OceanBase SQL 执行计划解读(三)── 标量子查询、分析函数

2024年 5月 7日 63.9k 0

前文《OceanBase SQL 执行计划解读(二)──── 表连接和子查询》介绍了子查询的执行计划特点,还没有完全说完。本文继续介绍子查询的执行计划以及分析函数(窗口函数)的执行计划特点。

熟悉常用 SQL 的执行计划是为了反过来快速解读分析复杂 SQL 的执行计划。

子查询

本文不讨论非相关子查询。

标量子查询表达式(Scalar Subquery Expression)是一类从一行返回一列值的子查询。标量子查询表达式的值是子查询的查询列的值。如果子查询返回 0 行,则标量子查询表达式的值是 NULL。如果子查询返回多行,则标量子查询表达式返回一个错误。

SUBPLAN FILTER 和 SCALAR GROUP BY

EXPLAIN extended_noaddr 
SELECT (SELECT w_name FROM BMSQL_WAREHOUSE w WHERE w.w_id = c.C_W_ID) ware_name  
    , c.C_D_ID ,c.C_FIRST ,c.C_LAST 
    , (SELECT count(*) FROM BMSQL_OORDER o WHERE o.O_C_ID =c.C_ID ) order_cnt
    , (SELECT sum(o.O_OL_CNT) FROM BMSQL_OORDER o WHERE o.O_C_ID =c.C_ID ) item_cnt
FROM BMSQL_CUSTOMER c
;

==================================================================
|ID|OPERATOR        |NAME                |EST. ROWS|COST         |
------------------------------------------------------------------
|0 |SUBPLAN FILTER  |                    |30000000 |2.236362e+12 |
|1 | TABLE SCAN     |C                   |30000000 |34332847     |
|2 | TABLE GET      |W                   |1        |36           |
|3 | SCALAR GROUP BY|                    |1        |5159         |
|4 |  TABLE SCAN    |O(BMSQL_OORDER_IDX4)|12106    |2847         |
|5 | SCALAR GROUP BY|                    |1        |69350        |
|6 |  TABLE SCAN    |O(BMSQL_OORDER_IDX4)|12106    |67037        |
==================================================================

Outputs & filters: 
-------------------------------------
  0 - output([subquery(1)], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [subquery(2)], [subquery(3)]), filter(nil), 
      exec_params_([C.C_W_ID], [C.C_ID], [C.C_ID]), onetime_exprs_(nil), init_plan_idxs_(nil)
  1 - output([C.C_W_ID], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [C.C_ID]), filter(nil), 
      access([C.C_W_ID], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [C.C_ID]), partitions(p0), 
      is_index_back=false, 
      range_key([C.C_W_ID], [C.C_D_ID], [C.C_ID]), range(MIN,MIN,MIN ; MAX,MAX,MAX)always true
  2 - output([W.W_NAME]), filter(nil), 
      access([W.W_NAME]), partitions(p0), 
      is_index_back=false, 
      range_key([W.W_ID]), range(MIN ; MAX)always true, 
      range_cond([W.W_ID = ?])
  3 - output([T_FUN_COUNT(*)]), filter(nil), 
      group(nil), agg_func([T_FUN_COUNT(*)])
  4 - output([1]), filter(nil), 
      access([O.O_C_ID]), partitions(p0), 
      is_index_back=false, 
      range_key([O.O_C_ID], [O.O_W_ID], [O.O_D_ID], [O.O_ID]), range(MIN,MIN,MIN,MIN ; MAX,MAX,MAX,MAX)always true, 
      range_cond([O.O_C_ID = ?])
  5 - output([T_FUN_SUM(O.O_OL_CNT)]), filter(nil), 
      group(nil), agg_func([T_FUN_SUM(O.O_OL_CNT)])
  6 - output([O.O_OL_CNT]), filter(nil), 
      access([O.O_OL_CNT]), partitions(p0), 
      is_index_back=true, 
      range_key([O.O_C_ID], [O.O_W_ID], [O.O_D_ID], [O.O_ID]), range(MIN,MIN,MIN,MIN ; MAX,MAX,MAX,MAX)always true, 
      range_cond([O.O_C_ID = ?])

