SQL (Hive):在使用 GROUP BY 进行聚合时使用窗口函数
Posted
技术标签:
【中文标题】SQL (Hive):在使用 GROUP BY 进行聚合时使用窗口函数【英文标题】:SQL (Hive): using window functions while aggregating with GROUP BY 【发布时间】:2020-12-16 13:00:54 【问题描述】:我在Athena(Hive/Presto)中有下表:
CREATE EXTERNAL TABLE tmp (
id STRING,
updated_at TIMESTAMP,
location STRING,
direction STRING
)
LOCATION 's3://path';
我需要聚合和计数id
字段,同时选择与组内最新的timestamp
相关的location
和direction
(分区再次位于id
)。
到目前为止,我想出了以下查询,首先利用窗口函数,然后再分组:
SELECT
b.id,
MAX(b.latest_location) AS "latest_location", -- It seems it is not possible to use first_value() on GROUP BY
MAX(b.latest_direction) AS "latest_direction",
COUNT(*) AS "total"
FROM (
SELECT
a.id,
first_value(a.location) OVER (PARTITION BY a.id ORDER BY a.updated_at DESC) AS "latest_location",
first_value(a.direction) OVER (PARTITION BY a.id ORDER BY a.updated_at DESC) AS "latest_direction"
FROM tmp a
) b
GROUP BY b.id;
我第一次尝试同时做group by aggregation和window aggregation,但是好像引擎不允许这样做。是否可以编写更高效的查询(可能没有子查询)?
【问题讨论】:
您在内部查询中执行select distinct
并添加count(*) over (partition by a.id)
。它会更短,但我不确定内部执行效率是否会发生很大变化。
【参考方案1】:
SELECT DISTINCT
id,
first_value(a.location) OVER (PARTITION BY id ORDER BY updated_at DESC) AS latest_location,
first_value(a.direction) OVER (PARTITION BY id ORDER BY updated_at DESC) AS latest_direction,
count(*) OVER (PARTITION BY id) as total
FROM tmp
在您的原始查询中,max
基本上是一个虚拟聚合,因为所有行都具有相同的值。 group by
基本上是在做 distinct
在这里所做的事情。
【讨论】:
【参考方案2】:添加到首选答案 - 考虑更正式的窗口定义支持 DRY(不要重复自己)偏好:
SELECT DISTINCT
id,
first_value(a.location) OVER w AS latest_location,
first_value(a.direction) OVER w AS latest_direction,
count(*) OVER (PARTITION BY id) as total
FROM tmp
WINDOW w AS (PARTITION BY id ORDER BY updated_at DESC)
这将允许将更复杂的窗口定义精确地维护在一个地方,并保证两个列计算使用相同的窗口逻辑。
【讨论】:
【参考方案3】:您可以混合使用窗口函数和聚合函数。 . .但在另一个方向:先聚合,然后是窗口函数。
也就是说,如果您消除聚合,您的查询应该会更快。只需使用row_number()
和过滤:
SELECT a.id, a.location, a.updated_at
FROM (SELECT a.*,
ROW_NUMBER() OVER (PARTITION BY a.id ORDER BY a.updated_at DESC) AS seqnum
FROM tmp a
) a
WHERE seqnum = 1;
【讨论】:
以上是关于SQL (Hive):在使用 GROUP BY 进行聚合时使用窗口函数的主要内容,如果未能解决你的问题,请参考以下文章
hive高阶1--sql和hive语句执行顺序explain查看执行计划group by生成MR
Hive mapreduce SQL实现原理——SQL最终分解为MR任务,而group by在MR里和单词统计MR没有区别了