我正在查询一个大表,该表分区在一个名为day
的字段上。
如果我运行查询:select *
from my_table
where day in ('2016-04-01', '2016-03-01')
我有许多映射器和化简器,查询需要很长时间才能运行。
但是,如果我写一个查询:select *
from my_table
where day = '2016-04-01'
or day = '2016-03-01'
我得到的映射器和缩减器少得多,查询运行很快。 对我来说,这表明in
没有利用表中的分区。谁能证实这一点并解释原因?
hive 版本:1.2.1
Hadoop版本:2.3.4.7-4
细节:
我相信执行计划的相关部分是...
使用Where or
完全没有过滤器运算符
使用Where in
Filter Operator
predicate: (day) IN ('2016-04-01', '2016-03-01') (type: boolean)
Statistics: Num rows: 100000000 Data size: 9999999999
The hive docs只是说:
“查询中要使用的分区由系统根据分区列上的where子句条件自动确定。”
但不要详细说明。我找不到任何与此直接相关的SO帖子。
谢谢!
最佳答案
tl; dr
根据我在Hue中运行的解释计划,我将 Hive 1.1.0 与一起使用Cloudera 5.13.3 和IN
遵循与等于运算符(=
)相同的优化。
例子
我的表在LOAD_YEAR (SMALLINT)
和LOAD_MONTH (TINYINT)
上进行了分区,并具有以下两个分区:
load_year=2018/load_month=10
(19,828,71行)load_year=2018/load_month=11
(702,856行)以下是各种查询及其说明计划。
1.等于(
=
)运算符询问:
SELECT ID
FROM TBL
WHERE LOAD_MONTH = 11Y
说明计划:
STAGE DEPENDENCIES:
Stage-0 is a root stage
STAGE PLANS:
Stage: Stage-0
Fetch Operator
limit: -1
Processor Tree:
TableScan
alias: tbl
filterExpr: (load_month = 11) (type: boolean)
Statistics: Num rows: 702856 Data size: 84342720 Basic stats: COMPLETE Column stats: NONE
Select Operator
expressions: id (type: string)
outputColumnNames: _col0
Statistics: Num rows: 702856 Data size: 84342720 Basic stats: COMPLETE Column stats: NONE
ListSink
2.
IN
运算符查询(请注意,数据中没有第12个月):
SELECT ID
FROM TBL
WHERE LOAD_MONTH IN (11Y, 12Y)
说明计划:
STAGE DEPENDENCIES:
Stage-0 is a root stage
STAGE PLANS:
Stage: Stage-0
Fetch Operator
limit: -1
Processor Tree:
TableScan
alias: tbl
filterExpr: (load_month = 11) (type: boolean)
Statistics: Num rows: 702856 Data size: 84342720 Basic stats: COMPLETE Column stats: NONE
Select Operator
expressions: id (type: string)
outputColumnNames: _col0
Statistics: Num rows: 702856 Data size: 84342720 Basic stats: COMPLETE Column stats: NONE
ListSink
3.与AND和OR一起等于(
=
)询问:
SELECT ID
FROM TBL
WHERE
(LOAD_YEAR = 2018S AND LOAD_MONTH = 11Y)
OR (LOAD_YEAR = 2019S AND LOAD_MONTH = 1Y)
说明计划:
STAGE DEPENDENCIES:
Stage-0 is a root stage
STAGE PLANS:
Stage: Stage-0
Fetch Operator
limit: -1
Processor Tree:
TableScan
alias: tbl
filterExpr: (((load_year = 2018) and (load_month = 11)) or ((load_year = 2019) and (load_month = 1))) (type: boolean)
Statistics: Num rows: 702856 Data size: 84342720 Basic stats: COMPLETE Column stats: NONE
Select Operator
expressions: id (type: string)
outputColumnNames: _col0
Statistics: Num rows: 702856 Data size: 84342720 Basic stats: COMPLETE Column stats: NONE
ListSink
4.算术运算
询问:
SELECT ID
FROM TBL
WHERE (LOAD_YEAR * 100 + LOAD_MONTH) IN (201811, 201901)
边注:
100没有后缀,所以它是
INT
,而(LOAD_YEAR * 100 + LOAD_MONTH)
也是INT
。这样可以确保结果准确。由于LOAD_YEAR
是SMALLINT
,而LOAD_MONTH
是TINYINT
,因此对两者的算术计算将SMALLINT
用于结果,并且存储的最大值为32,767(对于yyyymm
来说,这是不够的,它需要6位数字,即,至少为999,999)。以INT
为100时,将使用INT
类型进行计算,并允许数字最大为2147483647。说明计划:
STAGE DEPENDENCIES:
Stage-0 is a root stage
STAGE PLANS:
Stage: Stage-0
Fetch Operator
limit: -1
Processor Tree:
TableScan
alias: tbl
filterExpr: (201811) IN (201811, 201901) (type: boolean)
Statistics: Num rows: 702856 Data size: 84342720 Basic stats: COMPLETE Column stats: NONE
Select Operator
expressions: id (type: string)
outputColumnNames: _col0
Statistics: Num rows: 702856 Data size: 84342720 Basic stats: COMPLETE Column stats: NONE
ListSink
摘要
所有这些查询仅扫描第二个分区,从而避免了另一个分区中的〜2000万行。
关于 hive :+中不使用分区?,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/36896251/