SUBPLAN FILTER 用于驱动表达式中的子查询,OceanBase 会以 NESTED-LOOP 算法来执行 SUBPLAN FILTER 算子。即循环遍历左边的记录集,然后去右边结果集中取数据。所以,子查询是否能命中索引,对性能影响很大。

说明:

  • 标量子查询要求只返回一笔记录。可以直接取列,也可以用统计函数( count 、minmaxsum)。
  • 算子 0 是 SUBPLAN FILTER 。 output 表示输出列,后面包括 3 个子查询结果。filter表示算子上的过滤条件,这里是空(nil)。 exec_params_ 表示左表(结果集)传递给右表(结果集)的参数,一般关联子查询这里都是连接条件,如果是非关联子查询,这里就是空(nil)。onetime_exprs_表示只计算一次的对象(如子查询1),通常非关联的子查询结果集只需要计算一次。这里是关联子查询,所以值是空(nil)。
  • 算子 2 是第一个子查询,直接主键访问,用 TABLE GET .
  • 算子 3 和 4 是第二个子查询,扫描索引(TABLE SCAN),然后再聚合 。不过这里没有分组逻辑,所以 group 参数是空。

算子 SCALAR GROUP BY 是聚合函数生成标量结果常用的算法,用在没有 GROUP BY 语句的时候。当有GROUP BY语句时,使用的就是 HASH GROUP BY 或者 MERGE GROUP BY 算子。

EXPLAIN extended_noaddr
SELECT c.C_W_ID , count(*)
FROM BMSQL_CUSTOMER c
GROUP BY c.C_W_ID 
HAVING count(*) > 1000;

===========================================
|ID|OPERATOR      |NAME|EST. ROWS|COST    |
-------------------------------------------
|0 |MERGE GROUP BY|    |50       |41251449|
|1 | TABLE SCAN   |C   |30000000 |33009350|
===========================================

Outputs & filters: 
-------------------------------------
  0 - output([C.C_W_ID], [T_FUN_COUNT(*)]), filter([T_FUN_COUNT(*) > 1000]), 
      group([C.C_W_ID]), agg_func([T_FUN_COUNT(*)])
  1 - output([C.C_W_ID]), filter(nil), 
      access([C.C_W_ID]), partitions(p0), 
      is_index_back=false, 
      range_key([C.C_W_ID], [C.C_D_ID], [C.C_ID]), range(MIN,MIN,MIN ; MAX,MAX,MAX)always true

说明:

  • 如上,有明显的 GROUP BY子句,使用的是 MERGE GROUP BY算子, 分组表达式是 C.C_W_ID 。( group([C.C_W_ID]) )
  • HAVING 子句,会在算子 MERGE GROUP BY 上产生一个 filter 。

MERGE GROUP BY 和 HASH GROUP BY

有时候,没有 GROUP BY子句,也会用到 MERGE GROUP BY算子。

EXPLAIN extended_noaddr 
SELECT c.C_W_ID ,c.C_D_ID ,c.C_FIRST ,c.C_LAST ,c.C_BALANCE ,c.C_PAYMENT_CNT 
FROM BMSQL_CUSTOMER c 
WHERE (SELECT count(*) FROM BMSQL_OORDER o WHERE o.O_C_ID=c.C_ID) > 10;

==================================================================
|ID|OPERATOR             |NAME                |EST. ROWS|COST    |
------------------------------------------------------------------
|0 |HASH RIGHT OUTER JOIN|                    |10000000 |84866769|
|1 | SUBPLAN SCAN        |VIEW1               |3008     |18473351|
|2 |  MERGE GROUP BY     |                    |3008     |18472936|
|3 |   TABLE SCAN        |O(BMSQL_OORDER_IDX4)|36414597 |8468513 |
|4 | TABLE SCAN          |C                   |30000000 |35656345|
==================================================================

