Java multi-threaded batch operations, some people do not do transaction control?

foreword

There is a requirement in the company's business that up to about 50,000 pieces of data need to be modified at the same time, and batch or asynchronous modification operations are not supported. So I can only write a for loop operation, but the operation takes too long, and I can only find other solutions step by step.

The specific operations are as follows:

First, the code of the loop operation

Write a simple for loop code first, and see how the time-consuming situation is.

/***
 * 一条一条依次对50000条数据进行更新操作
 * 耗时：2m27s,1m54s
 */
@Test
void updateStudent() {
    List<Student> allStudents = studentMapper.getAll();
    allStudents.forEach(s -> {
        //更新教师信息
        String teacher = s.getTeacher();
        String newTeacher = "TNO_" + new Random().nextInt(100);
        s.setTeacher(newTeacher);
        studentMapper.update(s);
    });
}
复制代码

The whole cycle modification takes about 1 minute and 54 seconds, and there is no manual transaction control in the code. It should be automatic transaction submission, so each operation transaction will be submitted, so the operation is relatively slow. We first add manual transaction control to the code to see how efficient the query is. .

2. Operation code using manual transactions

The modified code is as follows:

@Autowired
private DataSourceTransactionManager dataSourceTransactionManager;

@Autowired
private TransactionDefinition transactionDefinition;

/**
 * 由于希望更新操作 一次性完成，需要手动控制添加事务
 * 耗时：24s
 * 从测试结果可以看出，添加事务后插入数据的效率有明显的提升
 */
@Test
void updateStudentWithTrans() {
    List<Student> allStudents = studentMapper.getAll();
    TransactionStatus transactionStatus = dataSourceTransactionManager.getTransaction(transactionDefinition);
    try {
        allStudents.forEach(s -> {
            //更新教师信息
            String teacher = s.getTeacher();
            String newTeacher = "TNO_" + new Random().nextInt(100);
            s.setTeacher(newTeacher);
            studentMapper.update(s);
        });
        dataSourceTransactionManager.commit(transactionStatus);
    } catch (Throwable e) {
        dataSourceTransactionManager.rollback(transactionStatus);
        throw e;
    }
}
复制代码

After adding manual transaction control, the overall time is about 24 seconds, which is about 5 times faster than the code of automatic transaction submission. For a large number of cyclic database submission operations, adding manual transactions can effectively improve the operation efficiency.

3. Try multi-threading for data modification

After adding the database manual transaction, the operation efficiency has been improved in detail, but it is still relatively long. Next, try multi-threaded submission to see if it can be faster.

First add a Service to integrate the batch modification operations. The specific code is as follows:

StudentServiceImpl.java

@Service
public class StudentServiceImpl implements StudentService {
    @Autowired
    private StudentMapper studentMapper;
 
    @Autowired
    private DataSourceTransactionManager dataSourceTransactionManager;
 
    @Autowired
    private TransactionDefinition transactionDefinition;
 
    @Override
    public void updateStudents(List<Student> students, CountDownLatch threadLatch) {
        TransactionStatus transactionStatus = dataSourceTransactionManager.getTransaction(transactionDefinition);
        System.out.println("子线程：" + Thread.currentThread().getName());
        try {
            students.forEach(s -> {
                // 更新教师信息
                // String teacher = s.getTeacher();
                String newTeacher = "TNO_" + new Random().nextInt(100);
                s.setTeacher(newTeacher);
                studentMapper.update(s);
            });
            dataSourceTransactionManager.commit(transactionStatus);
            threadLatch.countDown();
        } catch (Throwable e) {
            e.printStackTrace();
            dataSourceTransactionManager.rollback(transactionStatus);
        }
    }
}
复制代码

Batch test code, we use multi-thread to submit, the modified test code is as follows:

@Autowired
private DataSourceTransactionManager dataSourceTransactionManager;

@Autowired
private TransactionDefinition transactionDefinition;

@Autowired
private StudentService studentService;

