使用Hive分区表优化联接性能

时间:2019-07-08 16:43:56

标签: performance hive query-optimization partitioning

我有一个带有某些示例数据的Hive orc test_dev_db.TransactionUpdateTable表,该表将保存需要更新到主表(test_dev_db.TransactionMainHistoryTable)的增量数据,该主表已划分为Country,Tran_date列。

Hive增量负载表架构:它包含19行需要合并。

CREATE TABLE IF NOT EXISTS test_dev_db.TransactionUpdateTable 
(
Transaction_date timestamp,
Product       string,
Price         int,
Payment_Type  string,
Name          string, 
City          string,
State         string,
Country       string
)
ROW FORMAT DELIMITED
FIELDS TERMINATED BY ','
STORED AS orc
;

Hive主表模式:总行数为77。

CREATE TABLE IF NOT EXISTS test_dev_db.TransactionMainHistoryTable
(
Transaction_date timestamp,
Product       string,
Price         int,
Payment_Type  string,
Name          string,
City          string,
State         string
)
PARTITIONED BY (Country string,Tran_date string) 
ROW FORMAT DELIMITED
FIELDS TERMINATED BY ','
STORED AS orc
;

我正在查询下面运行,以将增量数据与主表合并。

SELECT
  case when i.transaction_date is not null then cast(substring(current_timestamp(),0,19) as timestamp)  
  else t.transaction_date   end as transaction_date,
  t.product,
  case when i.price is not null then i.price else t.price end as price,
  t.payment_type,
  t.name,
  t.city,
  t.state,
  t.country,
  case when i.transaction_date is not null then substring(current_timestamp(),0,10) 
  else t.tran_date end as tran_date
  from
test_dev_db.TransactionMainHistoryTable t
full join test_dev_db.TransactionUpdateTable i on (t.Name=i.Name)
;
/hdfs/path/database/test_dev_db.db/transactionmainhistorytable/country=Australia/tran_date=2009-03-01
/hdfs/path/database/test_dev_db.db/transactionmainhistorytable/country=Australia/tran_date=2009-05-01

并在下面的查询中运行以过滤出需要合并的特定分区,以免重写未更新的分区。

SELECT
  case when i.transaction_date is not null then cast(substring(current_timestamp(),0,19) as timestamp)  
  else t.transaction_date   end as transaction_date,
  t.product,
  case when i.price is not null then i.price else t.price end as price,
  t.payment_type,
  t.name,
  t.city,
  t.state,
  t.country,
  case when i.transaction_date is not null then substring(current_timestamp(),0,10) else t.tran_date end as tran_date
  from
(SELECT 
  *
  FROM 
test_dev_db.TransactionMainHistoryTable
where Tran_date in
(select distinct  from_unixtime(to_unix_timestamp (Transaction_date,'yyyy-MM-dd HH:mm'),'yyyy-MM-dd') from test_dev_db.TransactionUpdateTable
))t
full join test_dev_db.TransactionUpdateTable i on (t.Name=i.Name)
;
在两种情况下,仅

仅需更新Transaction_date,Price和分区列tran_date。尽管横向查询需要更长的时间才能执行,但这两个查询都运行良好。

分区表的执行计划为:

 Stage: Stage-5
    Map Reduce
      Map Operator Tree:
          TableScan
            alias: transactionmainhistorytable
            filterExpr: tran_date is not null (type: boolean)
            Statistics: Num rows: 77 Data size: 39151 Basic stats: COMPLETE Column stats: COMPLETE
            Map Join Operator
              condition map:
                   Left Semi Join 0 to 1
              keys:
                0 tran_date (type: string)
                1 _col0 (type: string)
              outputColumnNames: _col0, _col1, _col2, _col3, _col4, _col5, _col6, _col7, _col8

我在第二次查询中做错什么了吗?是否需要同时使用两个partition列才能进行更好的修剪。 任何帮助或建议,我们将不胜感激。

1 个答案:

答案 0 :(得分:1)

也许这不是一个完整的答案,但我希望这些想法会有所帮助。

where tran_date IN (select ... )

实际上与

相同
LEFT SEMI JOIN (SELECT ...)

这反映在计划中:

Map Join Operator
              condition map:
                   Left Semi Join 0 to 1
              keys:
                0 tran_date (type: string)
                1 _col0 (type: string) 

它作为map-join执行。首先,选择子查询数据集,然后将其放置在分布式缓存中,并加载到要在map-join中使用的内存中。所有这些步骤:选择,加载到内存,映射联接比读取和覆盖所有表都要慢,因为它是如此之小且过度分区:统计数据显示行数:77数据大小:39151-太小,无法被两个分区列,甚至太小而根本无法分区。尝试使用更大的表,并使用EXPLAIN EXTENDED检查真正被扫描的内容。

此外,替换为:

from_unixtime(to_unix_timestamp (Transaction_date,'yyyy-MM-dd HH:mm'),'yyyy-MM-dd')

带有substr(Transaction_date,0,10)date(Transaction_date)

使用substring(current_timestamp,0,10)current_date只是为了简化代码。

如果要在计划中显示分区过滤器,请尝试替换作为分区列表传递的分区过滤器,您可以在单独的会话中选择该分区过滤器,并使用shell将分区列表传递给where子句,请参见以下答案: https://stackoverflow.com/a/56963448/2700344