文档首页/ 云数据库 GaussDB/ 特性指南（分布式_8.x）/ 分区表/ 分区表查询优化/ 分区剪枝/ 分区表动态剪枝/ PBE动态剪枝
更新时间：2024-06-07 GMT+08:00
查看PDF
分享

PBE动态剪枝

PBE动态剪枝支持范围如下所示：
  1. 支持分区类型：范围分区、哈希分区、列表分区。
  2. 支持表达式类型：比较表达式（<，<=，=，>=，>）、逻辑表达式、数组表达式。
  3. 支持隐式类型转换和函数：对于类型可以相互转换的场景和immutable函数可以支持PBE动态剪枝
  • PBE动态剪枝支持表达式、隐式转换、immutable函数和stable函数，不支持子查询表达式和volatile函数。对于stable函数，如to_timestamp等类型转换函数，可能会受GUC参数变化，影响剪枝结果。为了保持性能优化，此情况可以通过analyze表重新生成gplan解决。
  • 由于PBE动态剪枝是基于generic plan的剪枝，所以判断语句是否能PBE动态剪枝时，需要设置参数plan_cache_mode = 'force_generic_plan'，排除custom plan的干扰。
  • PBE动态剪枝支持的典型场景具体示例如下：
    1. 比较表达式。 
      gaussdb=#
--创建分区表
CREATE TABLE t1 (c1 int, c2 int)
PARTITION BY RANGE (c1)
(
    PARTITION p1 VALUES LESS THAN(10),
    PARTITION p2 VALUES LESS THAN(20),
    PARTITION p3 VALUES LESS THAN(MAXVALUE)
);

gaussdb=# PREPARE p1(int) AS SELECT * FROM t1 WHERE c1 = $1;
PREPARE
gaussdb=# EXPLAIN (VERBOSE ON, COSTS OFF) EXECUTE p1(1);
                         QUERY PLAN                         
------------------------------------------------------------
 Data Node Scan
   Output: t1.c1, t1.c2
   Node/s: datanode1
   Node expr: $1
   Remote query: SELECT c1, c2 FROM public.t1 WHERE c1 = $1

 Remote SQL: SELECT c1, c2 FROM public.t1 WHERE c1 = $1
 Datanode Name: datanode1
   Partition Iterator
     Output: c1, c2
     Iterations: PART
     ->  Partitioned Seq Scan on public.t1
           Output: c1, c2
           Filter: (t1.c1 = $1)
           Selected Partitions:  1 (pbe-pruning)

(16 rows)
    2. 逻辑表达式。 
      gaussdb=# PREPARE p2(INT, INT) AS SELECT * FROM t1 WHERE c1 = $1 AND c2 = $2;
PREPARE
gaussdb=# EXPLAIN (VERBOSE ON, COSTS OFF) EXECUTE p2(1, 2);
                               QUERY PLAN                               
------------------------------------------------------------------------
 Data Node Scan
   Output: t1.c1, t1.c2
   Node/s: datanode1
   Node expr: $1
   Remote query: SELECT c1, c2 FROM public.t1 WHERE c1 = $1 AND c2 = $2

 Remote SQL: SELECT c1, c2 FROM public.t1 WHERE c1 = $1 AND c2 = $2
 Datanode Name: datanode1
   Partition Iterator
     Output: c1, c2
     Iterations: PART
     ->  Partitioned Seq Scan on public.t1
           Output: c1, c2
           Filter: ((t1.c1 = $1) AND (t1.c2 = $2))
           Selected Partitions:  1 (pbe-pruning)

(16 rows)
    3. 类型转换触发隐式转换。 
      gaussdb=# set plan_cache_mode = 'force_generic_plan';
gaussdb=# PREPARE p3(TEXT) AS SELECT * FROM t1 WHERE c1 = $1;
PREPARE
gaussdb=# EXPLAIN (VERBOSE ON, COSTS OFF) EXECUTE p3('12');
                             QUERY PLAN                             
--------------------------------------------------------------------
 Data Node Scan
   Output: t1.c1, t1.c2
   Node/s: datanode1
   Node expr: $1
   Remote query: SELECT c1, c2 FROM public.t1 WHERE c1 = $1::bigint