Outputs & filters: 
-------------------------------------
  0 - output([C.C_W_ID], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [C.C_BALANCE], [C.C_PAYMENT_CNT]), filter([CASE WHEN (T_OP_IS_NOT, VIEW1.O.O_C_ID, NULL, 0) THEN VIEW1.COUNT(*) ELSE 0 END > 10]), 
      equal_conds([VIEW1.O.O_C_ID = C.C_ID]), other_conds(nil)
  1 - output([VIEW1.COUNT(*)], [VIEW1.O.O_C_ID]), filter(nil), 
      access([VIEW1.COUNT(*)], [VIEW1.O.O_C_ID])
  2 - output([T_FUN_COUNT(*)], [O.O_C_ID]), filter(nil), 
      group([O.O_C_ID]), agg_func([T_FUN_COUNT(*)])
  3 - output([O.O_C_ID]), filter(nil), 
      access([O.O_C_ID]), partitions(p0), 
      is_index_back=false, 
      range_key([O.O_C_ID], [O.O_W_ID], [O.O_D_ID], [O.O_ID]), range(MIN,MIN,MIN,MIN ; MAX,MAX,MAX,MAX)always true
  4 - output([C.C_ID], [C.C_W_ID], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [C.C_BALANCE], [C.C_PAYMENT_CNT]), filter(nil), 
      access([C.C_ID], [C.C_W_ID], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [C.C_BALANCE], [C.C_PAYMENT_CNT]), partitions(p0), 
      is_index_back=false, 
      range_key([C.C_W_ID], [C.C_D_ID], [C.C_ID]), range(MIN,MIN,MIN ; MAX,MAX,MAX)always true

说明:

  • 没有 GROUP BY子句,但是针对 WHERE 条件中的关联子查询,优化器改写了算法为 HASH RIGHT OUTER JOIN ,事先将子查询结果分组统计出来 (按 o.o_c_id 做 GROUP BY),所以有算子 3 MERGE GROUP BY 。使用 MERGE 是利用了索引的有序性。
  • 算子 1 SUBPLAN SCAN 从子查询视图扫描数据。在算子 0 HASH RIGHT OUTER JOIN 使用 filter 应用子查询的过滤条件 (>10) .

如果子查询中结果集没有好的索引可以使用,优化器算法会调整为使用 HASH GROUP BY 。

EXPLAIN extended_noaddr 
SELECT c.C_W_ID ,c.C_D_ID ,c.C_FIRST ,c.C_LAST ,c.C_BALANCE ,c.C_PAYMENT_CNT 
FROM BMSQL_CUSTOMER c 
WHERE (SELECT count(*) FROM BMSQL_HISTORY h WHERE h.H_C_ID = c.C_ID) > 100;

===================================================
|ID|OPERATOR             |NAME |EST. ROWS|COST    |
---------------------------------------------------
|0 |HASH RIGHT OUTER JOIN|     |10000000 |88584899|
|1 | SUBPLAN SCAN        |VIEW1|3008     |22191481|
|2 |  HASH GROUP BY      |     |3008     |22191066|
|3 |   TABLE SCAN        |H    |36189654 |8391834 |
|4 | TABLE SCAN          |C    |30000000 |35656345|
===================================================

Outputs & filters: 
-------------------------------------
  0 - output([C.C_W_ID], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [C.C_BALANCE], [C.C_PAYMENT_CNT]), filter([CASE WHEN (T_OP_IS_NOT, VIEW1.H.H_C_ID, NULL, 0) THEN VIEW1.COUNT(*) ELSE 0 END > 100]), 
      equal_conds([VIEW1.H.H_C_ID = C.C_ID]), other_conds(nil)
  1 - output([VIEW1.COUNT(*)], [VIEW1.H.H_C_ID]), filter(nil), 
      access([VIEW1.COUNT(*)], [VIEW1.H.H_C_ID])
  2 - output([T_FUN_COUNT(*)], [H.H_C_ID]), filter(nil), 
      group([H.H_C_ID]), agg_func([T_FUN_COUNT(*)])
  3 - output([H.H_C_ID]), filter(nil), 
      access([H.H_C_ID]), partitions(p0), 
      is_index_back=false, 
      range_key([H.__pk_increment]), range(MIN ; MAX)always true
  4 - output([C.C_ID], [C.C_W_ID], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [C.C_BALANCE], [C.C_PAYMENT_CNT]), filter(nil), 
      access([C.C_ID], [C.C_W_ID], [C.C_D_ID], [C.C_FIRST], [C.C_LAST], [C.C_BALANCE], [C.C_PAYMENT_CNT]), partitions(p0), 
      is_index_back=false, 
      range_key([C.C_W_ID], [C.C_D_ID], [C.C_ID]), range(MIN,MIN,MIN ; MAX,MAX,MAX)always true

