): RpcConstants.INVALID_CALL_ID;

final Call call = newCall(method, args, isRpc, callId);
  while (true) {
    final CallReturn c = call.invokeOnce();
    final CallReturn.State state = c.getState();
    if (state == CallReturn.State.ASYNC_INVOKED) {
      return null; // return null for async calls
    } else if (c.getState() != CallReturn.State.RETRY) {
      return c.getReturnValue();
    }
  }
}
```

这里我们看到一个循环， call.invokeOnce() ，直到call返回成功。

```java
synchronized CallReturn invokeOnce() {
  try {
    if (retryInfo != null) {
      return processWaitTimeAndRetryInfo();
    }

// The number of times this invocation handler has ever been failed over
    // before this method invocation attempt. Used to prevent concurrent
    // failed method invocations from triggering multiple failover attempts.
    final long failoverCount = retryInvocationHandler.getFailoverCount();
    try {
      //调用RPC
      return invoke();
    } catch (Exception e) {
      if (LOG.isTraceEnabled()) {
        LOG.trace(toString(), e);
      }
      if (Thread.currentThread().isInterrupted()) {
        // If interrupted, do not retry.
        throw e;
      }

retryInfo = retryInvocationHandler.handleException(
          method, callId, retryPolicy, counters, failoverCount, e);
      //切换连接
      return processWaitTimeAndRetryInfo();
    }
  } catch(Throwable t) {
    return new CallReturn(t);
  }
}

```

这里我们看到调用invoke 异常时，就是使用processWaitTimeAndRetryInfo方法

```java
CallReturn processWaitTimeAndRetryInfo() throws InterruptedIOException {
  final Long waitTime = getWaitTime(Time.monotonicNow());
  LOG.trace("#{} processRetryInfo: retryInfo={}, waitTime={}",
      callId, retryInfo, waitTime);
  if (waitTime != null && waitTime > 0) {
    try {
      Thread.sleep(retryInfo.delay);
    } catch (InterruptedException e) {
      Thread.currentThread().interrupt();
      if (LOG.isDebugEnabled()) {
        LOG.debug("Interrupted while waiting to retry", e);
      }
      InterruptedIOException intIOE = new InterruptedIOException(
          "Retry interrupted");
      intIOE.initCause(e);
      throw intIOE;
    }
  }
  processRetryInfo();
  return CallReturn.RETRY;
}

synchronized void processRetryInfo() {
      counters.retries++;
      if (retryInfo.isFailover()) {
        retryInvocationHandler.proxyDescriptor.failover(
            retryInfo.expectedFailoverCount, method, callId);
        counters.failovers++;
      }
      retryInfo = null;
    }
    
    
    synchronized void failover(long expectedFailoverCount, Method method,
                               int callId) {
      // Make sure that concurrent failed invocations only cause a single
      // actual failover.
      if (failoverCount == expectedFailoverCount) {
        fpp.performFailover(proxyInfo.proxy);
        failoverCount++;
      } else {
        LOG.warn("A failover has occurred since the start of call #" + callId
            + " " + proxyInfo.getString(method.getName()));
      }
      proxyInfo = fpp.getProxy();
    }
```

一直到failover方法中，我们终于看到performFailover 切换代理，与获取新的代理方法调用

回帖

消灭零回复