数据库路由中间件MyCat - 源代码篇(9)

此文已由做者张镐薪受权网易云社区发布。
html

欢迎访问网易云社区,了解更多网易技术产品运营经验。前端

3. 链接模块

3.5 后端链接

3.5.1 后端链接获取与负载均衡

上一节咱们讲了后端链接的基本创建和响应处理,那么这些后端链接是何时创建的呢? 首先,MyCat配置文件中,DataHost标签中有minIdle这个属性。表明在MyCat初始化时,会在这个DataHost上初始化维护多少个链接(这些链接能够理解为链接池)。每一个前端Client链接会建立Session,而Session会根据命令的不一样而建立不一样的Handler。每一个Handler会从链接池中拿出所须要的链接并使用。在链接池大小不够时,RW线程会异步驱使新建所需的链接补充链接池,可是链接数最大不能超过配置的maxCon。同时,如以前所述,有定时线程检查并回收空闲后端链接。但池中最小不会小于minCon。 咱们能够经过后端链接的工厂方法的调用链来理解: 这里写图片描述 看这个调用链,咱们简述下大概的流程。java

st=>start: MyCat接受客户端链接并为之创建惟一绑定的Session
e=>end: 将请求发送给对应链接,处理完以后归还链接
op1=>operation: MyCat接受客户端的请求,计算路由
op2=>operation: 根据请求和路由建立合适的handler,这里为SingleNodeHandler
op3=>operation: 从PhysicalDBNode中获取后端链接
cond=>condition: 尝试获取链接,链接够用?
op4=>operation: 尝试异步建立新的链接
op5=>operation: 经过DelegateResponseHandler将链接与以前的Handler,这里是SingleNodeHandler绑定
st->op1->op2->op3->condcond(yes)->econd(no)->op4->op5->e

咱们先从Session看起,在MyCat中实现类为NonBlockingSession。在前端链接创建时,会建立绑定惟一的Session: ServerConnectionFactory.java:node

protected FrontendConnection getConnection(NetworkChannel channel) throws IOException {
    SystemConfig sys = MycatServer.getInstance().getConfig().getSystem();
    ServerConnection c = new ServerConnection(channel);
    MycatServer.getInstance().getConfig().setSocketParams(c, true);
    c.setPrivileges(MycatPrivileges.instance());
    c.setQueryHandler(new ServerQueryHandler(c));
    c.setLoadDataInfileHandler(new ServerLoadDataInfileHandler(c));    // c.setPrepareHandler(new ServerPrepareHandler(c));
    c.setTxIsolation(sys.getTxIsolation());    //建立绑定惟一Session
    c.setSession2(new NonBlockingSession(c));    return c;
}

Session主要处理事务,多节点转发协调等,由不一样的ResponseHandler实现; 这里写图片描述 这些ResponseHandler咱们以后会在对应的模块去细细分析。这里先跳过。 查看SingleNodeHanlder的处理方法 SingleNodeHanlder.java:es6

public void execute() throws Exception {    //从这里开始计算处理时间
    startTime=System.currentTimeMillis();
    ServerConnection sc = session.getSource();    this.isRunning = true;    this.packetId = 0;    final BackendConnection conn = session.getTarget(node);    //以前是否获取过Connection而且Connection有效
    if (session.tryExistsCon(conn, node)) {
        _execute(conn);
    } else {        // create new connection
        MycatConfig conf = MycatServer.getInstance().getConfig();        //从config中获取DataNode
        PhysicalDBNode dn = conf.getDataNodes().get(node.getName());        //获取对应的数据库链接
        dn.getConnection(dn.getDatabase(), sc.isAutocommit(), node, this,
                   node);
    }

}

从PhysicalDBNode中获取合适的链接:数据库

public void getConnection(String schema,boolean autoCommit, RouteResultsetNode rrs,
        ResponseHandler handler, Object attachment) throws Exception {
    checkRequest(schema);    //检查数据库链接池是否初始化成功,由于有reload命令
    if (dbPool.isInitSuccess()) {        //根据是否能在读节点上运行获取链接,通常是判断是否为读请求,而且读请求不在事务中
        if (rrs.canRunnINReadDB(autoCommit)) {            //根据负载均衡策略选择合适的后端链接
            dbPool.getRWBanlanceCon(schema,autoCommit, handler, attachment,                    this.database);
        } else {            //直接选择当前链接池中的的后端链接
            dbPool.getSource().getConnection(schema,autoCommit, handler, attachment);
        }

