GaussDB-算子级调优
算子级调优介绍
一个查询语句要经过多个算子步骤才会输出最终的结果。由于个别算子耗时过长导致整体查询性能下降的情况比较常见。这些算子是整个查询的瓶颈算子。通用的优化手段是EXPLAIN ANALYZE/PERFORMANCE命令查看执行过程的瓶颈算子，然后进行针对性优化。
如下面的执行过程信息中，Hashagg算子的执行时间占总时间的：(51016-13535)/ 56476 ≈66%，此处Hashagg算子就是这个查询的瓶颈算子，在进行性能优化时应当优先考虑此算子的优化。
算子级调优示例
示例1： 基表扫描时，对于点查询或者范围扫描等过滤大量数据的查询，如果使用SeqScan全表扫描会比较耗时，可以在条件列上建立索引选择IndexScan进行索引扫描提升扫描效率。
| ``` gaussdb=# explain (analyze on,costs off) select * from t1 where c2=10004; id | operation | A-time | A-rows | Peak Memory | A-width ----+------------------------------+-----------------+--------+-------------+--------- 1 | -> Streaming (type: GATHER) | 20.040 | 5 | 85KB | 2 | -> Seq Scan on t1 | [17.239,17.376] | 5 | [18KB,18KB] | (2 rows) Predicate Information (identified by plan id) ----------------------------------------------- 2 --Seq Scan on t1 Filter: (c2 = 10004) Rows Removed by Filter: 90002 (3 rows)
| ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| ```
gaussdb=# create index idx on t1(c2); CREATE INDEX gaussdb=# explain (analyze on,costs off) select * from t1 where c2=10004; id | operation | A-time | A-rows | Peak Memory | A-width ----+-----------------------------------+---------------+--------+-------------+--------- 1 | -> Streaming (type: GATHER) | 3.206 | 5 | 85KB | 2 | -> Index Scan using idx on t1 | [0.122,0.146] | 5 | [73KB,73KB] | (2 rows) Predicate Information (identified by plan id) ----------------------------------------------- 2 --Index Scan using idx on t1 Index Cond: (c2 = 10004) (2 rows)
``` |
| --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
上述例子中，全表扫描返回5条数据，过滤掉大量数据，在c2列上建立索引后，使用IndexScan扫描效率显著提高，从20毫秒降低到3毫秒。
**示例2：** 如果从执行计划中看，两表join选择了NestLoop，而实际行数比较大时，NestLoop Join可能执行比较慢。如下的例子中NestLoop耗时5秒，如果设置参数enable_mergejoin=off关掉Merge Join，同时设置参数enable_nestloop=off关掉NestLoop，让优化器选择HashJoin，则Join耗时降低至86毫秒。
gaussdb=# explain analyze select count(*) from t2,t1 where t1.c1=t2.c2; id | operation | A-time | A-rows | E-rows | Peak Memory | A-width | E-width | E-costs ----+--------------------------------------------------+---------------------+----------+--------+-------------+---------+---------+--------- 1 | -> Aggregate | 5070.296 | 1 | 1 | 14KB | | 8 | 2148.49 2 | -> Streaming (type: GATHER) | 5070.219 | 2 | 2 | 81KB | | 8 | 2148.49 3 | -> Aggregate | [4828.705,5062.289] | 2 | 2 | [11KB,11KB] | | 8 | 2148.40 4 | -> Nested Loop (5,6) | [4828.565,5062.142] | 996 | 40 | [4KB,4KB] | | 0 | 2148.34 5 | -> Seq Scan on t1 | [13.574,14.508] | 90007 | 20000 | [15KB,15KB] | | 4 | 184.00 6 | -> Materialize | [1508.956,1579.488] | 22413670 | 20 | [35KB,36KB] | | 4 | 14.37 7 | -> Streaming(type: REDISTRIBUTE) | [55.825,56.842] | 498 | 20 | [44KB,44KB] | | 4 | 14.31 8 | -> Seq Scan on t2 | [0.105,0.132] | 498 | 20 | [13KB,13KB] | | 4 | 13.13 (8 rows)
Predicate Information (identified by plan id)
