一次sparksql问题排查记录

问题： 在调试一个sparksql左连接查询时发现数据结果不正确，经过一天折腾才发现使用子查询方式能够得到正确的结果，分析执行计划发现第一种写法的优化后的执行计划将where t.ip is null and t.dn条件错误的加到了左表子查询中了，即红色标出的地方，这样导致左表子查询查不出数据来。

结论： 过滤条件写在where条件中时，spark会将sql优化为inner join， 如果连接条件中的字段出现在最后的where条件中，那么该条件在做谓词下推时也会被加到左表和右表中，此时就不符合预期结果，即会导致左表中的查不到预期的数据，但是将过滤数据用的限定条件写到子查询中时查出的结果是正确的，执行计划也是正确的，原因不详，怀疑是spark执行计划优化中的bug；

过程数据记录
1、条件在where中
select
oneday.dn, oneday.ip, '20201202', '20201202'
from
(
select
ip,dn
from dwd_dns.t_ip_dn_his_rel2
where dt = '20201202'
group by ip,dn
) oneday left join dwd_dns.t_ip_dn_first t on t.ip = oneday.ip and t.dn = oneday.dn
where t.ip is null and t.dn is null and t.dt = '20201001'

执行计划：

== Optimized Logical Plan ==
InsertIntoHiveTable dwd_dns.t_ip_dn_first, org.apache.hadoop.hive.ql.io.orc.OrcSerde, Map(dt -> None), true, false, [dn, ip, first_time, dt]
+- Project [dn#1, ip#2, 20201202 AS first_time#28, 20201202 AS dt#29]
+- Join Inner, ((ip#8 = ip#2) && (dn#7 = dn#1))
:- Aggregate [ip#2, dn#1], [ip#2, dn#1]
: +- Project [dn#1, ip#2]
: +- Filter (((((isnotnull(dt#6) && (dt#6 = 20201202)) && isnull(dn#1)) && isnull(ip#2)) && isnotnull(ip#2)) && isnotnull(dn#1))
: +- Relation[uid#0,dn#1,ip#2,cname#3,dnsip#4,probe_time#5,dt#6] orc
+- Project [dn#7, ip#8]
+- Filter (((((isnotnull(dt#10) && isnull(ip#8)) && isnull(dn#7)) && (dt#10 = 20201001)) && isnotnull(ip#8)) && isnotnull(dn#7))
+- Relation[dn#7,ip#8,first_time#9,dt#10] orc

2、条件在子查询中
select
/+ REPARTITION(10)/
oneday.dn, oneday.ip, '20201202', '20201202'
from
(
select
ip,dn
from dwd_dns.t_ip_dn_his_rel2
where dt = '20201202'
group by ip,dn
) oneday left join
(
select dn, ip
from
dwd_dns.t_ip_dn_first
where dt = '20201001'
) t on t.ip = oneday.ip and t.dn = oneday.dn
where t.ip is null and t.dn is null

执行计划：

== Optimized Logical Plan ==
InsertIntoHiveTable dwd_dns.t_ip_dn_first, org.apache.hadoop.hive.ql.io.orc.OrcSerde, Map(dt -> None), true, false, [dn, ip, first_time, dt]
+- Project [dn#1, ip#2, 20201202 AS first_time#28, 20201202 AS dt#29]
+- Repartition 10, true
+- Project [dn#1, ip#2]
+- Filter (isnull(ip#8) && isnull(dn#7))
+- Join LeftOuter, ((ip#8 = ip#2) && (dn#7 = dn#1))
:- Aggregate [ip#2, dn#1], [ip#2, dn#1]
: +- Project [dn#1, ip#2]
: +- Filter (isnotnull(dt#6) && (dt#6 = 20201202))
: +- Relation[uid#0,dn#1,ip#2,cname#3,dnsip#4,probe_time#5,dt#6] orc
+- Project [dn#7, ip#8]
+- Filter (((isnotnull(dt#10) && (dt#10 = 20201001)) && isnotnull(ip#8)) && isnotnull(dn#7))
+- Relation[dn#7,ip#8,first_time#9,dt#10] orc