 Remote SQL: SELECT c1, c2 FROM public.t1 WHERE c1 = $1::bigint
 Datanode Name: datanode1
   Partition Iterator
     Output: c1, c2
     Iterations: PART
     ->  Partitioned Seq Scan on public.t1
           Output: c1, c2
           Filter: (t1.c1 = ($1)::bigint)
           Selected Partitions:  2 (pbe-pruning)

(16 rows)
  • PBE动态剪枝不支持的典型场景具体示例如下：
    1. 子查询表达式。 
      gaussdb=# PREPARE p4(INT) AS SELECT * FROM t1 WHERE c1 = ALL(SELECT c2 FROM t1 WHERE c1 > $1);
PREPARE
gaussdb=# EXPLAIN (VERBOSE ON, COSTS OFF) EXECUTE p4(1);
                               QUERY PLAN                                
-------------------------------------------------------------------------
 Streaming (type: GATHER)
   Output: public.t1.c1, public.t1.c2
   Node/s: All datanodes
   ->  Partition Iterator
         Output: public.t1.c1, public.t1.c2
         Iterations: 3
         ->  Partitioned Seq Scan on public.t1
               Output: public.t1.c1, public.t1.c2
               Distribute Key: public.t1.c1
               Filter: (SubPlan 1)
               Selected Partitions:  1..3
               SubPlan 1
                 ->  Materialize
                       Output: public.t1.c2
                       ->  Streaming(type: BROADCAST)
                             Output: public.t1.c2
                             Spawn on: All datanodes
                             Consumer Nodes: All datanodes
                             ->  Partition Iterator
                                   Output: public.t1.c2
                                   Iterations: 3
                                   ->  Partitioned Seq Scan on public.t1
                                         Output: public.t1.c2
                                         Distribute Key: public.t1.c1
                                         Filter: (public.t1.c1 > 1)
                                         Selected Partitions:  1..3
(26 rows)
    2. 类型转换无法直接触发隐式转换。 
      gaussdb=# PREPARE p5(name) AS SELECT * FROM t1 WHERE c1 = $1;
PREPARE
gaussdb=# EXPLAIN (VERBOSE ON, COSTS OFF) EXECUTE p5('12');
                                QUERY PLAN                                
--------------------------------------------------------------------------
 Data Node Scan
   Output: t1.c1, t1.c2
   Node/s: All datanodes
   Remote query: SELECT c1, c2 FROM public.t1 WHERE c1::text = '12'::text

 Remote SQL: SELECT c1, c2 FROM public.t1 WHERE c1::text = '12'::text
 Datanode Name: datanode1
   Partition Iterator
     Output: c1, c2
     Iterations: 3
     ->  Partitioned Seq Scan on public.t1
           Output: c1, c2
           Filter: ((t1.c1)::text = '12'::text)
           Selected Partitions:  1..3

(15 rows)
    3. stable/volatile函数。 
      gaussdb=# create sequence seq;
gaussdb=# PREPARE p6(TEXT) AS SELECT * FROM t1 WHERE c1 = currval($1);--volatile函数不支持剪枝。
PREPARE
gaussdb=# EXPLAIN (VERBOSE ON, COSTS OFF) EXECUTE p6('seq');
                                 QUERY PLAN                                 
----------------------------------------------------------------------------
 Data Node Scan
   Output: t1.c1, t1.c2
   Node/s: All datanodes
   Remote query: SELECT c1, c2 FROM ONLY public.t1 WHERE true
   Coordinator quals: ((t1.c1)::numeric = currval(('seq'::text)::regclass))

 Remote SQL: SELECT c1, c2 FROM ONLY public.t1 WHERE true
 Datanode Name: datanode1
   Partition Iterator
     Output: c1, c2
     Iterations: 3
     ->  Partitioned Seq Scan on public.t1
           Output: c1, c2
           Selected Partitions:  1..3

(15 rows)

--清理示例环境。
gaussdb=# DROP TABLE t1;
父主题： 分区表动态剪枝