分布式执行计划的简单调度模型为，在计划生成的最后阶段，以 exchange 节点为界，拆分成多个子计划，每个子计划被封装成为一个 DFO, 在并行度大于 1 的场景下, 会一次调度两个 DFO, 依次完成DFO树的遍历执行； 在并行度等于 1 的场景下, 每个 DFO 会将产生的数据存入中间结果管理器, 按照后序遍历的形式完成整个 DFO 树的遍历执行。

单 DFO 调度示例

在并行度为 1 的场景下, 对于如下查询计划执行单 DFO 调度。

  1. ======================================================================================
  2. |ID|OPERATOR                             |NAME                 |EST. ROWS |COST      |
  3. --------------------------------------------------------------------------------------
  4. |0 |LIMIT                                |                     |10        |6956829987|
  5. |1 | PX COORDINATOR MERGE SORT           |                     |10        |6956829985|
  6. |2 |  EXCHANGE OUT DISTR                 |:EX10002             |10        |6956829976|
  7. |3 |   LIMIT                             |                     |10        |6956829976|
  8. |4 |    TOP-N SORT                       |                     |10        |6956829975|
  9. |5 |     HASH GROUP BY                   |                     |454381562 |5815592885|
  10. |6 |      HASH JOIN                      |                     |500918979 |5299414557|
  11. |7 |       EXCHANGE IN DISTR             |                     |225943610 |2081426759|
  12. |8 |        EXCHANGE OUT DISTR (PKEY)    |:EX10001             |225943610 |1958446695|
  13. |9 |         MATERIAL                    |                     |225943610 |1958446695|
  14. |10|          HASH JOIN                  |                     |225943610 |1480989849|
  15. |11|           JOIN FILTER CREATE        |                     |30142669  |122441311 |
  16. |12|            PX PARTITION ITERATOR    |                     |30142669  |122441311 |
  17. |13|             TABLE SCAN              |CUSTOMER             |30142669  |122441311 |
  18. |14|           EXCHANGE IN DISTR         |                     |731011898 |900388059 |
  19. |15|            EXCHANGE OUT DISTR (PKEY)|:EX10000             |731011898 |614947815 |
  20. |16|             JOIN FILTER USE         |                     |731011898 |614947815 |
  21. |17|              PX BLOCK ITERATOR      |                     |731011898 |614947815 |
  22. |18|               TABLE SCAN            |ORDERS               |731011898 |614947815 |
  23. |19|       PX PARTITION ITERATOR         |                     |3243094528|1040696710|
  24. |20|        TABLE SCAN                   |LINEITEM(I_L_Q06_001)|3243094528|1040696710|
  25. ======================================================================================

如下图所示， DFO 树除 ROOT DFO 外, 在垂直方向上被分别划分为 0 , 1 , 2 个 DFO, 从而后序遍历调度的顺序为 0->1->2, 即可完成整个计划树的迭代。

单 DFO.jpeg

两 DFO 调度示例

对于并行度大于 1 的计划, 调度方式会采用两 DFO 调度。对于如下查询计划执行两 DFO 调度：

  1. Query Plan
  2. =============================================================================
  3. |ID|OPERATOR                                   |NAME    |EST. ROWS|COST     |
  4. -----------------------------------------------------------------------------
  5. |0 |PX COORDINATOR MERGE SORT                  |        |9873917  |692436562|
  6. |1 | EXCHANGE OUT DISTR                        |:EX10002|9873917  |689632565|
  7. |2 |  SORT                                     |        |9873917  |689632565|
  8. |3 |   SUBPLAN SCAN                            |VIEW5   |9873917  |636493382|
  9. |4 |    WINDOW FUNCTION                        |        |29621749 |629924873|
  10. |5 |     HASH GROUP BY                         |        |29621749 |624266752|
  11. |6 |      HASH JOIN                            |        |31521003 |591048941|
  12. |7 |       JOIN FILTER CREATE                  |        |407573   |7476793  |
  13. |8 |        EXCHANGE IN DISTR                  |        |407573   |7476793  |
  14. |9 |         EXCHANGE OUT DISTR (BROADCAST)    |:EX10001|407573   |7303180  |
  15. |10|          HASH JOIN                        |        |407573   |7303180  |
  16. |11|           JOIN FILTER CREATE              |        |1        |53       |
  17. |12|            EXCHANGE IN DISTR              |        |1        |53       |
  18. |13|             EXCHANGE OUT DISTR (BROADCAST)|:EX10000|1        |53       |
  19. |14|              PX BLOCK ITERATOR            |        |1        |53       |
  20. |15|               TABLE SCAN                  |NATION  |1        |53       |
  21. |16|           JOIN FILTER USE                 |        |10189312 |3417602  |
  22. |17|            PX BLOCK ITERATOR              |        |10189312 |3417602  |
  23. |18|             TABLE SCAN                    |SUPPLIER|10189312 |3417602  |
  24. |19|       JOIN FILTER USE                     |        |803481600|276540086|
  25. |20|        PX PARTITION ITERATOR              |        |803481600|276540086|
  26. |21|         TABLE SCAN                        |PARTSUPP|803481600|276540086|
  27. =============================================================================

如下图所示，DFO 树除 ROOT DFO 外, 被划分为 3 个 DFO, 调度时会先调 0 和 1 对应的 DFO, 待 0 号 DFO 执行完毕后, 会再调度 1 和 2 号 DFO, 依次迭代完成执行。

两DFO.jpeg