/**
 * 对用户而言，27s 任是一个较长的时间，我们尝试用多线程的方式来经行修改操作看能否加快处理速度
 * 预计创建10个线程，每个线程进行5000条数据修改操作
 * 耗时统计
 * 1 线程数：1      耗时：25s
 * 2 线程数：2      耗时：14s
 * 3 线程数：5      耗时：15s
 * 4 线程数：10     耗时：15s
 * 5 线程数：100    耗时：15s
 * 6 线程数：200    耗时：15s
 * 7 线程数：500    耗时：17s
 * 8 线程数：1000    耗时：19s
 * 8 线程数：2000    耗时：23s
 * 8 线程数：5000    耗时：29s
 */
@Test
void updateStudentWithThreads() {
    //查询总数据
    List<Student> allStudents = studentMapper.getAll();
    // 线程数量
    final Integer threadCount = 100;

    //每个线程处理的数据量
    final Integer dataPartionLength = (allStudents.size() + threadCount - 1) / threadCount;

    // 创建多线程处理任务
    ExecutorService studentThreadPool = Executors.newFixedThreadPool(threadCount);
    CountDownLatch threadLatchs = new CountDownLatch(threadCount);

    for (int i = 0; i < threadCount; i++) {
        // 每个线程处理的数据
        List<Student> threadDatas = allStudents.stream()
                .skip(i * dataPartionLength).limit(dataPartionLength).collect(Collectors.toList());
        studentThreadPool.execute(() -> {
            studentService.updateStudents(threadDatas, threadLatchs);
        });
    }
    try {
        // 倒计时锁设置超时时间 30s
        threadLatchs.await(30, TimeUnit.SECONDS);
    } catch (Throwable e) {
        e.printStackTrace();
    }

    System.out.println("主线程完成");
}
复制代码

When submitting modifications in multiple threads, we tried the effect of different number of threads on the submission speed. For details, please refer to the following table.

Time-consuming comparison of different number of threads when modifying 50,000 pieces of data in multiple threads (seconds)

According to the table, when we increase the number of threads, the submission speed does not always increase. In the current situation, when the number of threads is about 2-5, the submission speed is the fastest (the actual number of threads still needs to be tested according to the server configuration).

Fourth, control multi-threaded transaction submission based on two CountDownLatch

Due to the multi-threaded submission, each thread transaction is separate and cannot guarantee consistency. We try to add transaction control to multi-threading to ensure that each thread commits the transaction after inserting data.

Here we use two CountDownLatch to control the main thread and sub-thread transaction commit, and set the timeout to 30 seconds. We modified the code a little bit:

@Override
public void updateStudentsThread(List<Student> students, CountDownLatch threadLatch, CountDownLatch mainLatch, StudentTaskError taskStatus) {
    TransactionStatus transactionStatus = dataSourceTransactionManager.getTransaction(transactionDefinition);
    System.out.println("子线程：" + Thread.currentThread().getName());
    try {
        students.forEach(s -> {
            // 更新教师信息
            // String teacher = s.getTeacher();
            String newTeacher = "TNO_" + new Random().nextInt(100);
            s.setTeacher(newTeacher);
            studentMapper.update(s);
        });
    } catch (Throwable e) {
        taskStatus.setIsError();
    } finally {
        threadLatch.countDown(); // 切换到主线程执行
    }
    try {
        mainLatch.await();  //等待主线程执行
    } catch (Throwable e) {
        taskStatus.setIsError();
    }
    // 判断是否有错误，如有错误 就回滚事务
    if (taskStatus.getIsError()) {
        dataSourceTransactionManager.rollback(transactionStatus);
    } else {
        dataSourceTransactionManager.commit(transactionStatus);
    }
}
复制代码

/**
 * 由于每个线程都是单独的事务，需要添加对线程事务的统一控制
 * 我们这边使用两个 CountDownLatch 对子线程的事务进行控制
 */
@Test
void updateStudentWithThreadsAndTrans() {
    //查询总数据
    List<Student> allStudents = studentMapper.getAll();
    // 线程数量
    final Integer threadCount = 4;

    //每个线程处理的数据量
    final Integer dataPartionLength = (allStudents.size() + threadCount - 1) / threadCount;

