【Hive】从执行计划DAG中执行慢的Task，找到对应SQL逻辑片段

**

【Hive】从执行计划DAG中执行慢的Task，找到对应SQL逻辑片段

**

一个稍微复杂的Hive SQL，在执行过程中发现某个Task非常慢，怎么去定位这个Task是属于哪段SQL逻辑呢

对于如下逻辑的一个SQL，在Spark引擎和Tez引擎中执行DAG的比较。在两个引擎下，执行慢的逻辑是一样的，分别看两种引擎的执行情况。

WITH tbl_a AS
         (
             SELECT 
                  a,
                  b,
                  name,
                  SUM(pv) AS pv
             FROM (
                      SELECT a,b,name
                           , count(1)              AS pv
                      FROM dataware.a
                      WHERE dt = '{@date}'
                      GROUP BY a,b,name 

                      UNION ALL

                      SELECT a,b,name
                           , count(1)    AS pv
                      FROM dateware.b
                      WHERE dt = '{@date}'
                      GROUP BY a,b,name 
                        

                      UNION ALL

                      SELECT a,b,name
                           , count(1)      AS pv
                      FROM dateware.c
                      WHERE dt = '{@date}'
                      GROUP BY a,b,name 

                      UNION ALL

                      SELECT a,b,name
                           , count(1)       AS pv
                      FROM dataware.d
                      WHERE dt = '{@date}'
                      GROUP BY a,b,name 
                  ) a
             GROUP BY a,
                      b,
                      name
         ),
     tbl_b AS (
         SELECT *
         FROM (
                  SELECT 
                       a,
                       ,b
                       ,row_number() OVER (PARTITION BY app ORDER BY min_time DESC) AS row_num
                  FROM dataware.demo_a
                  WHERE dt = '{@date}'
              ) tmp
         WHERE row_num = 1
     ),
     tbl_c AS (
         SELECT 
              a,
              b,
              user_id
         FROM dataware.demo_b
         WHERE dt = '{@date}'
     ),
     tbl_d AS (
         SELECT user_id
              , identity as role
         FROM dataware.demo_c
         WHERE dt = '{@date}'
     )

INSERT OVERWRITE TABLE dataware.result PARTITION ( dt = '{@date}')
SELECT a,
       b,
       name,
       other_fields,
     , SUM(pv) AS pv
FROM tbl_a
         LEFT JOIN tbl_b ON tbl_a.a = tbl_b.a AND tbl_a.b = tbl_b.b
         LEFT JOIN tbl_c ON tbl_a.a = tbl_c.a AND tbl_a.b = tbl_c.b
         LEFT JOIN tbl_d ON tbl_c.user_id = tbl_d.user_id
GROUP BY a,
         b,
         name,
         other_fields

Spark引擎当中，从执行结果上看。执行慢的Stage分别是Stage1和Stage8。刚入门要想看出来Stage1和Stage8分别对应哪部分SQL逻辑有一定困难。

在Spark引擎中，我总结可能会有如下办法找到对应SQL逻辑：

1、根据经验将SQL分开。例子中union all的逻辑很好对应。

Stage1是读取tbl_b表阶段

2、也可以单独执行某个SQL逻辑段，看看【单独执行SQL逻辑段的STAGE】与【整体Stage哪个部分】相符。

Stage7对应读取dataware.demo_a表逻辑

3、如果是个读取文件的过程，一般Task的并行度与文件块个数是对应的。所以查看这个表的block块，确定Stage对应SQL逻辑段

Stage5对应是读取dataware.demo_b表逻辑

4、参考Tez引擎

Tez引擎执行情况：

Tez引擎执行情况：下图按Vertices的开始时间排序。看出Map8和Reduce9阶段耗时长。

在 Graphical View中可以很容易看到Map8和Reduce9对应的读取tbl_b的阶段