分析函数

分析函数(某些数据库下也叫做窗口函数)与聚集函数类似,计算总是基于一组行的集合,不同的是,聚集函数一组只能返回一行,而分析函数每组可以返回多行,组内每一行都是基于窗口的逻辑计算的结果。分析函数可以显著优化需要 self-join的查询。有些分析函数也可以当聚集函数使用。

分析函数包括:

  • MAX 、MIN 、AVG
  • COUNTSUM
  • GROUP_CONCAT 、 LISTAGG
  • ROW_NUMBER 、RANKDENSE_RANKPERCENT_RANK
  • CUME_DIST
  • FIRST_VALUELAST_VALUE
  • NTH_VALUENTILE
  • LEADLAG

算子 WINDOW_FUNCTION

如下面示例,统计各个仓库下的各个区的销量在本仓库内的排名。

EXPLAIN extended_noaddr 
SELECT d.D_W_ID , d.D_ID , d.D_NAME , d.D_YTD ,ROW_NUMBER () OVER (PARTITION BY d.D_W_ID ORDER BY d.D_YTD DESC ) rn 
FROM BMSQL_DISTRICT d
ORDER BY rn  ;
;

==========================================
|ID|OPERATOR        |NAME|EST. ROWS|COST |
------------------------------------------
|0 |SORT            |    |10000    |82278|
|1 | WINDOW FUNCTION|    |10000    |32980|
|2 |  SORT          |    |10000    |31070|
|3 |   TABLE SCAN   |D   |10000    |4022 |
==========================================

Outputs & filters: 
-------------------------------------
  0 - output([D.D_W_ID], [D.D_ID], [D.D_NAME], [D.D_YTD], [T_WIN_FUN_ROW_NUMBER()]), filter(nil), sort_keys([T_WIN_FUN_ROW_NUMBER(), ASC])
  1 - output([D.D_W_ID], [D.D_ID], [D.D_NAME], [D.D_YTD], [T_WIN_FUN_ROW_NUMBER()]), filter(nil), 
      win_expr(T_WIN_FUN_ROW_NUMBER()), partition_by([D.D_W_ID]), order_by([D.D_YTD, DESC]), window_type(RANGE), upper(UNBOUNDED PRECEDING), lower(UNBOUNDED FOLLOWING)
  2 - output([D.D_W_ID], [D.D_ID], [D.D_NAME], [D.D_YTD]), filter(nil), sort_keys([D.D_W_ID, ASC], [D.D_YTD, DESC]), prefix_pos(1)
  3 - output([D.D_W_ID], [D.D_ID], [D.D_NAME], [D.D_YTD]), filter(nil), 
      access([D.D_W_ID], [D.D_ID], [D.D_NAME], [D.D_YTD]), partitions(p0), 
      is_index_back=false, 
      range_key([D.D_W_ID], [D.D_ID]), range(MIN,MIN ; MAX,MAX)always true

说明:

  • 分析函数对应的算子是 WINDOW FUNCTION ,依赖下层算子的有序输出,有分区表达式和排序表达式。 
  • output 是算子的输出表达式,包含分析函数的结果, filter 固定为 nil
  • win_expr 表示在窗口中使用哪个窗口函数 ,partition_by 表示窗口内的分组表达式, order_by 表示窗口每组内部统计时的排序表达式 。window_type
  • window_type 表示窗口类型,有两种:range 和 rows 。range 表示按照逻辑位置偏移进行计算窗口上下界限,rows 表示按照实际物理位置偏移进行计算窗口上下界限;默认使用 range 方式。
  • upper 和 lower 分别定义窗口的上限和下限。UNBOUNDED 表示无边界,按照最大的选择(默认)。CURRENT ROW 表示从当前行开始,如果出现数字则表示移动的行数。PRECEDING 表示向前取边界,FOLLOWING 则表示向后取边界。
  • 算子 2 的 SORT 会包含分区窗口表达式和窗口内的排序表达式。
  • 算子 0 的 SORT 是最外层的排序表达式。

下面可以看看不同分析函数下执行计划里算子 WINDOW_FUNCTION 的各个参数。

EXPLAIN extended_noaddr
SELECT w.W_STATE, w_id, w_name, w_ytd,  max(w_ytd) over (partition by W_STATE  order by w_ytd desc rows between 1 preceding and 1 following) max_ytd_in_3
from BMSQL_WAREHOUSE w
ORDER BY w.W_STATE ;

