摘要：本文就针对因USING子句的书写方式可能导致MERGE INTO语句的执行不下推的场景，对USING子句的SQL语句进行改写一遍，整个SQL语句可以下推。
本文分享自华为云社区《GaussDB(DWS)运维 -- values子句做MERGE数据源导致SQL执行不下推的改写方案》，作者： 譡里个檔。
现网做实时接入的时候，有的时候会使用MERGE INTO语句实现类似UPSERT的功能。这种场景下MERGE INTO语句的USING部分的数据位VALUES子句，为了后续的SQL语句中描述方便，需要对VALUES子句的输出命名别名。USING子句的书写方式可能导致MERGE INTO语句的执行不下推，本文就针对因此导致的不下推的场景，对USING子句的SQL语句进行改写一遍，整个SQL语句可以下推。
预置条件
CREATE TABLE t1(name text, id INT) DISTRIBUTE BY HASH(id);
原始语句
MERGE INTO t1 USING (
SELECT *
FROM (VALUES ('json', 1), ('sam', 2)) AS val(name, id)
) tmp ON (t1.id = tmp.id)
WHEN MATCHED THEN
UPDATE SET t1.name = tmp.name
WHEN NOT MATCHED THEN
INSERT (name, id) VALUES(tmp.name, tmp.id);
SQL语句不下推，导致执行低效
postgres=# EXPLAIN VERBOSE MERGE INTO t1 USING (
postgres(# SELECT *
postgres(# FROM (VALUES ('json', 1), ('sam', 2)) AS val(name, id)
postgres(# ) tmp ON (t1.id = tmp.id)
postgres-# WHEN MATCHED THEN
postgres-# UPDATE SET t1.name = tmp.name
postgres-# WHEN NOT MATCHED THEN
postgres-# INSERT (name, id) VALUES(tmp.name, tmp.id);
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------------------
id | operation | E-rows | E-distinct | E-width | E-costs
----+-------------------------------------------------------+--------+------------+---------+---------
1 | -> Merge on public.t1 | 2 | | 54 | 0.08
2 | -> Nested Loop Left Join (3, 4) | 2 | | 54 | 0.08
3 | -> Values Scan on "*VALUES*" | 2 | | 36 | 0.03
4 | -> Data Node Scan on t1 "_REMOTE_TABLE_QUERY_" | 2 | | 18 | 0.00
SQL Diagnostic Information
------------------------------------------------------------
SQL is not plan-shipping
reason: Type of Record in non-real table can not be shipped
Predicate Information (identified by plan id)
-------------------------------------------------
1 --Merge on public.t1
Node expr: : $10
2 --Nested Loop Left Join (3, 4)
Join Filter: (t1.id = "*VALUES*".column2)
Targetlist Information (identified by plan id)
-----------------------------------------------------------------------------------------------------------------------------------------------------------------
1 --Merge on public.t1
Node/s: All datanodes
Remote query: UPDATE ONLY public.t1 SET name = $7, id = $8 WHERE t1.ctid = $5 AND t1.xc_node_id = $6
Node/s: All datanodes
Remote query: INSERT INTO public.t1 (name, id) VALUES ($9, $10)
2 --Nested Loop Left Join (3, 4)
Output: "*VALUES*".column1, "*VALUES*".column2, t1.name, t1.id, t1.ctid, t1.xc_node_id, "*VALUES*".column1, t1.id, "*VALUES*".column1, "*VALUES*".column2
3 --Values Scan on "*VALUES*"
Output: "*VALUES*".column1, "*VALUES*".column2
4 --Data Node Scan on t1 "_REMOTE_TABLE_QUERY_"
Output: t1.name, t1.id, t1.ctid, t1.xc_node_id
Node/s: All datanodes
Remote query: SELECT name, id, ctid, xc_node_id FROM ONLY public.t1 WHERE true
====== Query Summary =====
--------------------------
Parser runtime: 0.079 ms
Planner runtime: 1.392 ms
Unique SQL Id: 1657855173
(40 rows)
改写方案
MERGE INTO t1 USING (
WITH val(name, id) AS(
VALUES ('json', 1), ('sam', 2)
)
SELECT * FROM val
) tmp ON (t1.id = tmp.id)
WHEN MATCHED THEN
UPDATE SET t1.name = tmp.name
WHEN NOT MATCHED THEN
INSERT (name, id) VALUES(tmp.name, tmp.id);
改写后下推
postgres=# EXPLAIN VERBOSE MERGE INTO t1 USING (
postgres(# WITH val(name, id) AS(
postgres(# VALUES ('json', 1), ('sam', 2)
postgres(# )
postgres(# SELECT * FROM val
postgres(# ) tmp ON (t1.id = tmp.id)
postgres-# WHEN MATCHED THEN
postgres-# UPDATE SET t1.name = tmp.name
postgres-# WHEN NOT MATCHED THEN
postgres-# INSERT (name, id) VALUES(tmp.name, tmp.id);
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------
id | operation | E-rows | E-distinct | E-memory | E-width | E-costs
----+----------------------------------------------+--------+------------+----------+---------+---------
1 | -> Streaming (type: GATHER) | 1 | | | 54 | 1.56
2 | -> Merge on public.t1 | 2 | | | 54 | 1.15
3 | -> Streaming(type: REDISTRIBUTE) | 2 | | 2MB | 54 | 1.15
4 | -> Nested Loop Left Join (5, 7) | 2 | | 1MB | 54 | 1.11
5 | -> Subquery Scan on tmp | 2 | | 1MB | 36 | 0.08
6 | -> Values Scan on "*VALUES*" | 24 | | 1MB | 36 | 0.03
7 | -> Seq Scan on public.t1 | 2 | | 1MB | 18 | 1.01
Predicate Information (identified by plan id)
---------------------------------------------
4 --Nested Loop Left Join (5, 7)
Join Filter: (t1.id = tmp.id)
5 --Subquery Scan on tmp
Filter: (Hash By tmp.id)
Targetlist Information (identified by plan id)
----------------------------------------------------------------------------------------------------------------------------------------------------
1 --Streaming (type: GATHER)
Node/s: All datanodes
3 --Streaming(type: REDISTRIBUTE)
Output: tmp.name, tmp.id, t1.name, t1.id, t1.ctid, t1.xc_node_id, tmp.name, tmp.id, (CASE WHEN (t1.ctid IS NULL) THEN tmp.id ELSE t1.id END)
Distribute Key: (CASE WHEN (t1.ctid IS NULL) THEN tmp.id ELSE t1.id END)
Spawn on: All datanodes
Consumer Nodes: All datanodes
4 --Nested Loop Left Join (5, 7)
Output: tmp.name, tmp.id, t1.name, t1.id, t1.ctid, t1.xc_node_id, tmp.name, tmp.id, CASE WHEN (t1.ctid IS NULL) THEN tmp.id ELSE t1.id END
5 --Subquery Scan on tmp
Output: tmp.name, tmp.id
6 --Values Scan on "*VALUES*"
Output: "*VALUES*".column1, "*VALUES*".column2
7 --Seq Scan on public.t1
Output: t1.name, t1.id, t1.ctid, t1.xc_node_id
Distribute Key: t1.id
====== Query Summary =====
-------------------------------
System available mem: 3112960KB
Query Max mem: 3112960KB
Query estimated mem: 6336KB
Parser runtime: 0.107 ms
Planner runtime: 1.185 ms
Unique SQL Id: 780461632
(44 rows)