云原生数据库 PolarDB：分区表扫描算子

前提条件

支持的PolarDB PostgreSQL版的版本如下：

PostgreSQL 14（内核小版本14.9.15.0及以上）

select version();

背景信息

对数据库的分区表进行扫描时，优化器会对每个子分区生成最优的执行计划，然后通过Append算子将子分区的执行计划串联起来，作为分区表扫描的最优执行计划。如果子分区的数量不多，上述过程将会快速完成。然而，PolarDB PostgreSQL版对分区表的分区数量没有限制，当子分区数量过多时，优化器所使用的时间和SQL执行过程中所使用的内存使用将会急剧增大，与扫描相同大小的普通表相比差距尤为明显。

为了解决该问题，PolarDB PostgreSQL版提供了分区表扫描算子（PartitionedTableScan，简称PTS）。与Append算子相比，PTS能够明显减少优化器生成执行计划的时间，且在SQL执行过程中使用更少的内存，能够有效避免OOM。

使用限制

• PTS当前仅支持SELECT，暂不支持DML语句。

• PTS不支持智能分区连接（Partition Wise Join），如果您打开了enable_partitionwise_join，将不会生成带有PTS算子的执行计划。

参数说明

使用说明

通过参数开启PTS算子

SET polar_num_parts_for_pts TO 64;

使用HINT

使用HINT语法PTScan（tablealias），示例如下：

EXPLAIN (COSTS OFF, ANALYZE) /*+ PTScan(part_range) */ SELECT * FROM part_range;
                                   QUERY PLAN
--------------------------------------------------------------------------------
 PartitionedTableScan on part_range (actual time=86.404..86.405 rows=0 loops=1)
   Scan 1000 Partitions: part_range_p0, part_range_p1, part_range_p2,...
   ->  Seq Scan on part_range
 Planning Time: 36.613 ms
 Execution Time: 89.246 ms
(5 rows)

并行查询

PTS算子支持并行查询，并行的方式包括分区间并行和混合并行两种，均已默认开启，无需调整。

• 分区间并行：每个工作进程查询一个分区。

• 混合并行：分区间和分区内都可以并行执行。

示例

1. 创建两张分区表，并分别创建1000个子分区。

   CREATE TABLE part_range (a INT, b VARCHAR, c NUMERIC, d INT8) PARTITION BY RANGE (a);
   SELECT 'CREATE TABLE part_range_p' || i || ' PARTITION OF part_range FOR VALUES FROM (' || 10 * i || ') TO (' || 10 * (i + 1) || ');'
   FROM generate_series(0,999) i;\gexec
   
   CREATE TABLE part_range2 (a INT, b VARCHAR, c NUMERIC, d INT8) PARTITION BY RANGE (a);
   SELECT 'CREATE TABLE part_range2_p' || i || ' PARTITION OF part_range2 FOR VALUES FROM (' || 10 * i || ') TO (' || 10 * (i + 1) || ');'
   FROM generate_series(0,999) i;\gexec

2. 对分区表进行全表扫描的执行计划如下。

   SET polar_num_parts_for_pts TO 0;
   EXPLAIN (COSTS OFF, ANALYZE) SELECT * FROM part_range;
                                            QUERY PLAN
   ---------------------------------------------------------------------------------------------
    Append (actual time=8.376..8.751 rows=0 loops=1)
      ->  Seq Scan on part_range_p0 part_range_1 (actual time=0.035..0.036 rows=0 loops=1)
      ->  Seq Scan on part_range_p1 part_range_2 (actual time=0.009..0.009 rows=0 loops=1)
      ->  Seq Scan on part_range_p2 part_range_3 (actual time=0.010..0.011 rows=0 loops=1)
     ...
     ...
     ...
      ->  Seq Scan on part_range_p997 part_range_998 (actual time=0.009..0.009 rows=0 loops=1)
      ->  Seq Scan on part_range_p998 part_range_999 (actual time=0.010..0.010 rows=0 loops=1)
      ->  Seq Scan on part_range_p999 part_range_1000 (actual time=0.009..0.009 rows=0 loops=1)
    Planning Time: 785.169 ms
    Execution Time: 163.534 ms
   (1003 rows)