==========================================
|ID|OPERATOR       |NAME|EST. ROWS|COST  |
------------------------------------------
|0 |WINDOW FUNCTION|    |1000     |148632|
|1 | SORT          |    |1000     |148441|
|2 |  TABLE SCAN   |W   |1000     |141926|
==========================================

Outputs & filters: 
-------------------------------------
  0 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD], [T_FUN_MAX(W.W_YTD)]), filter(nil), 
      win_expr(T_FUN_MAX(W.W_YTD)), partition_by([W.W_STATE]), order_by([W.W_YTD, DESC]), window_type(ROWS), upper(1 PRECEDING), lower(1 FOLLOWING)
  1 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD]), filter(nil), sort_keys([W.W_STATE, ASC], [W.W_YTD, DESC])
  2 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD]), filter(nil), 
      access([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD]), partitions(p0), 
      is_index_back=false, 
      range_key([W.W_ID]), range(MIN ; MAX)always true

说明:

  • window_type 是 ROWS ,upper 是同一个分组内向前一笔, lower是同一个分组内向后一笔。如果同一个分组内没有向前或向后一笔,那就是空。

下面输出各个仓库的销量以及包括前2笔在内的最大销量。

EXPLAIN extended_noaddr
SELECT w.W_STATE, w_id, w_name, w_ytd,  max(w_ytd) over (order by w_ytd desc rows between 2 PRECEDING AND current row  ) max_ytd_in_3
from BMSQL_WAREHOUSE w
ORDER BY w.W_YTD  DESC ;

==========================================
|ID|OPERATOR       |NAME|EST. ROWS|COST  |
------------------------------------------
|0 |WINDOW FUNCTION|    |1000     |288860|
|1 | SORT          |    |1000     |288669|
|2 |  TABLE SCAN   |W   |1000     |285169|
==========================================

Outputs & filters: 
-------------------------------------
  0 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD], [T_FUN_MAX(W.W_YTD)]), filter(nil), 
      win_expr(T_FUN_MAX(W.W_YTD)), partition_by(nil), order_by([W.W_YTD, DESC]), window_type(ROWS), upper(2 PRECEDING), lower(CURRENT ROW)
  1 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD]), filter(nil), sort_keys([W.W_YTD, DESC])
  2 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD]), filter(nil), 
      access([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD]), partitions(p0), 
      is_index_back=false, 
      range_key([W.W_ID]), range(MIN ; MAX)always true

说明:

  • 分区表达式 partition_by 并不是必须的,可以为空。

下面看看行转列函数的执行计划。

EXPLAIN extended_Noaddr
SELECT d.D_W_ID , d.D_ID ,d.d_name, d.D_YTD , listagg(d.D_NAME,',') WITHIN GROUP (ORDER BY d.D_YTD DESC ) OVER (PARTITION BY d.D_W_ID) d_names
FROM  BMSQL_DISTRICT  d 
WHERE d.D_W_ID = 10
;

========================================
|ID|OPERATOR       |NAME|EST. ROWS|COST|
----------------------------------------
|0 |WINDOW FUNCTION|    |10       |40  |
|1 | TABLE SCAN    |D   |10       |39  |
========================================

Outputs & filters: 
-------------------------------------
  0 - output([D.D_W_ID], [D.D_ID], [D.D_NAME], [D.D_YTD], [T_FUN_GROUP_CONCAT(D.D_NAME, ',') order_items(D.D_YTD)]), filter(nil), 
      win_expr(T_FUN_GROUP_CONCAT(D.D_NAME, ',') order_items(D.D_YTD)), partition_by([D.D_W_ID]), order_by(nil), window_type(RANGE), upper(UNBOUNDED PRECEDING), lower(UNBOUNDED FOLLOWING)
  1 - output([D.D_W_ID], [D.D_ID], [D.D_NAME], [D.D_YTD]), filter(nil), 
      access([D.D_W_ID], [D.D_ID], [D.D_NAME], [D.D_YTD]), partitions(p0), 
      is_index_back=false, 
      range_key([D.D_W_ID], [D.D_ID]), range(10,MIN ; 10,MAX), 
      range_cond([D.D_W_ID = 10])