    // 创建多线程处理任务
    ExecutorService studentThreadPool = Executors.newFixedThreadPool(threadCount);
    CountDownLatch threadLatchs = new CountDownLatch(threadCount); // 用于计算子线程提交数量
    CountDownLatch mainLatch = new CountDownLatch(1); // 用于判断主线程是否提交
    StudentTaskError taskStatus = new StudentTaskError(); // 用于判断子线程任务是否有错误

    for (int i = 0; i < threadCount; i++) {
        // 每个线程处理的数据
        List<Student> threadDatas = allStudents.stream()
                .skip(i * dataPartionLength).limit(dataPartionLength)
                .collect(Collectors.toList());
        studentThreadPool.execute(() -> {
            studentService.updateStudentsThread(threadDatas, threadLatchs, mainLatch, taskStatus);
        });
    }
    try {
        // 倒计时锁设置超时时间 30s
        boolean await = threadLatchs.await(30, TimeUnit.SECONDS);
        if (!await) { // 等待超时，事务回滚
            taskStatus.setIsError();
        }
    } catch (Throwable e) {
        e.printStackTrace();
        taskStatus.setIsError();
    }
    mainLatch.countDown(); // 切换到子线程执行
    studentThreadPool.shutdown(); //关闭线程池

    System.out.println("主线程完成");
}
复制代码

When I wanted to test the effect of different number of threads on execution efficiency again, I found that when the number of threads exceeds 10, an error will be reported during execution. The specific error content is as follows:

Exception in thread "pool-1-thread-2" org.springframework.transaction.CannotCreateTransactionException: Could not open JDBC Connection for transaction; nested exception is java.sql.SQLTransientConnectionException: HikariPool-1 - Connection is not available, request timed out after 30055ms.
 at org.springframework.jdbc.datasource.DataSourceTransactionManager.doBegin(DataSourceTransactionManager.java:309)
 at org.springframework.transaction.support.AbstractPlatformTransactionManager.startTransaction(AbstractPlatformTransactionManager.java:400)
 at org.springframework.transaction.support.AbstractPlatformTransactionManager.getTransaction(AbstractPlatformTransactionManager.java:373)
 at com.example.springbootmybatis.service.Impl.StudentServiceImpl.updateStudentsThread(StudentServiceImpl.java:58)
 at com.example.springbootmybatis.StudentTest.lambda$updateStudentWithThreadsAndTrans$3(StudentTest.java:164)
 at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
 at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
 at java.lang.Thread.run(Thread.java:748)
Caused by: java.sql.SQLTransientConnectionException: HikariPool-1 - Connection is not available, request timed out after 30055ms.
 at com.zaxxer.hikari.pool.HikariPool.createTimeoutException(HikariPool.java:696)
 at com.zaxxer.hikari.pool.HikariPool.getConnection(HikariPool.java:197)
 at com.zaxxer.hikari.pool.HikariPool.getConnection(HikariPool.java:162)
 at com.zaxxer.hikari.HikariDataSource.getConnection(HikariDataSource.java:128)
 at org.springframework.jdbc.datasource.DataSourceTransactionManager.doBegin(DataSourceTransactionManager.java:265)
 ... 7 more
复制代码

错误的大致意思时，不能为数据库事务打开 jdbc Connection，连接在30s的时候超时了。由于前面启动的十个线程需要等待主线程完成后才能提交，所以一直占用连接未释放，造成后面的进程创建连接超时。

看错误日志中错误的来源是 HikariPool ，我们来重新配置一下这个连接池的参数，将最大连接数修改为100，具体配置如下：

# 连接池中允许的最小连接数。缺省值：10
spring.datasource.hikari.minimum-idle=10
# 连接池中允许的最大连接数。缺省值：10
spring.datasource.hikari.maximum-pool-size=100
# 自动提交
spring.datasource.hikari.auto-commit=true
# 一个连接idle状态的最大时长（毫秒），超时则被释放（retired），缺省:10分钟
spring.datasource.hikari.idle-timeout=30000
# 一个连接的生命时长（毫秒），超时而且没被使用则被释放（retired），缺省:30分钟，建议设置比数据库超时时长少30秒
spring.datasource.hikari.max-lifetime=1800000
# 等待连接池分配连接的最大时长（毫秒），超过这个时长还没可用的连接则发生SQLException， 缺省:30秒
复制代码

再次执行测试发现没有报错，修改线程数为20又执行了一下，同样执行成功了。

五、基于TransactionStatus集合来控制多线程事务提交

在同事推荐下我们使用事务集合来进行多线程事务控制，主要代码如下

@Service
public class StudentsTransactionThread {
 
    @Autowired
    private StudentMapper studentMapper;
    @Autowired
    private StudentService studentService;
    @Autowired
    private PlatformTransactionManager transactionManager;
 
    List<TransactionStatus> transactionStatuses = Collections.synchronizedList(new ArrayList<TransactionStatus>());
 
