导购APP高可用数据库设计：主从分离与分库分表在返利系统中的应用

大家好，我是省赚客APP研发者阿可！省赚客APP（juwatech.cn）作为日活百万级的导购返利平台，核心数据如用户、订单、佣金记录等面临高写入、高查询压力。为保障系统高可用与低延迟，我们在MySQL层面实施了主从读写分离 + ShardingSphere分库分表的混合架构。本文将结合实际代码，详解关键设计与实现。

读写分离：基于ShardingSphere-JDBC动态路由

我们使用ShardingSphere-JDBC实现透明读写分离，写操作路由至主库，读操作负载均衡至多个从库：

yaml 复制代码

# application.yml
spring:
  shardingsphere:
    datasource:
      names: ds-master,ds-slave0,ds-slave1
      ds-master:
        type: com.zaxxer.hikari.HikariDataSource
        driver-class-name: com.mysql.cj.jdbc.Driver
        jdbc-url: jdbc:mysql://master.db.juwatech.cn:3306/juwatech_commission
        username: root
        password: ***
      ds-slave0:
        type: com.zaxxer.hikari.HikariDataSource
        driver-class-name: com.mysql.cj.jdbc.Driver
        jdbc-url: jdbc:mysql://slave0.db.juwatech.cn:3306/juwatech_commission
        username: root
        password: ***
      ds-slave1:
        # 类似配置
    rules:
      readwrite-splitting:
        data-sources:
          rw_ds:
            write-data-source-name: ds-master
            read-data-source-names: ds-slave0, ds-slave1
            load-balancer-name: round-robin
        load-balancers:
          round-robin:
            type: ROUND_ROBIN

业务代码无需感知底层拓扑：

java 复制代码

package juwatech.cn.service;

@Service
public class CommissionService {
    @Autowired
    private CommissionMapper commissionMapper;

    // 自动走主库
    @Transactional
    public void createCommission(Commission record) {
        commissionMapper.insert(record);
    }

    // 自动走从库（除非事务中）
    public List<Commission> getCommissionByUser(Long userId) {
        return commissionMapper.selectByUserId(userId);
    }
}

分库分表：按用户ID哈希拆分佣金表

佣金记录表 commission_record 数据量超亿级，采用 user_id % 4 分4库，每库再按 order_id % 16 分16表：

yaml 复制代码

spring:
  shardingsphere:
    rules:
      sharding:
        tables:
          commission_record:
            actual-data-nodes: ds-master-${0..3}.commission_record_${0..15}
            table-strategy:
              standard:
                sharding-column: order_id
                sharding-algorithm-name: table-inline
            database-strategy:
              standard:
                sharding-column: user_id
                sharding-algorithm-name: db-inline
        sharding-algorithms:
          db-inline:
            type: INLINE
            props:
              algorithm-expression: ds-master-$->{user_id % 4}
          table-inline:
            type: INLINE
            props:
              algorithm-expression: commission_record_$->{order_id % 16}

实体类与Mapper保持原样：

java 复制代码

package juwatech.cn.entity;

@TableName("commission_record")
public class CommissionRecord {
    private Long id;
    private Long userId;   // 分库键
    private String orderId; // 分表键（存储为字符串，但取模时转Long）
    private BigDecimal amount;
    // getter/setter
}

java 复制代码

package juwatech.cn.mapper;

@Mapper
public interface CommissionRecordMapper {
    @Insert("INSERT INTO commission_record (user_id, order_id, amount) VALUES (#{userId}, #{orderId}, #{amount})")
    int insert(CommissionRecord record);

    // 注意：必须带 user_id 和 order_id 才能精准路由
    @Select("SELECT * FROM commission_record WHERE user_id = #{userId} AND order_id = #{orderId}")
    CommissionRecord selectByUserAndOrder(@Param("userId") Long userId, @Param("orderId") String orderId);
}

全局唯一ID生成：雪花算法避免冲突

分库分表后自增主键失效，我们采用改良版雪花算法生成分布式ID：

java 复制代码

package juwatech.cn.util;

@Component
public class SnowflakeIdGenerator {
    private final long datacenterId; // 从配置获取，0-31
    private final long machineId;    // 0-31
    private long sequence = 0L;
    private long lastTimestamp = -1L;

    public synchronized long nextId() {
        long timestamp = System.currentTimeMillis();
        if (timestamp < lastTimestamp) {
            throw new RuntimeException("Clock moved backwards");
        }
        if (timestamp == lastTimestamp) {
            sequence = (sequence + 1) & 4095L;
            if (sequence == 0) {
                timestamp = waitNextMillis(lastTimestamp);
            }
        } else {
            sequence = 0L;
        }
        lastTimestamp = timestamp;
        return ((timestamp - 1672502400000L) << 22)
                | (datacenterId << 17)
                | (machineId << 12)
                | sequence;
    }

    private long waitNextMillis(long lastTimestamp) {
        long timestamp = System.currentTimeMillis();
        while (timestamp <= lastTimestamp) {
            timestamp = System.currentTimeMillis();
        }
        return timestamp;
    }
}

在插入前赋值：

java 复制代码

CommissionRecord record = new CommissionRecord();
record.setId(snowflakeIdGenerator.nextId());
record.setUserId(userId);
record.setOrderId(orderId);
commissionRecordMapper.insert(record);

跨分片查询优化：异步并行+结果归并

对于管理后台的全局统计需求，我们通过异步任务预聚合，避免实时跨分片扫描：

java 复制代码

@Scheduled(cron = "0 0 1 * * ?")
public void aggregateDailyCommission() {
    List<CompletableFuture<Void>> futures = new ArrayList<>();
    for (int db = 0; db < 4; db++) {
        for (int tbl = 0; tbl < 16; tbl++) {
            final int finalDb = db;
            final int finalTbl = tbl;
            CompletableFuture<Void> future = CompletableFuture.runAsync(() -> {
                String sql = "INSERT INTO daily_commission_summary (date, total_amount) " +
                    "SELECT '2025-04-05', SUM(amount) FROM ds-master-" + finalDb + ".commission_record_" + finalTbl +
                    " WHERE create_time >= '2025-04-05 00:00:00' AND create_time < '2025-04-06 00:00:00'";
                jdbcTemplate.update(sql);
            }, executor);
            futures.add(future);
        }
    }
    CompletableFuture.allOf(futures.toArray(new CompletableFuture[0])).join();
}

数据一致性保障：Binlog监听补偿

为确保主从延迟不导致读取脏数据，关键路径（如提现前查询）强制走主库：

java 复制代码

public class MasterRouteHintManager {
    public static void forceMaster() {
        HintManager.getInstance().setWriteRouteOnly(true);
    }
}

// 使用
public BigDecimal getAvailableCommission(Long userId) {
    MasterRouteHintManager.forceMaster();
    return commissionMapper.selectAvailableByUser(userId);
}

同时，通过 Canal 监听 Binlog，将佣金变动同步至 Elasticsearch 供复杂查询使用，形成多副本冗余。

本文著作权归聚娃科技省赚客app开发者团队，转载请注明出处！