说明:

  • 使用 LISTAGG 语法时,窗口函数表达式是 T_FUN_GROUP_CONCAT 。

再看一个 窗口类型为 RANGE 的示例。 

EXPLAIN extended_noaddr
SELECT w.W_STATE, w_id, w_name, w_ytd,  max(w_ytd) over (order by w_ytd desc range between 1000000 PRECEDING AND current row  ) max_ytd_in_3
from BMSQL_WAREHOUSE w
ORDER BY w.W_YTD  DESC ;

==========================================
|ID|OPERATOR       |NAME|EST. ROWS|COST  |
------------------------------------------
|0 |WINDOW FUNCTION|    |1000     |191810|
|1 | SORT          |    |1000     |191619|
|2 |  TABLE SCAN   |W   |1000     |188120|
==========================================

Outputs & filters: 
-------------------------------------
  0 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD], [T_FUN_MAX(W.W_YTD)]), filter(nil), 
      win_expr(T_FUN_MAX(W.W_YTD)), partition_by(nil), order_by([W.W_YTD, DESC]), window_type(RANGE), upper(1000000 PRECEDING), lower(CURRENT ROW)
  1 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD], [W.W_YTD + 1000000], [W.W_YTD - 1000000]), filter(nil), sort_keys([W.W_YTD, DESC])
  2 - output([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD]), filter(nil), 
      access([W.W_STATE], [W.W_ID], [W.W_NAME], [W.W_YTD]), partitions(p0), 
      is_index_back=false, 
      range_key([W.W_ID]), range(MIN ; MAX)always true


说明:

  • window_type 是 RANGE 。是按实际值计算窗口大小,不是按行数固定窗口大小。算子 1 的output 里多了两列 ( [W.W_YTD + 1000000], [W.W_YTD - 1000000] )。

分析函数的代价

分析函数看起来很酷,不过也有代价,那就是每次调用分析函数都可能会有一次排序,排序需要内存,可能需要增大内部参数 _sort_area_size 的值。为了性能还建议使用并行( OB 的并行会在下篇文章介绍)。

下面这个示例会涉及到一次全表扫描和三次排序。

EXPLAIN extended_noaddr
SELECT  c_w_id, c_d_id, c_id, c.C_LAST ,c.C_FIRST , C_YTD_PAYMENT ,
    rank() OVER (PARTITION BY C_W_ID, c_d_id ORDER BY C_YTD_PAYMENT DESC  ) rank_in_district,
    rank() OVER (PARTITION BY c_w_id ORDER BY C_YTD_PAYMENT DESC ) rank_in_warehouse,
    rank() OVER (ORDER BY C_YTD_PAYMENT DESC ) rank_in_all
FROM BMSQL_CUSTOMER c
WHERE c.C_YTD_PAYMENT >= 1000000
ORDER BY c.C_YTD_PAYMENT DESC ;
;

=================================================
|ID|OPERATOR           |NAME|EST. ROWS|COST     |
-------------------------------------------------
|0 |WINDOW FUNCTION    |    |15488448 |314812916|
|1 | SORT              |    |15488448 |311854430|
|2 |  WINDOW FUNCTION  |    |15488448 |185456913|
|3 |   SORT            |    |15488448 |182498427|
|4 |    WINDOW FUNCTION|    |15488448 |118046685|
|5 |     SORT          |    |15488448 |115088200|
|6 |      TABLE SCAN   |C   |15488448 |50636458 |
=================================================

