假设存在表格如下：
代码语言：javascript复制select 'a' as category, 19 as duration
union all
select 'b' as category, 15 as duration
union all
select 'c' as category, 12 as duration
union all
select 'd' as category, 53 as duration
union all
select 'e' as category, 27 as duration
union all
select 'f' as category, 9 as duration;
category | duration
b | 15
f | 9
e | 27
c | 12
d | 53
a | 19 想要多行转一行并且按照duration排序，形成如下效果d,e,a,b,c,f
首先排序：row_number() over (partition by category order by cast(duration as int) desc) duration_rank，然后拼接concat_ws(',',collect_set(category))，但是得到的结果却是乱序的，产生这个问题的根本原因自然在MapReduce，如果启动了多于一个mapper/reducer来处理数据，select出来的数据顺序就几乎肯定与原始顺序不同了。
解决方法可以把mapper数固定成1，或者把rank加进来再进行一次排序，拼接完之后把rank去掉：
代码语言：javascript复制select
regexp_replace(
concat_ws(',',
sort_array(
collect_list(
concat_ws(':',lpad(cast(duration_rank as string),5,'0'),cast(category as string))
)
)
),
'\d :','')
from
(select
category
,row_number() over (order by cast(duration as int) desc) duration_rank
from
(select 'a' as category, 19 as duration
union all
select 'b' as category, 15 as duration
union all
select 'c' as category, 12 as duration
union all
select 'd' as category, 53 as duration
union all
select 'e' as category, 27 as duration
union all
select 'f' as category, 9 as duration) t
) T;duration_rank 必须要在高位补足够的0对齐，因为排序的是字符串而不是数字，如果不补0的话，按字典序排序就会变成1, 10, 11, 12, 13, 2, 3, 4...，又不对了。将排序的结果拼起来之后，用regexp_replace函数替换掉冒号及其前面的数字，大功告成。