4 --Nested Loop (5,6) Join Filter: (t2.c2 = t1.c1) Rows Removed by Join Filter: 22412672 (3 rows)
设置参数后：
gaussdb=# set enable_mergejoin=off; SET gaussdb=# set enable_nestloop=off; SET gaussdb=# explain analyze select count(*) from t2,t1 where t1.c1=t2.c2; id | operation | A-time | A-rows | E-rows | Peak Memory | A-width | E-width | E-costs ----+--------------------------------------------------+-----------------+--------+--------+---------------+---------+---------+--------- 1 | -> Aggregate | 92.911 | 1 | 1 | 14KB | | 8 | 224.45 2 | -> Streaming (type: GATHER) | 92.855 | 2 | 2 | 81KB | | 8 | 224.45 3 | -> Aggregate | [84.295,87.102] | 2 | 2 | [11KB,11KB] | | 8 | 224.36 4 | -> Hash Join (5,6) | [84.171,86.966] | 996 | 40 | [6KB,6KB] | | 0 | 224.30 5 | -> Seq Scan on t1 | [11.885,13.103] | 90007 | 20000 | [15KB,15KB] | | 4 | 184.00 6 | -> Hash | [55.895,56.072] | 498 | 21 | [292KB,292KB] | [20,20] | 4 | 14.31 7 | -> Streaming(type: REDISTRIBUTE) | [55.601,55.771] | 498 | 20 | [44KB,44KB] | | 4 | 14.31 8 | -> Seq Scan on t2 | [0.118,0.143] | 498 | 20 | [13KB,13KB] | | 4 | 13.13 (8 rows)
Predicate Information (identified by plan id)
4 --Hash Join (5,6) Hash Cond: (t1.c1 = t2.c2) (2 rows)
**示例3**：通常情况下Agg选择HashAgg性能较好，如果大结果集选择了Sort+GroupAgg，则需要设置enable_sort=off，HashAgg耗时优于Sort+GroupAgg。
gaussdb=# explain analyze select count(*) from t1 group by c2; id | operation | A-time | A-rows | E-rows | Peak Memory | E-memory | A-width | E-width | E-costs ----+------------------------------------+-----------------+--------+--------+-------------+----------+-----------------+---------+--------- 1 | -> GroupAggregate | 244.817 | 40000 | 5000 | 15KB | | | 12 | 2131.52 2 | -> Sort | 156.344 | 40000 | 10000 | 5603KB | | | 12 | 2131.52 3 | -> Streaming (type: GATHER) | 91.595 | 40000 | 10000 | 82KB | | | 12 | 1442.14 4 | -> GroupAggregate | [90.317,96.852] | 40000 | 10000 | [12KB,12KB] | 16MB | | 12 | 973.39 5 | -> Sort | [59.775,64.724] | 90007 | 20000 | [5MB,5MB] | 16MB | [896220,903920] | 4 | 873.39 6 | -> Seq Scan on t1 | [18.092,21.033] | 90007 | 20000 | [12KB,12KB] | 1MB | | 4 | 184.00 (6 rows)
设置参数后：
gaussdb=# set enable_sort=off; SET gaussdb=# explain analyze select count(*) from t1 group by c2; id | operation | A-time | A-rows | E-rows | Peak Memory | E-memory | A-width | E-width | E-costs ----+---------------------------------+-----------------+--------+--------+-------------+----------+---------+---------+--------- 1 | -> HashAggregate | 228.260 | 40000 | 5000 | 6663KB | | | 12 | 752.75 2 | -> Streaming (type: GATHER) | 95.506 | 40000 | 10000 | 82KB | | | 12 | 752.75 3 | -> HashAggregate | [63.974,71.290] | 40000 | 10000 | [3MB,3MB] | 16MB | [20,20] | 12 | 284.00 4 | -> Seq Scan on t1 | [17.578,21.204] | 90007 | 20000 | [12KB,12KB] | 1MB | | 4 | 184.00 (4 rows)
更多详情请参考GaussDB 文档中心：<https://doc.hcs.huawei.com/db/zh-cn/gaussdbqlh/24.1.30/productdesc/qlh_03_0001.html>
