SQL窗口函数详解：从基础语法到实战应用

1. 窗口函数核心概念解析

窗口函数（Window Function）是SQL中一种强大的分析工具，它能够在保留原始行明细的同时，对数据的特定"窗口"进行计算。与传统的GROUP BY聚合不同，窗口函数不会合并行，而是将计算结果附加到每一行上。

1.1 窗口函数与GROUP BY的本质区别

传统GROUP BY聚合的工作原理：

• 将数据按指定列分组
• 每组合并为单行结果
• 原始行细节丢失
• 适合汇总统计但不保留明细

窗口函数的典型特征：

• 保持原始行不变
• 计算结果作为新列添加
• 可同时查看明细和聚合结果
• 适合需要保留完整数据场景

实际经验：在电商分析中，GROUP BY适合计算每日总销售额，而窗口函数适合分析每笔订单在用户所有订单中的排名或占比。

1.2 窗口函数语法结构详解

完整窗口函数语法包含三个核心部分：

sql复制函数名(参数) OVER(
  PARTITION BY 分组字段
  ORDER BY 排序字段
  [窗口范围FRAME子句]
)

• PARTITION BY：定义计算的分组范围，类似GROUP BY但不合并行
• ORDER BY：确定窗口内的排序规则，影响排名和累积计算
• FRAME子句：精确控制计算涉及的行范围（如"前3行到当前行"）

2. 排名类窗口函数实战

2.1 ROW_NUMBER() 唯一序号

基本特性：

• 为每行分配唯一序号
• 即使值相同也会分配不同序号
• 不会出现并列情况

典型应用场景：

• 为每个用户的订单按时间编号
• 获取每组最新/最早的记录

sql复制SELECT 
  user_id,
  order_id,
  created_at,
  ROW_NUMBER() OVER(PARTITION BY user_id ORDER BY created_at) AS order_seq
FROM orders;

2.2 RANK() 与 DENSE_RANK() 的区别

RANK()特点：

• 相同值获得相同排名
• 后续排名会"跳号"
• 如：1,1,3,4,4,6...

DENSE_RANK()特点：

• 相同值获得相同排名
• 后续排名不跳号
• 如：1,1,2,3,3,4...

实际案例对比：

sql复制SELECT
  product_id,
  sales_amount,
  RANK() OVER(ORDER BY sales_amount DESC) AS rank_val,
  DENSE_RANK() OVER(ORDER BY sales_amount DESC) AS dense_rank_val
FROM products;

2.3 TOP N查询最佳实践

获取每个用户支付金额最高的3笔订单（允许并列）：

sql复制SELECT *
FROM (
  SELECT
    o.*,
    RANK() OVER(PARTITION BY user_id ORDER BY pay_amount DESC) AS rnk
  FROM orders
  WHERE paid_at IS NOT NULL
) t
WHERE t.rnk <= 3;

注意事项：RANK()可能返回多于N条记录（因并列），如需严格N条应使用ROW_NUMBER()

3. 聚合类窗口函数深度应用

3.1 基础窗口聚合

计算每个用户的总GMV（同时保留订单明细）：

sql复制SELECT
  user_id,
  order_id,
  pay_amount,
  SUM(pay_amount) OVER(PARTITION BY user_id) AS user_gmv,
  pay_amount/SUM(pay_amount) OVER(PARTITION BY user_id) AS amount_ratio
FROM orders
WHERE paid_at IS NOT NULL;

3.2 累积计算与移动平均

3.2.1 累积求和实现

按时间顺序计算用户累计消费金额：

sql复制SELECT
  user_id,
  order_id,
  created_at,
  pay_amount,
  SUM(pay_amount) OVER(
    PARTITION BY user_id
    ORDER BY created_at
    ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW
  ) AS running_total
FROM orders
WHERE paid_at IS NOT NULL;

3.2.2 移动平均计算

计算每个用户最近3笔订单的平均金额：

sql复制SELECT
  user_id,
  order_id,
  created_at,
  pay_amount,
  AVG(pay_amount) OVER(
    PARTITION BY user_id
    ORDER BY created_at
    ROWS BETWEEN 2 PRECEDING AND CURRENT ROW
  ) AS avg_last_3
FROM orders
WHERE paid_at IS NOT NULL;

3.3 FRAME子句详解

FRAME子句定义了窗口函数计算的行范围，主要语法元素：

• ROWS vs RANGE：

  • ROWS：按物理行偏移
  • RANGE：按逻辑值范围

• 边界定义：

  • UNBOUNDED PRECEDING：分区第一行
  • n PRECEDING：当前行前n行
  • CURRENT ROW：当前行
  • n FOLLOWING：当前行后n行
  • UNBOUNDED FOLLOWING：分区最后一行

