SQL Oracle 脚本优化答案

【问题标题】：SQL Oracle script optimizationSQL Oracle 脚本优化
【发布时间】：2026-01-14 21:20:03
【问题描述】：

我有一个表 TRANSACTIONS，其中包含近 3000 万个事务（13 列）。如何优化以下代码？我尝试了自我加入，但它似乎不太有效。

逻辑：如果 receiver_2 存在，我想通过 sender-receiver_2 获取最后交易，否则通过 sender-receiver + 计算一些统计信息（ 10/30/90 天）

SELECT T.* FROM
(SELECT T.*, row_number() over (partition by T.SENDER, (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END) order by T.DATE_ACCEPT desc) as seqnum 
FROM 
(
SELECT T.*
      ,(SELECT COUNT(DISTINCT T2.ID_TRAN)
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 10  AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
              AND
              T2.SENDER = T.SENDER
        ) CNT_10
      ,(SELECT COUNT(DISTINCT T2.ID_TRAN)
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 30 AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
              AND
              T2.SENDER = T.SENDER 
        ) CNT_30
      ,(SELECT COUNT(DISTINCT T2.ID_TRAN)
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 90  AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
              AND
              T2.SENDER = T.SENDER 
        ) CNT_90 
        ,(SELECT DISTINCT AVG(CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END) OVER()
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 10 AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
             AND
              T2.SENDER = T.SENDER
        GROUP BY T2.ID_TRAN, (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END)
        ) AVG_AMOUNT_10
      ,(SELECT DISTINCT AVG(CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END) OVER()
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 30 AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
              AND
              T2.SENDER = T.SENDER
        GROUP BY T2.ID_TRAN, (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END)
        ) AVG_AMOUNT_30
        ,(SELECT DISTINCT AVG(CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END) OVER()
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 90 AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
              AND
              T2.SENDER = T.SENDER
        GROUP BY T2.ID_TRAN, (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END)
        ) AVG_AMOUNT_90
        ,(SELECT MAX(CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END)
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 10 AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
              AND
              T2.SENDER = T.SENDER
        ) MAX_AMOUNT_10
        ,(SELECT MAX(CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END)
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 30 AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
              AND
              T2.SENDER = T.SENDER 
        ) MAX_AMOUNT_30
        ,(SELECT MAX(CASE WHEN T.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END)
        FROM TRANSACTIONS T2
        WHERE T2.DATE_ACCEPT > T.DATE_ACCEPT - 90 AND
              T2.DATE_ACCEPT < T.DATE_ACCEPT AND
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T2.RECEIVER ELSE T2.RECEIVER_2 END) =
              (CASE WHEN T.RECEIVER_2 IS NULL THEN T.RECEIVER ELSE T.RECEIVER_2 END)
              AND
              T2.SENDER = T.SENDER 
        ) MAX_AMOUNT_90
FROM TRANSACTIONS T
) T ) T
WHERE T.SEQNUM = 1

我还在 (SENDER, DATE_ACCEPT) 上创建了索引。

Query plan

TABLE EXAMPLE

【问题讨论】：

请添加带索引的查询计划
您的 SQL 中可能存在复制粘贴错误：您的子查询中总是有 WHEN T.RECEIVER_2 IS NULL THEN，但是，也应该使用 WHEN T2.RECEIVER_2 IS NULL THEN。
您能否提供所需的结果（基于您的样本数据）？请将表格数据提供为表格，而不是截图，请参阅meta.*.com/q/285551

