gRPC(Java) keepAlive机制研究

时间:2022-11-18 13:06:06

基于java gRPC 1.24.2 分析

结论

  1. gRPC keepAlive是grpc框架在应用层面连接保活的一种措施。即当grpc连接上没有业务数据时,是否发送pingpong,以保持连接活跃性,不因长时间空闲而被Server或操作系统关闭
  2. gRPC keepAlive在client与server都有,client端默认关闭(keepAliveTime为Long.MAX_VALUE), server端默认打开,keepAliveTime为2小时,即每2小时向client发送一次ping
// io.grpc.internal.GrpcUtil
public static final long DEFAULT_SERVER_KEEPALIVE_TIME_NANOS = TimeUnit.HOURS.toNanos(2L);
  1. KeepAlive的管理使用类io.grpc.internal.KeepAliveManager, 用于管理KeepAlive状态,ping任务调度与执行.

Client端KeepAlive

使用入口

  1. 我们在使用io.grpc框架创建grpc连接的时候,可以设置keeplive, 例如下面:
NettyChannelBuilder builder = NettyChannelBuilder.forTarget(String.format("grpc://%s", provider)) //
      .usePlaintext() //
      .defaultLoadBalancingPolicy(props.getBalancePolicy()) //
      .maxInboundMessageSize(props.getMaxInboundMessageSize()) //
      .keepAliveTime(1,TimeUnit.MINUTES)
      .keepAliveWithoutCalls(true)
      .keepAliveTimeout(10,TimeUnit.SECONDS)
      .intercept(channelManager.getInterceptors()); //
  1. 其中与keepAlive相关的参数有三个,keepAliveTime,keepAliveTimeout,keepAliveWithoutCalls。这三个变量有什么作用呢?
  • keepAliveTime: 表示当grpc连接没有数据传递时,多久之后开始向server发送ping packet
  • keepAliveTimeout: 表示当发送完ping packet后多久没收到server回应算超时
  • keepAliveTimeoutCalls: 表示如果grpc连接没有数据传递时,是否keepAlive,默认为false

简要时序列表

Create & Start

NettyChannelBuilder
   -----> NettyTransportFactory
   ---------> NettyClientTransport
   -------------> KeepAliveManager & NettyClientHandler

响应各种事件
当Active、Idle、DataReceived、Started、Termination事件发生时,更改KeepAlive状态,调度发送ping任务。

Server端KeepAlive

使用入口

// 只截取关键代码,详细代码请看`NettyServerBuilder`
ServerImpl server = new ServerImpl(
    this,
    buildTransportServers(getTracerFactories()),
    Context.ROOT);
for (InternalNotifyOnServerBuild notifyTarget : notifyOnBuildList) {
  notifyTarget.notifyOnBuild(server);
}
return server;

// 在buildTransportServers方法中创建NettyServer
List<NettyServer> transportServers = new ArrayList<>(listenAddresses.size());
for (SocketAddress listenAddress : listenAddresses) {
  NettyServer transportServer = new NettyServer(
      listenAddress, resolvedChannelType, channelOptions, bossEventLoopGroupPool,
      workerEventLoopGroupPool, negotiator, streamTracerFactories,
      getTransportTracerFactory(), maxConcurrentCallsPerConnection, flowControlWindow,
      maxMessageSize, maxHeaderListSize, keepAliveTimeInNanos, keepAliveTimeoutInNanos,
      maxConnectionIdleInNanos, maxConnectionAgeInNanos, maxConnectionAgeGraceInNanos,
      permitKeepAliveWithoutCalls, permitKeepAliveTimeInNanos, getChannelz());
  transportServers.add(transportServer);
}

简要时序列表

Create & Start

NettyServerBuilder
    ---> NettyServer
    ---------> NettyServerTransport
    -------------> NettyServerHandler
    -----------------> KeepAliveEnforcer

连接准备就绪
调用 io.netty.channel.ChannelHandler的handlerAdded方法,关于此方法的描述:

Gets called after the ChannelHandler was added to the actual context and it's ready to handle events.
NettyServerHandler(handlerAdded)
   ---> 创建KeepAliveManager对象

响应各种事件
同Client

KeepAliveEnforcer

在上面Server端的简要时序图中,可以看见,server端有一个特有的io.grpc.netty.KeepAliveEnforcer
此类的作用是监控clinet ping的频率,以确保其在一个合理范围内。

package io.grpc.netty;

import com.google.common.annotations.VisibleForTesting;
import com.google.common.base.Preconditions;
import java.util.concurrent.TimeUnit;
import javax.annotation.CheckReturnValue;