常见组合示例：

sql复制-- 从开始到当前行（默认）
ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW

-- 最近3行（包括当前）
ROWS BETWEEN 2 PRECEDING AND CURRENT ROW

-- 前后各1行（共3行）
ROWS BETWEEN 1 PRECEDING AND 1 FOLLOWING

-- 整个分区
ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING

4. 高级技巧与性能优化

4.1 CTE与子查询的应用

窗口函数结果不能直接在WHERE中过滤，必须使用子查询或CTE：

sql复制-- 使用CTE方式
WITH ranked_orders AS (
  SELECT
    o.*,
    ROW_NUMBER() OVER(PARTITION BY user_id ORDER BY created_at DESC) AS rn
  FROM orders o
)
SELECT * FROM ranked_orders WHERE rn = 1;

-- 使用子查询方式
SELECT * FROM (
  SELECT
    o.*,
    ROW_NUMBER() OVER(PARTITION BY user_id ORDER BY created_at DESC) AS rn
  FROM orders o
) t WHERE t.rn = 1;

4.2 多窗口函数组合使用

单次查询可包含多个窗口函数，各自独立定义：

sql复制SELECT
  user_id,
  order_id,
  created_at,
  pay_amount,
  ROW_NUMBER() OVER(PARTITION BY user_id ORDER BY created_at) AS order_seq,
  SUM(pay_amount) OVER(PARTITION BY user_id) AS user_gmv,
  pay_amount/SUM(pay_amount) OVER(PARTITION BY user_id) AS amount_pct,
  AVG(pay_amount) OVER(
    PARTITION BY user_id
    ORDER BY created_at
    ROWS BETWEEN 2 PRECEDING AND CURRENT ROW
  ) AS avg_last_3
FROM orders
WHERE paid_at IS NOT NULL;

4.3 性能优化建议

1. 减少分区数量：PARTITION BY列的选择影响性能，高基数列会导致大量小分区
2. 合理使用索引：确保ORDER BY和PARTITION BY列有适当索引
3. 限制窗口范围：明确指定FRAME子句而非依赖默认值
4. 避免嵌套窗口：多层嵌套窗口函数会显著降低性能

5. 实际业务场景案例

5.1 用户行为分析

计算用户访问间隔及会话划分：

sql复制WITH user_visits AS (
  SELECT
    user_id,
    visit_time,
    LAG(visit_time) OVER(PARTITION BY user_id ORDER BY visit_time) AS prev_visit
  FROM user_activity
)
SELECT
  user_id,
  visit_time,
  prev_visit,
  TIMESTAMPDIFF(MINUTE, prev_visit, visit_time) AS mins_since_last,
  CASE WHEN TIMESTAMPDIFF(MINUTE, prev_visit, visit_time) > 30 OR prev_visit IS NULL 
       THEN 1 ELSE 0 END AS is_new_session
FROM user_visits;

5.2 销售漏斗分析

计算各阶段转化率：

sql复制WITH funnel_steps AS (
  SELECT
    user_id,
    MAX(CASE WHEN event_type = 'view' THEN 1 ELSE 0 END) AS viewed,
    MAX(CASE WHEN event_type = 'cart' THEN 1 ELSE 0 END) AS carted,
    MAX(CASE WHEN event_type = 'purchase' THEN 1 ELSE 0 END) AS purchased
  FROM user_events
  GROUP BY user_id
),
funnel_counts AS (
  SELECT
    SUM(viewed) AS total_views,
    SUM(carted) AS total_carts,
    SUM(purchased) AS total_purchases
  FROM funnel_steps
)
SELECT
  total_views,
  total_carts,
  total_purchases,
  total_carts/total_views AS view_to_cart_rate,
  total_purchases/total_carts AS cart_to_purchase_rate
FROM funnel_counts;

5.3 时间序列分析

计算7日移动平均和同比变化：

sql复制WITH daily_metrics AS (
  SELECT
    report_date,
    SUM(sales) AS daily_sales
  FROM sales_data
  GROUP BY report_date
)
SELECT
  report_date,
  daily_sales,
  AVG(daily_sales) OVER(
    ORDER BY report_date
    ROWS BETWEEN 6 PRECEDING AND CURRENT ROW
  ) AS weekly_moving_avg,
  daily_sales/LAG(daily_sales, 364) OVER(ORDER BY report_date) AS yoy_change
FROM daily_metrics
ORDER BY report_date;

窗口函数是SQL数据分析的利器，掌握它能大幅提升数据处理效率。在实际应用中，建议先从简单场景入手，逐步尝试复杂分析模式。根据我的经验，合理使用窗口函数通常能将原本需要多步处理的任务简化为单次查询，同时保持代码的可读性和维护性。