标签： sql oracle optimization

【解决方案1】：

你知道Analytic Functions Windowing Clause吗？

我不明白您查询的逻辑，但我想可能没有任何自联接。看看这个查询，它可能是一个起点：

SELECT 
    COUNT(ID_TRAN) OVER (PARTITION BY SENDER, NVL(RECEIVER_2, RECEIVER) ORDER BY DATE_ACCEPT RANGE BETWEEN INTERVAL '10' DAY PRECEDING AND CURRENT ROW) AS CNT_10,
    COUNT(ID_TRAN) OVER (PARTITION BY SENDER, NVL(RECEIVER_2, RECEIVER) ORDER BY DATE_ACCEPT RANGE BETWEEN INTERVAL '30' DAY PRECEDING AND CURRENT ROW) AS CNT_30,
    COUNT(ID_TRAN) OVER (PARTITION BY SENDER, NVL(RECEIVER_2, RECEIVER) ORDER BY DATE_ACCEPT RANGE BETWEEN INTERVAL '90' DAY PRECEDING AND CURRENT ROW) AS CNT_90,
    AVG(NVL(T.AMOUNT_2, T.AMOUNT)) OVER (PARTITION BY SENDER, NVL(RECEIVER_2, RECEIVER) ORDER BY DATE_ACCEPT RANGE BETWEEN INTERVAL '30' DAY PRECEDING AND CURRENT ROW) AS AVG_30,
    AVG(NVL2(T.RECEIVER_2, T.AMOUNT_2, T.AMOUNT)) OVER (PARTITION BY SENDER, NVL(RECEIVER_2, RECEIVER) ORDER BY DATE_ACCEPT RANGE BETWEEN INTERVAL '90' DAY PRECEDING AND CURRENT ROW) AS AVG_90
FROM TRANSACTIONS

注意，RANGE BETWEEN INTERVAL '10' DAY PRECEDING AND CURRENT ROW) 等于 RANGE INTERVAL '10' DAY PRECEDING)

另一个注意事项，当我对示例数据运行查询时，我得到了

+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
|ID_TRAN|SENDER|RECEIVER|RECEIVER_2|AMOUNT|AMOUNT_2|DATE_ACCEPT        |CNT_10|CNT_30|CNT_90|AVG_AMOUNT_10|AVG_AMOUNT_30|AVG_AMOUNT_90|MAX_AMOUNT_10|MAX_AMOUNT_30|MAX_AMOUNT_90|SEQNUM|
+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
|1      |00010 |22222   |1112      |3000  |1000    |16.04.2021 14:01:00|0     |0     |0     |             |             |             |             |             |             |1     |
|1      |00010 |22222   |2114      |3000  |2000    |16.04.2021 14:01:00|0     |0     |0     |             |             |             |             |             |             |1     |
|2      |01236 |45872   |          |4000  |        |01.04.2021 22:01:00|0     |0     |0     |             |             |             |             |             |             |1     |
|3      |45872 |00010   |          |5000  |        |17.04.2021 14:01:00|0     |0     |0     |             |             |             |             |             |             |1     |
+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+

这看起来毫无意义。

【讨论】：

我尝试使用分析函数，但平均值计算不正确，我将按照您的示例尝试。并且示例表只是用来理解表中数据的一块，不是用来检查计算的
“计算一些统计数据（10/30/90 天）”并不是很精确。但是窗口函数实际上是为这样的计算而设计的。
这是 windowing_clause 的一个很好的用法（不确定是谁以及为什么 dw），不幸的是它在 ORA-30487 计算时失败了 count distinct .但是，如果事务表包含 unique ID_TRANS @SunMoon，则 count distinct 可能是 overkill
OP 的示例实际上实现的不是AND CURRENT ROW ，而是AND CURRENT ROW EXCLUDE TIES（因为他只考虑lover 日期的交易而不是last 交易）。这就是样本数据结果中出现nulls 的原因。不确定这是有意还是仅已实现。 EXCLUDE TIES 需要 21c。 @SunMoon 另请参阅我的回答中的相关说明。
@MarmiteBomber，使用RANGE BETWEEN INTERVAL '90' DAY PRECEDING AND INTERVAL '0.001' SECOND PRECEDING 也应该可以。

【解决方案2】：

查询中的主要问题是谓词中的CASE。它使任何索引的使用无效。因此，您需要使用虚拟列：

ALTER TABLE Transactions ADD rec AS (
     CASE WHEN RECEIVER_2 IS NULL 
     THEN RECEIVER ELSE RECEIVER_2 END
);

第二步是用该列创建索引：

CREATE INDEX ix_transactions_sender_rec 
    ON Transactions(sender, rec, date_accept)

