Flink：左加入带有静态列表的流

Question

我想将尝试流加入到被阻止电子邮件的静态列表中，并按 IP 对结果进行分组，以便以后可以计算一组相关统计信息。结果应以每 10 秒后 30 分钟的滑动窗口形式提供。以下是我尝试实现的几种方法之一：

override fun performQuery(): Table {
    val query = "SELECT ip, " +
        "COUNT(CASE WHEN success IS false THEN 1 END) AS fails, " +
        "COUNT(CASE WHEN success IS true THEN 1 END) AS successes, " +
        "COUNT(DISTINCT id) accounts, " +
        "COUNT(CASE WHEN id = 0 THEN 1 END) AS non_existing_accounts, " +
        "COUNT(CASE WHEN blockedEmail IS NOT NULL THEN 1 END) AS blocked_accounts " +
        "FROM Attempts " +
        "LEFT JOIN LATERAL TABLE(blockedEmailsList()) AS T(blockedEmail) ON TRUE " +
        "WHERE Attempts.email <> '' AND Attempts.createdAt < CURRENT_TIMESTAMP " +
        "GROUP BY HOP(Attempts.createdAt, INTERVAL '10' SECOND, INTERVAL '30' MINUTE), ip"

    return runQuery(query)
        .select("ip, accounts, fails, successes, non_existing_accounts, blocked_accounts")
}

这使用下面的用户定义表函数，它已经在我的tableEnv注册为blockedEmailsList：

public class BlockedEmailsList extends TableFunction<Row> {
    private Collection<String> emails;

    public BlockedEmailsList(Collection<String> emails) {
        this.emails = emails;
    }

    public Row read(String email) {
        return Row.of(email);
    }

    public void eval() {
        this.emails.forEach(email -> collect(read(email)));
    }
}

但是，它返回以下错误：

Caused by: org.apache.flink.table.api.TableException: Rowtime attributes must not be in the input rows of a regular join. As a workaround you can cast the time attributes of input tables to TIMESTAMP before.

如果我按照建议进行操作并将created_at 转换为TIMESTAMP，我会得到这个：

org.apache.flink.client.program.ProgramInvocationException: The main method caused an error: Window can only be defined over a time attribute column.

我在 Stack Overflow 上发现了与这些异常相关的其他问题，但它们涉及流和时态表，并且它们都没有解决将流加入静态列表的情况。

有什么想法吗？

编辑：对于我的用例，Flink 项目中似乎有一个未解决的问题：https://cwiki.apache.org/confluence/display/FLINK/FLIP-17+Side+Inputs+for+DataStream+API

所以，我也接受解决方法建议。

Answer 1

Caused by: org.apache.flink.table.api.TableException: Rowtime attributes must not be in the input rows of a regular join. As a workaround you can cast the time attributes of input tables to TIMESTAMP before.

原因是横向表函数是Flink正则连接，正则连接会发送空值，例如

left:（K0, A）, right（K1, T1）  => send    (K0, A, NULL, NULL)
left:         , right（K0, T2） => retract (K0, A, NULL, NULL )  
                                   send   (K0, A, K0, T2)

因此输入流中的时间属性在加入后会丢失。

在您的情况下，您不需要 TableFunction，您可以使用标量函数 喜欢：

 public static class BlockedEmailFunction extends ScalarFunction {
     private static List<String> blockedEmails = ...;
     public Boolean eval(String email) {
        return blockedEmails.contains(attempt.getEmail());
     }
 }


// register function
env.createTemporarySystemFunction("blockedEmailFunction", BlockedEmailFunction.class);

// call registered function in SQL and do window operation as your expected
env.sqlQuery("SELECT blockedEmailFunction(email) as status, ip, createdAt FROM Attempts");
 

Answer 2

我设法实施了解决我的问题的解决方法！

我没有将流式尝试与电子邮件的静态列表连接起来，而是预先将每个尝试映射到一个带有 blockedEmail 属性的新尝试。如果静态列表blockedEmails 包含当前的尝试电子邮件，我将其blockedEmail 属性设置为true。

DataStream<Attempt> attemptsStream = sourceApi.<Attempt>startStream().map(new MapFunction<Attempt, Attempt>() {
    @Override
    public Attempt map(Attempt attempt) throws Exception {
        if (blockedEmails.contains(attempt.getEmail())) {
            attempt.setBlockedEmail(true);
        }
        return attempt;
    }
});

静态列表blockedEmails 的类型为HashSet，因此查找时间为O(1)。

最后，分组查询调整为：

override fun performQuery(): Table {
    val query = "SELECT ip, " +
        "COUNT(CASE WHEN success IS false THEN 1 END) AS fails, " +
        "COUNT(CASE WHEN success IS true THEN 1 END) AS successes, " +
        "COUNT(DISTINCT id) accounts, " +
        "COUNT(CASE WHEN id = 0 THEN 1 END) AS non_existing_accounts, " +
        "COUNT(CASE WHEN blockedEmail IS true THEN 1 END) AS blocked_accounts " +
        "FROM Attempts " +
        "WHERE Attempts.email <> '' " +
        "GROUP BY HOP(Attempts.createdAt, INTERVAL '10' SECOND, INTERVAL '30' MINUTE), ip"

    return runQuery(query)
        .select("ip, accounts, fails, successes, non_existing_accounts, blocked_accounts")
}

到目前为止，流和静态列表之间的连接问题似乎尚未解决，但在我的情况下，上述解决方法解决了它。