Outputs & filters: 
-------------------------------------
  0 - output([C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST], [C.C_YTD_PAYMENT], [T_WIN_FUN_RANK()], [T_WIN_FUN_RANK()], [T_WIN_FUN_RANK()]), filter(nil), 
      win_expr(T_WIN_FUN_RANK()), partition_by(nil), order_by([C.C_YTD_PAYMENT, DESC]), window_type(RANGE), upper(UNBOUNDED PRECEDING), lower(UNBOUNDED FOLLOWING)
  1 - output([C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST], [C.C_YTD_PAYMENT], [T_WIN_FUN_RANK()], [T_WIN_FUN_RANK()]), filter(nil), sort_keys([C.C_YTD_PAYMENT, DESC])
  2 - output([C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST], [C.C_YTD_PAYMENT], [T_WIN_FUN_RANK()], [T_WIN_FUN_RANK()]), filter(nil), 
      win_expr(T_WIN_FUN_RANK()), partition_by([C.C_W_ID]), order_by([C.C_YTD_PAYMENT, DESC]), window_type(RANGE), upper(UNBOUNDED PRECEDING), lower(UNBOUNDED FOLLOWING)
  3 - output([C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST], [C.C_YTD_PAYMENT], [T_WIN_FUN_RANK()]), filter(nil), sort_keys([C.C_W_ID, ASC], [C.C_YTD_PAYMENT, DESC]), prefix_pos(1)
  4 - output([C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST], [C.C_YTD_PAYMENT], [T_WIN_FUN_RANK()]), filter(nil), 
      win_expr(T_WIN_FUN_RANK()), partition_by([C.C_W_ID], [C.C_D_ID]), order_by([C.C_YTD_PAYMENT, DESC]), window_type(RANGE), upper(UNBOUNDED PRECEDING), lower(UNBOUNDED FOLLOWING)
  5 - output([C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST], [C.C_YTD_PAYMENT]), filter(nil), sort_keys([C.C_W_ID, ASC], [C.C_D_ID, ASC], [C.C_YTD_PAYMENT, DESC]), prefix_pos(2)
  6 - output([C.C_YTD_PAYMENT], [C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST]), filter([C.C_YTD_PAYMENT >= 1000000]), 
      access([C.C_YTD_PAYMENT], [C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST]), partitions(p0), 
      is_index_back=false, filter_before_indexback[false], 
      range_key([C.C_W_ID], [C.C_D_ID], [C.C_ID]), range(MIN,MIN,MIN ; MAX,MAX,MAX)always true

执行结果如下图:

图片

OceanBase SQL 执行计划解读(三)── 标量子查询、分析函数-1

说明:

  • 执行的顺序,首先算子 6 是针对表的扫描,先执行过滤条件(filter)。
  • 算子 5 是第一次排序,排序列是分区列加上排序列(sort_keys([C.C_W_ID, ASC], [C.C_D_ID, ASC], [C.C_YTD_PAYMENT, DESC]))。
  • 算子 4 是第一个窗口函数,分区列是 partition_by([C.C_W_ID], [C.C_D_ID]) , 排序列是 order_by([C.C_YTD_PAYMENT, DESC])
  • 算子 3 是一个优化,利用了第一个窗口函数的结果继续进行排序。
  • 算子 1 在算子 2 结果集基础上进一步排序。

上面示例 3 个分析函数使用的窗口函数算子都是 T_WIN_FUN_RANK,只是分区列不同所以还是有三次排序。如果分区列和排序列一样的话,是可以规避多次排序的。如下面示例。

EXPLAIN extended_noaddr
SELECT * FROM (
    SELECT c_w_id, c_d_id, c_id, c.C_LAST ,c.C_FIRST , C_YTD_PAYMENT 
        ,rank() OVER (PARTITION BY c_w_id,c_d_id ORDER BY C_YTD_PAYMENT) ytd_rank
        ,first_value(C_YTD_PAYMENT) OVER (PARTITION BY c_w_id,c_d_id ORDER BY C_YTD_PAYMENT) first_ytd
--      ,last_value(C_YTD_PAYMENT) OVER (PARTITION BY C_W_ID,c_d_id ORDER BY C_YTD_PAYMENT ) last_ytd
        ,last_value(C_YTD_PAYMENT) OVER (PARTITION BY C_W_ID,c_d_id ORDER BY C_YTD_PAYMENT rows between unbounded preceding and unbounded following) last_ytd_all
    FROM BMSQL_CUSTOMER  c
    WHERE c.C_YTD_PAYMENT >= 1000000
 ) t 
WHERE t.c_w_id = 23
;

===========================================
|ID|OPERATOR        |NAME|EST. ROWS|COST  |
-------------------------------------------
|0 |SUBPLAN SCAN    |T   |1533     |147862|
|1 | WINDOW FUNCTION|    |15477    |144430|
|2 |  SORT          |    |15477    |141474|
|3 |   TABLE SCAN   |C   |15477    |41003 |
===========================================