/** Monitors the client's PING usage to make sure the rate is permitted. */
class KeepAliveEnforcer {
  @VisibleForTesting
  static final int MAX_PING_STRIKES = 2;
  @VisibleForTesting
  static final long IMPLICIT_PERMIT_TIME_NANOS = TimeUnit.HOURS.toNanos(2);

  private final boolean permitWithoutCalls;
  private final long minTimeNanos;
  private final Ticker ticker;
  private final long epoch;

  private long lastValidPingTime;
  private boolean hasOutstandingCalls;
  private int pingStrikes;

  public KeepAliveEnforcer(boolean permitWithoutCalls, long minTime, TimeUnit unit) {
    this(permitWithoutCalls, minTime, unit, SystemTicker.INSTANCE);
  }

  @VisibleForTesting
  KeepAliveEnforcer(boolean permitWithoutCalls, long minTime, TimeUnit unit, Ticker ticker) {
    Preconditions.checkArgument(minTime >= 0, "minTime must be non-negative");

    this.permitWithoutCalls = permitWithoutCalls;
    this.minTimeNanos = Math.min(unit.toNanos(minTime), IMPLICIT_PERMIT_TIME_NANOS);
    this.ticker = ticker;
    this.epoch = ticker.nanoTime();
    lastValidPingTime = epoch;
  }

  /** Returns {@code false} when client is misbehaving and should be disconnected. */
  @CheckReturnValue
  public boolean pingAcceptable() {
    long now = ticker.nanoTime();
    boolean valid;
    if (!hasOutstandingCalls && !permitWithoutCalls) {
      valid = compareNanos(lastValidPingTime + IMPLICIT_PERMIT_TIME_NANOS, now) <= 0;
    } else {
      valid = compareNanos(lastValidPingTime + minTimeNanos, now) <= 0;
    }
    if (!valid) {
      pingStrikes++;
      return !(pingStrikes > MAX_PING_STRIKES);
    } else {
      lastValidPingTime = now;
      return true;
    }
  }

  /**
   * Reset any counters because PINGs are allowed in response to something sent. Typically called
   * when sending HEADERS and DATA frames.
   */
  public void resetCounters() {
    lastValidPingTime = epoch;
    pingStrikes = 0;
  }

  /** There are outstanding RPCs on the transport. */
  public void onTransportActive() {
    hasOutstandingCalls = true;
  }

  /** There are no outstanding RPCs on the transport. */
  public void onTransportIdle() {
    hasOutstandingCalls = false;
  }

  /**
   * Positive when time1 is greater; negative when time2 is greater; 0 when equal. It is important
   * to use something like this instead of directly comparing nano times. See {@link
   * System#nanoTime}.
   */
  private static long compareNanos(long time1, long time2) {
    // Possibility of overflow/underflow is on purpose and necessary for correctness
    return time1 - time2;
  }

  @VisibleForTesting
  interface Ticker {
    long nanoTime();
  }

  @VisibleForTesting
  static class SystemTicker implements Ticker {
    public static final SystemTicker INSTANCE = new SystemTicker();

    @Override
    public long nanoTime() {
      return System.nanoTime();
    }
  }
}

  1. 先来看pingAcceptable方法,此方法是判断是否接受client ping。
  • lastValidPingTime是上次client valid ping的时间, 连接建立时此时间等于KeepAliveEnforcer对象创建的时间。当client ping有效时,其等于当时ping的时间
  • hasOutstandingCalls其初始值为false,当连接activie时,其值为true,当连接idle时,其值为false。如果grpc调用为阻塞时调用,则调用时连接变为active,调用完成,连接变为idle.
  • permitWithoutCalls其值是创建NettyServer时传入,默认为false.
  • IMPLICIT_PERMIT_TIME_NANOS其值为常量,2h
  • minTimeNanos其值是创建NettyServer时传入,默认为5min.
  • MAX_PING_STRIKES其值为常量2
  1. resetCounters方法是当grpc当中有数据时会被调用,即有grpc调用时lastValidPingTime和pingStrikes会被重置。
  2. 如果client要想使用keepAlive,permitWithoutCalls值需要设置为true,而且cient keepAliveTime需要>=minTimeNanos