3. 将两张分区表进行连接查询时，执行计划的生成时间长和SQL执行过程中内存使用多的问题将会更加明显。

   => SET polar_num_parts_for_pts TO 0;
   => EXPLAIN (COSTS OFF, ANALYZE)
         SELECT COUNT(*) FROM part_range a
         JOIN part_range2 b ON a.a = b.a
         WHERE b.c = '0001';
                                                            QUERY PLAN
   ----------------------------------------------------------------------------------------------------------------------------
    Finalize Aggregate (actual time=3191.718..3212.437 rows=1 loops=1)
      ->  Gather (actual time=2735.417..3212.288 rows=3 loops=1)
            Workers Planned: 2
            Workers Launched: 2
            ->  Partial Aggregate (actual time=2667.247..2667.789 rows=1 loops=3)
                  ->  Parallel Hash Join (actual time=1.957..2.497 rows=0 loops=3)
                        Hash Cond: (a.a = b.a)
                        ->  Parallel Append (never executed)
                              ->  Parallel Seq Scan on part_range_p0 a_1 (never executed)
                              ->  Parallel Seq Scan on part_range_p1 a_2 (never executed)
                              ->  Parallel Seq Scan on part_range_p2 a_3 (never executed)
                              ...
                              ...
                              ...
                              ->  Parallel Seq Scan on part_range_p997 a_998 (never executed)
                              ->  Parallel Seq Scan on part_range_p998 a_999 (never executed)
                              ->  Parallel Seq Scan on part_range_p999 a_1000 (never executed)
                        ->  Parallel Hash (actual time=0.337..0.643 rows=0 loops=3)
                              Buckets: 4096  Batches: 1  Memory Usage: 0kB
                              ->  Parallel Append (actual time=0.935..1.379 rows=0 loops=1)
                                    ->  Parallel Seq Scan on part_range2_p0 b_1 (actual time=0.001..0.001 rows=0 loops=1)
                                          Filter: (c = '1'::numeric)
                                    ->  Parallel Seq Scan on part_range2_p1 b_2 (actual time=0.001..0.001 rows=0 loops=1)
                                          Filter: (c = '1'::numeric)
                                    ->  Parallel Seq Scan on part_range2_p2 b_3 (actual time=0.001..0.001 rows=0 loops=1)
                                          Filter: (c = '1'::numeric)
                                    ...
                                    ...
                                    ...
                                    ->  Parallel Seq Scan on part_range2_p997 b_998 (actual time=0.001..0.001 rows=0 loops=1)
                                          Filter: (c = '1'::numeric)
                                    ->  Parallel Seq Scan on part_range2_p998 b_999 (actual time=0.000..0.001 rows=0 loops=1)
                                          Filter: (c = '1'::numeric)
                                    ->  Parallel Seq Scan on part_range2_p999 b_1000 (actual time=0.002..0.002 rows=0 loops=1)
                                          Filter: (c = '1'::numeric)
    Planning Time: 1900.615 ms
    Execution Time: 3694.320 ms
   (3013 rows)

4. 从上述示例可以看出，分区表的全表查询由于缺少分区键作为过滤条件，无法将查询集中在部分分区，相比于普通表是没有任何优势的。分区表比普通表更加低效。分区表的最佳实践是尽可能使查询发生分区裁剪，使查询可以集中在少部分分区上。但是对于一些OLAP场景，不得不对分区表进行全表扫描。此时，使用PTS算子比Append算子更加高效。

   SET polar_num_parts_for_pts TO 10;
   EXPLAIN (COSTS OFF, ANALYZE) SELECT * FROM part_range;
                                      QUERY PLAN
   --------------------------------------------------------------------------------
    PartitionedTableScan on part_range (actual time=86.404..86.405 rows=0 loops=1)
      Scan 1000 Partitions: part_range_p0, part_range_p1, part_range_p2,...
      ->  Seq Scan on part_range
    Planning Time: 36.613 ms
    Execution Time: 89.246 ms
   (5 rows)

   SET polar_num_parts_for_pts TO 10;
   EXPLAIN (COSTS OFF, ANALYZE)
         SELECT COUNT(*) FROM part_range a
         JOIN part_range2 b ON a.a = b.a
         WHERE b.c = '0001';
                                                QUERY PLAN
   ----------------------------------------------------------------------------------------------------
    Aggregate (actual time=61.384..61.388 rows=1 loops=1)
      ->  Merge Join (actual time=61.378..61.381 rows=0 loops=1)
            Merge Cond: (a.a = b.a)
            ->  Sort (actual time=61.377..61.378 rows=0 loops=1)
                  Sort Key: a.a
                  Sort Method: quicksort  Memory: 25kB
                  ->  PartitionedTableScan on part_range a (actual time=61.342..61.343 rows=0 loops=1)
                        Scan 1000 Partitions: part_range_p0, part_range_p1, part_range_p2, ...
                        ->  Seq Scan on part_range a
            ->  Sort (never executed)
                  Sort Key: b.a
                  ->  PartitionedTableScan on part_range2 b (never executed)
                        ->  Seq Scan on part_range2 b
                              Filter: (c = '1'::numeric)
    Planning Time: 96.675 ms
    Execution Time: 64.913 ms
   (16 rows)

   结果显示，使用PTS算子后，执行计划的生成时间得到了明显的缩短。

性能对比

单条SQL的执行计划生成时间

单条SQL的内存使用量

PGBench QPS