但是，由于查询语法的原因，可能不会使用索引。将CASE 语法替换为新创建的列rec，并将greatest per group 解决方案重写为自联接。我添加了简化的 SQL 示例来说明如何操作。

select t.*,
    (
           select count(DISTINCT T2.id_tran)
           from transactions T2
           where T2.date_accept > T.date_accept - 10
                 AND T2.date_accept < T.date_accept
                 AND T2.rec = T.rec
                 AND T2.sender = T.sender
    ) CNT_10
from (
    select sender, rec, max(date_accept)
    from transactions
    group sender, rec
) tmax 
join transactions t on t.sender = tmax.sender and
                       t.rec = tmax.rec and
                       t.date_accept = tmax.date_accept

如果您希望统计子查询速度超快，还可以添加其中使用的其他列：

CREATE INDEX ix_transactions_sender_rec 
    ON Transactions(sender, rec, date_accept, id_tran, amount)

【讨论】：

问题是receiver_2可以在receiver的位置。我需要将它们分开，因为总和会明显不同。 Receiver - 组织主管，receiver_2 - 员工
@SunMoon 我知道您需要将它们分开。我的建议包括一个虚拟列，它简化了查询并启用了索引使用。我不建议摆脱 receiver_2 和 receiver 属性。
我明白了。谢谢我试试

【解决方案3】：

在 (SENDER, DATE_ACCEPT) 上建立索引可能会有所帮助。

您可以通过使用带有条件聚合的LATERAL JOIN 来简化和加速查询。
它允许计算超过 1 个 COUNT/AVG/MAX。

例如：

SELECT T.*, LT.*
FROM (
  SELECT SENDER
    , RECEIVER, RECEIVER_2
    , DATE_ACCEPT
    , AMOUNT, AMOUNT_2 
  FROM (
    SELECT SENDER
    , RECEIVER, RECEIVER_2
    , DATE_ACCEPT
    , AMOUNT, AMOUNT_2 
    , ROW_NUMBER() OVER (PARTITION BY SENDER, NVL(RECEIVER_2, RECEIVER) ORDER BY DATE_ACCEPT DESC) AS RN
    FROM TRANSACTIONS
  ) TRANS
  WHERE RN = 1
) T
CROSS JOIN LATERAL (
  SELECT 
    COUNT(DISTINCT
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 10 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN T2.ID_TRAN
    END) AS CNT_10
  , COUNT(DISTINCT
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 30 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN T2.ID_TRAN
    END) AS CNT_30
  , COUNT(DISTINCT
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 90 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN T2.ID_TRAN
    END) AS CNT_90
  , NVL(AVG(
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 10 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN CASE WHEN T2.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END
    END), 0) AS AVG_AMOUNT_10
  , NVL(AVG(
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 30 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN CASE WHEN T2.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END
    END), 0) AS AVG_AMOUNT_30
  , NVL(AVG(
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 90 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN CASE WHEN T2.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END
    END), 0) AS AVG_AMOUNT_90
  , NVL(MAX(
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 10 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN CASE WHEN T2.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END
    END), 0) AS MAX_AMOUNT_10
  , NVL(MAX(
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 30 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN CASE WHEN T2.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END
    END), 0) AS MAX_AMOUNT_30
  , NVL(MAX(
    CASE WHEN T2.DATE_ACCEPT > T.DATE_ACCEPT - 90 
          AND T2.DATE_ACCEPT < T.DATE_ACCEPT
    THEN CASE WHEN T2.RECEIVER_2 IS NULL THEN T2.AMOUNT ELSE T2.AMOUNT_2 END
    END), 0) AS MAX_AMOUNT_90
  FROM TRANSACTIONS T2
  WHERE T2.SENDER = T.SENDER 
    AND T2.DATE_ACCEPT > T.DATE_ACCEPT - 90
    AND NVL(T2.RECEIVER_2, T2.RECEIVER) = NVL(T.RECEIVER_2, T.RECEIVER)
) LT;