    } else {        throw new IllegalArgumentException("Invalid DataSource:"
                + dbPool.getActivedIndex());
    }
}

PhysicalDBPool类中有负载均衡,切换writeHost,控制write方式等(分别对应balance,writeType等标签)的实现。首先咱们看若是有负载均衡策略(配置了balance)的获取链接的方式:canvas

public void getRWBanlanceCon(String schema, boolean autocommit,
                             ResponseHandler handler, Object attachment, String database) throws Exception {

    PhysicalDatasource theNode = null;
    ArrayList<PhysicalDatasource> okSources = null;    switch (banlance) {        //全部读写节点参与read请求的负载均衡,除了当前活跃的写节点,balance=1
        case BALANCE_ALL_BACK: {            //返回全部写节点和符合条件的读节点,不包括当前的写节点
            okSources = getAllActiveRWSources(true, false, checkSlaveSynStatus());            if (okSources.isEmpty()) {                //若是结果即为空,返回当前写节点
                theNode = this.getSource();
            } else {                //不为空,随机选一个
                theNode = randomSelect(okSources);
            }            break;
        }        //全部读写节点参与read请求的负载均衡,balance=2
        case BALANCE_ALL: {            //返回全部写节点和符合条件的读节点
            okSources = getAllActiveRWSources(true, true, checkSlaveSynStatus());            //随机选一个
            theNode = randomSelect(okSources);            break;
        }        case BALANCE_ALL_READ: {            //返回全部符合条件的读节点
            okSources = getAllActiveRWSources(false, false, checkSlaveSynStatus());            //随机取一个
            theNode = randomSelect(okSources);            break;
        }        //不作负载均衡,balance=0或其余不为以上的值
        case BALANCE_NONE:        default:
            // return default write data source
            theNode = this.getSource();
    }    if (LOGGER.isDebugEnabled()) {
        LOGGER.debug("select read source " + theNode.getName() + " for dataHost:" + this.getHostName());
    }
    theNode.getConnection(schema, autocommit, handler, attachment);
}

其中涉及到的方法:后端

  1. 返回符合条件节点集:安全

private ArrayList<PhysicalDatasource> getAllActiveRWSources(            boolean includeWriteNode, boolean includeCurWriteNode, boolean filterWithSlaveThreshold) {        int curActive = activedIndex;
        ArrayList<PhysicalDatasource> okSources = new ArrayList<PhysicalDatasource>(this.allDs.size());        //判断写节点
        for (int i = 0; i < this.writeSources.length; i++) {
            PhysicalDatasource theSource = writeSources[i];            //判断写节点是不是active,可能reload会置为inactive,可能多个写节点可是只有一个是活跃在用的(writeType=0)
            if (isAlive(theSource)) {                //负载均衡策略是否包含写节点
                if (includeWriteNode) {                    //判断是否包含当前活跃的写入节点
                    if (i == curActive && includeCurWriteNode == false) {                        // not include cur active source
                    } else if (filterWithSlaveThreshold) {                        //若是包含从节点同步延迟限制,检查同步状态
                        if (canSelectAsReadNode(theSource)) {
                            okSources.add(theSource);
                        } else {                            //若是同步状态不对,则不添加这个写节点
                            continue;
                        }

                    } else {
                        okSources.add(theSource);
                    }
                }                //检查theSource对应的读节点
                if (!readSources.isEmpty()) {                    // 检查theSource对应的读节点(从节点)
                    PhysicalDatasource[] allSlaves = this.readSources.get(i);                    if (allSlaves != null) {                        for (PhysicalDatasource slave : allSlaves) {                            if (isAlive(slave)) {                                //若是包含从节点同步延迟限制,检查同步状态
                                if (filterWithSlaveThreshold) {                                    if (canSelectAsReadNode(slave)) {                                        //若是同步状态正确,则把读节点加入
                                        okSources.add(slave);
                                    } else {                                        continue;
                                    }

                                } else {
                                    okSources.add(slave);
                                }
                            }
                        }
                    }
                }

