Spark2.1和2.2 SQL物理执行策略关键源码分析

1. 文章开始之前

先附上一句SQL，使用tpc-ds的表结构，我们围绕这句SQL讲。

• SQL：

SQL> select
avg(cs_ext_discount_amt)
from
catalog_sales, date_dim
where
d_date between '1999-02-22'
and
cast('1999-05-22' as date)
and
d_date_sk = cs_sold_date_sk
group by cs_sold_date_sk;

• 逻辑计划：

Aggregate [cs_sold_date_sk#24], [cast((avg(UnscaledValue(cs_ext_discount_amt#46)) / 100.0) as decimal(11,6)) AS avg(cs_ext_discount_amt)#149]
+- Project [cs_sold_date_sk#24, cs_ext_discount_amt#46]
   +- Join Inner, (d_date_sk#58 = cs_sold_date_sk#24)
      :- Project [cs_sold_date_sk#24, cs_ext_discount_amt#46]
      :  +- Filter isnotnull(cs_sold_date_sk#24)
      :     +- Relation[cs_sold_date_sk#24,cs_sold_time_sk#25,cs_ship_date_sk#26,cs_bill_customer_sk#27,cs_bill_cdemo_sk#28,cs_bill_hdemo_sk#29,cs_bill_addr_sk#30,cs_ship_customer_sk#31,cs_ship_cdemo_sk#32,cs_ship_hdemo_sk#33,cs_ship_addr_sk#34,cs_call_center_sk#35,cs_catalog_page_sk#36,cs_ship_mode_sk#37,cs_warehouse_sk#38,cs_item_sk#39,cs_promo_sk#40,cs_order_number#41,cs_quantity#42,cs_wholesale_cost#43,cs_list_price#44,cs_sales_price#45,cs_ext_discount_amt#46,cs_ext_sales_price#47,... 10 more fields]
      +- Project [d_date_sk#58]
         +- Filter (((isnotnull(d_date#60) && (cast(d_date#60 as string) >= 1999-02-22)) && (d_date#60 <= 10733)) && isnotnull(d_date_sk#58))
            +- Relation[d_date_sk#58,d_date_id#59,d_date#60,d_month_seq#61,d_week_seq#62,d_quarter_seq#63,d_year#64,d_dow#65,d_moy#66,d_dom#67,d_qoy#68,d_fy_year#69,d_fy_quarter_seq#70,d_fy_week_seq#71,d_day_name#72,d_quarter_name#73,d_holiday#74,d_weekend#75,d_following_holiday#76,d_first_dom#77,d_last_dom#78,d_same_day_ly#79,d_same_day_lq#80,d_current_day#81,... 4 more fields]

2. 物理计划源码分析

2.1 物理策略

def strategies: Seq[Strategy] =
      extraStrategies ++ (
      FileSourceStrategy ::
      DataSourceStrategy ::
      DDLStrategy ::
      SpecialLimits ::
      Aggregation ::
      JoinSelection ::
      InMemoryScans ::
      BasicOperators :: Nil)

其中，extraStrategies是提供给外部人员可以自己添加的策略。调用这些strategies的代码如下：

// Collect physical plan candidates.
val candidates = strategies.iterator.flatMap(_(plan))

将strategies逐个去应用在逻辑计划上，然后做flat操作，返回一个PhysicalPlan的iterator。那么每个策略什么作用？

2.1.1 FileSourceStrategy

一个针对Hadoop文件系统做的策略，当执行计划的底层Relation是HadoopFsRelation时会调用到，用来扫描文件。

2.1.2 DataSourceStrategy

Spark针对DataSource预定义了四种scan接口，TableScan、PrunedScan、PrunedFilteredScan、CatalystScan(其中Catalys