SENDER	RECEIVER	RECEIVER_2	DATE_ACCEPT	AMOUNT	AMOUNT_2	CNT_10	CNT_30	CNT_90	AVG_AMOUNT_10	AVG_AMOUNT_30	AVG_AMOUNT_90	MAX_AMOUNT_10	MAX_AMOUNT_30	MAX_AMOUNT_90
1	2	3	30-MAR-21	10	20	1	2	3	11.2	21.65	45.5	11.2	32.1	93.2

dbfiddle here

上的演示

【讨论】：

【解决方案4】：

您与receiver / receiver2 的逻辑使它只是令人困惑，这不是性能问题的原因。

对于我在示例中使用的 SENDER, RECEIVER, AMOUNT 和 DATE_ACCEPT 的简单模型，您会遇到同样的问题 - 适应您的目的。

首先你应该明白问题的原因是什么。

您正在加入一个带有历史记录的大型事务表，产生的结果很大聚合它并计算聚合度量。

关键思想是先聚合，第二步加入事务表。

下面的查询首先计算每个发送者/接收者的max_date_accept，以计算使用历史窗口在下一步中聚合度量（例如窗口 10 天 - 根据需要进行调整）。

请注意，我通过添加谓词 DATE_ACCEPT < max_date_accept 复制了您在计算中忽略 last 事务的逻辑。

如果在计算的时间间隔内只有一笔交易，这将导致NULL 上的结果，这可能不是您想要的。

with trans as (
select 
 ID_TRAN, SENDER,  RECEIVER, AMOUNT,  DATE_ACCEPT,
 max(DATE_ACCEPT) over (partition by T.SENDER,  T.RECEIVER) max_date_accept
from TRANSACTIONS t
)
select 
  SENDER, RECEIVER,
  count(distinct case when DATE_ACCEPT > max_date_accept - 10 and DATE_ACCEPT < max_date_accept then ID_TRAN end) CNT_10,
  avg(case when DATE_ACCEPT > max_date_accept - 10 and DATE_ACCEPT < max_date_accept then AMOUNT end) AVG_AMOUNT_10,
  max(case when DATE_ACCEPT > max_date_accept - 10 and DATE_ACCEPT < max_date_accept then AMOUNT end) MAX_AMOUNT_10
from trans  
group by SENDER, RECEIVER;

可能这个结果已经是你想要的了，但是如果你真的想要完整的集合来自事务表的列与第一个事务的值，简单地将聚合结果连接到事务表：

with trans as (
select 
 ID_TRAN, SENDER,  RECEIVER, AMOUNT,  DATE_ACCEPT,
 max(DATE_ACCEPT) over (partition by T.SENDER,  T.RECEIVER) max_date_accept
from TRANSACTIONS t
),
agg as (
select 
  SENDER, RECEIVER,
  count(distinct case when DATE_ACCEPT > max_date_accept - 10 and DATE_ACCEPT < max_date_accept then ID_TRAN end) CNT_10,
  avg(case when DATE_ACCEPT > max_date_accept - 10 and DATE_ACCEPT < max_date_accept then AMOUNT end) AVG_AMOUNT_10,
  max(case when DATE_ACCEPT > max_date_accept - 10 and DATE_ACCEPT < max_date_accept then AMOUNT end) MAX_AMOUNT_10
from trans  
group by SENDER, RECEIVER),
trans2 as (
select
 t.*,
 row_number() over (partition by SENDER, RECEIVER order by DATE_ACCEPT desc) as seqnum
from TRANSACTIONS t)
select
 trans2.*,
 agg.CNT_10, agg.AVG_AMOUNT_10, agg.MAX_AMOUNT_10
from trans2
join agg on trans2.SENDER = agg.SENDER and trans2.RECEIVER = agg.RECEIVER
where seqnum = 1;

性能说明 - 检查查询的 execution plan。

您应该只看到TABLE ACCESS FULL 和HASH JOIN。如果他们使用NESTED LOOPS 或FILTER 与INDEX ACCESS 连接，您的类型的查询经常会出现问题。

【讨论】：