            } else {                // TODO : add by zhuam
                // 若是写节点不OK, 也要保证临时的读服务正常
                if (this.dataHostConfig.isTempReadHostAvailable()) {                    if (!readSources.isEmpty()) {                        // check all slave nodes
                        PhysicalDatasource[] allSlaves = this.readSources.get(i);                        if (allSlaves != null) {                            for (PhysicalDatasource slave : allSlaves) {                                if (isAlive(slave)) {                                    if (filterWithSlaveThreshold) {                                        if (canSelectAsReadNode(slave)) {
                                            okSources.add(slave);
                                        } else {                                            continue;
                                        }

                                    } else {
                                        okSources.add(slave);
                                    }
                                }
                            }
                        }
                    }
                }
            }

        }        return okSources;
    }
  1. 检查是否判断主从延迟:性能优化

private boolean checkSlaveSynStatus() {        return (dataHostConfig.getSlaveThreshold() != -1)
                && (dataHostConfig.getSwitchType() == DataHostConfig.SYN_STATUS_SWITCH_DS);
    }
  1. 随机选择节点:

/**
     * TODO: modify by zhuam
     * <p/>
     * 随机选择,按权重设置随机几率。
     * 在一个截面上碰撞的几率高,但调用量越大分布越均匀,并且按几率使用权重后也比较均匀,有利于动态调整提供者权重。
     *
     * @param okSources
     * @return
     */
    public PhysicalDatasource randomSelect(ArrayList<PhysicalDatasource> okSources) {        if (okSources.isEmpty()) {            return this.getSource();

        } else {            int length = okSources.size();    // 总个数
            int totalWeight = 0;            // 总权重
            boolean sameWeight = true;        // 权重是否都同样
            for (int i = 0; i < length; i++) {                int weight = okSources.get(i).getConfig().getWeight();
                totalWeight += weight;        // 累计总权重
                if (sameWeight && i > 0
                        && weight != okSources.get(i - 1).getConfig().getWeight()) {      // 计算全部权重是否同样
                    sameWeight = false;
                }
            }            if (totalWeight > 0 && !sameWeight) {                // 若是权重不相同且权重大于0则按总权重数随机
                int offset = random.nextInt(totalWeight);                // 并肯定随机值落在哪一个片段上
                for (int i = 0; i < length; i++) {
                    offset -= okSources.get(i).getConfig().getWeight();                    if (offset < 0) {                        return okSources.get(i);
                    }
                }
            }            // 若是权重相同或权重为0则均等随机
            return okSources.get(random.nextInt(length));            //int index = Math.abs(random.nextInt()) % okSources.size();
            //return okSources.get(index);
        }
    }
  1. 根据writeType获取当前writeHost方法:

public PhysicalDatasource getSource() {        switch (writeType) {            //writeType=0,返回当前active的writeHost
            case WRITE_ONLYONE_NODE: {                return writeSources[activedIndex];
            }            //writeType=1,随机发到一个writeHost
            case WRITE_RANDOM_NODE: {                int index = Math.abs(wnrandom.nextInt()) % writeSources.length;
                PhysicalDatasource result = writeSources[index];                if (!this.isAlive(result)) {                    // find all live nodes
                    ArrayList<Integer> alives = new ArrayList<Integer>(writeSources.length - 1);                    for (int i = 0; i < writeSources.length; i++) {                        if (i != index) {                            if (this.isAlive(writeSources[i])) {
                                alives.add(i);
                            }
                        }
                    }                    if (alives.isEmpty()) {
                        result = writeSources[0];
                    } else {                        // random select one
                        index = Math.abs(wnrandom.nextInt()) % alives.size();
                        result = writeSources[alives.get(index)];

                    }
                }                if (LOGGER.isDebugEnabled()) {
                    LOGGER.debug("select write source " + result.getName()
                            + " for dataHost:" + this.getHostName());
                }                return result;
            }            //参数不正确
            default: {                throw new java.lang.IllegalArgumentException("writeType is "
                        + writeType + " ,so can't return one write datasource ");
            }
        }

    }


免费体验云安全(易盾)内容安全、验证码等服务

更多网易技术、产品、运营经验分享请点击




相关文章:
【推荐】 canvas 动画库 CreateJs 之 EaselJS(下篇)
【推荐】 针对低网速的性能优化
【推荐】 nej+regular环境使用es6的低成本方案

相关文章
相关标签/搜索