Outputs & filters: 
-------------------------------------
  0 - output([T.C_W_ID], [T.C_D_ID], [T.C_ID], [T.C_LAST], [T.C_FIRST], [T.C_YTD_PAYMENT], [T.YTD_RANK], [T.FIRST_YTD], [T.LAST_YTD_ALL]), filter([T.C_D_ID = 10]), 
      access([T.C_W_ID], [T.C_D_ID], [T.C_ID], [T.C_LAST], [T.C_FIRST], [T.C_YTD_PAYMENT], [T.YTD_RANK], [T.FIRST_YTD], [T.LAST_YTD_ALL])
  1 - output([C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST], [C.C_YTD_PAYMENT], [T_WIN_FUN_RANK()], [T_WIN_FUN_NTH_VALUE(C.C_YTD_PAYMENT,1)], [T_WIN_FUN_NTH_VALUE(C.C_YTD_PAYMENT,1)]), filter(nil), 
      win_expr(T_WIN_FUN_RANK()), partition_by([C.C_W_ID]), order_by([C.C_YTD_PAYMENT, ASC]), window_type(RANGE), upper(UNBOUNDED PRECEDING), lower(UNBOUNDED FOLLOWING)
      win_expr(T_WIN_FUN_NTH_VALUE(C.C_YTD_PAYMENT,1)), partition_by([C.C_W_ID]), order_by([C.C_YTD_PAYMENT, ASC]), window_type(RANGE), upper(UNBOUNDED PRECEDING), lower(CURRENT ROW)
      win_expr(T_WIN_FUN_NTH_VALUE(C.C_YTD_PAYMENT,1)), partition_by([C.C_W_ID]), order_by([C.C_YTD_PAYMENT, ASC]), window_type(ROWS), upper(UNBOUNDED PRECEDING), lower(UNBOUNDED FOLLOWING)
  2 - output([C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST], [C.C_YTD_PAYMENT]), filter(nil), sort_keys([C.C_YTD_PAYMENT, ASC])
  3 - output([C.C_YTD_PAYMENT], [C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST]), filter([C.C_YTD_PAYMENT >= 1000000]), 
      access([C.C_YTD_PAYMENT], [C.C_W_ID], [C.C_D_ID], [C.C_ID], [C.C_LAST], [C.C_FIRST]), partitions(p0), 
      is_index_back=false, filter_before_indexback[false], 
      range_key([C.C_W_ID], [C.C_D_ID], [C.C_ID]), range(23,MIN,MIN ; 23,MAX,MAX), 
      range_cond([C.C_W_ID = 23])

OceanBase SQL 执行计划解读(三)── 标量子查询、分析函数-2

说明:

  • 这里 where 条件有两个(t.c_w_id = 23 AND t.c_d_id = 10),但实际条件下推到子查询里只有 t.c_w_id=23 。 这是因为子查询使用的分析函数里的分区列只包含列t.c_w_id 。看算子 3 的 range_cond 和算子 1的 win_expr 。 最后一部的 filter 才是条件 t.c_d_id = 10
  • 算子 2 的 sort_keys([C.C_YTD_PAYMENT, ASC]) 只包含了列 c.c_ytd_payment 这是因为算子 3 返回的数据在列 c_w_id 上已经是有序的。
  • 算子 1 的一共用了 3 个窗口函数表达式(win_expr),分别是:T_WIN_FUN_RANK()T_WIN_FUN_NTH_VALUE(C.C_YTD_PAYMENT,1)T_WIN_FUN_NTH_VALUE(C.C_YTD_PAYMENT,1),它们的 order_by 条件都是一样的,只是窗口的下限不一样,可以共用一个 SORT 操作。
  • 从这个例子还看出,默认的窗口范围是 upper(UNBOUNDED PRECEDING), lower(CURRENT ROW) 。被注释掉的 last_value 使用默认的窗口范围,只会返回当前行值。

参考

   更多执行计划总结,请参考前面文章。下一篇会重点介绍 OB 特有的并行和分布式执行计划。

相关文章

Oracle如何使用授予和撤销权限的语法和示例
Awesome Project: 探索 MatrixOrigin 云原生分布式数据库
下载丨66页PDF,云和恩墨技术通讯(2024年7月刊)
社区版oceanbase安装
Oracle 导出CSV工具-sqluldr2
ETL数据集成丨快速将MySQL数据迁移至Doris数据库

发布评论