    @Transactional(propagation = Propagation.REQUIRED, rollbackFor = {Exception.class})
    public void updateStudentWithThreadsAndTrans() throws InterruptedException {
 
        //查询总数据
        List<Student> allStudents = studentMapper.getAll();
 
        // 线程数量
        final Integer threadCount = 2;
 
        //每个线程处理的数据量
        final Integer dataPartionLength = (allStudents.size() + threadCount - 1) / threadCount;
 
        // 创建多线程处理任务
        ExecutorService studentThreadPool = Executors.newFixedThreadPool(threadCount);
        CountDownLatch threadLatchs = new CountDownLatch(threadCount);
        AtomicBoolean isError = new AtomicBoolean(false);
        try {
            for (int i = 0; i < threadCount; i++) {
                // 每个线程处理的数据
                List<Student> threadDatas = allStudents.stream()
                        .skip(i * dataPartionLength).limit(dataPartionLength).collect(Collectors.toList());
                studentThreadPool.execute(() -> {
                    try {
                        try {
                            studentService.updateStudentsTransaction(transactionManager, transactionStatuses, threadDatas);
                        } catch (Throwable e) {
                            e.printStackTrace();
                            isError.set(true);
                        }finally {
                            threadLatchs.countDown();
                        }
                    } catch (Exception e) {
                        e.printStackTrace();
                        isError.set(true);
                    }
                });
            }
 
            // 倒计时锁设置超时时间 30s
            boolean await = threadLatchs.await(30, TimeUnit.SECONDS);
            // 判断是否超时
            if (!await) {
                isError.set(true);
            }
        } catch (Throwable e) {
            e.printStackTrace();
            isError.set(true);
        }
 
        if (!transactionStatuses.isEmpty()) {
            if (isError.get()) {
                transactionStatuses.forEach(s -> transactionManager.rollback(s));
            } else {
                transactionStatuses.forEach(s -> transactionManager.commit(s));
            }
        }
 
        System.out.println("主线程完成");
    }
}
复制代码

@Override
@Transactional(propagation = Propagation.REQUIRED, rollbackFor = {Exception.class})
public void updateStudentsTransaction(PlatformTransactionManager transactionManager, List<TransactionStatus> transactionStatuses, List<Student> students) {
    // 使用这种方式将事务状态都放在同一个事务里面
    DefaultTransactionDefinition def = new DefaultTransactionDefinition();
    def.setPropagationBehavior(TransactionDefinition.PROPAGATION_REQUIRES_NEW); // 事物隔离级别，开启新事务，这样会比较安全些。
    TransactionStatus status = transactionManager.getTransaction(def); // 获得事务状态
    transactionStatuses.add(status);

    students.forEach(s -> {
        // 更新教师信息
        // String teacher = s.getTeacher();
        String newTeacher = "TNO_" + new Random().nextInt(100);
        s.setTeacher(newTeacher);
        studentMapper.update(s);
    });
    System.out.println("子线程：" + Thread.currentThread().getName());
}
复制代码

由于这个中方式去前面方式相同，需要等待线程执行完成后才会提交事务，所有任会占用Jdbc连接池，如果线程数量超过连接池最大数量会产生连接超时。所以在使用过程中任要控制线程数量，

六、使用union连接多个select实现批量update

有些情况写不支持，批量update，但支持insert 多条数据，这个时候可尝试将需要更新的数据拼接成多条select 语句，然后使用union 连接起来，再使用update 关联这个数据进行update，具体代码演示如下：

update student,(
 (select  1 as id,'teacher_A' as teacher) union
 (select  2 as id,'teacher_A' as teacher) union
 (select  3 as id,'teacher_A' as teacher) union
 (select  4 as id,'teacher_A' as teacher)
    /* ....more data ... */
    ) as new_teacher
set
 student.teacher=new_teacher.teacher
where
 student.id=new_teacher.id
复制代码

这种方式在Mysql 数据库没有配置 allowMultiQueries=true 也可以实现批量更新。

总结

• 对于大批量数据库操作，使用手动事务提交可以很多程度上提高操作效率
• 多线程对数据库进行操作时，并非线程数越多操作时间越快，按上述示例大约在2-5个线程时操作时间最快。
• 对于多线程阻塞事务提交时，线程数量不能过多。
• 如果能有办法实现批量更新那是最好