基于代价的路径选择 ============================== 在基于规则的路径选择之后，如果存在多个可以选择的路径，那么 OceanBase 数据库会按个计算每个路径的代价，并从中选择代价最小的路径作为最终选择的路径。OceanBase 数据库的代价模型考虑了 CPU 代价(比如处理一个谓词的 CPU 开销)和 IO 代价(比如顺序和随机读取宏块和微块的代价)，CPU 代价和 IO 代价最终相加得到一个总的代价。在 OceanBase 数据库中，每个访问路径的代价在计划中都会展示出来。 ```javascript obclient> create table t1(a int primary key, b int, c int, index k1(b)); Query OK, 0 rows affected (0.35 sec) -- 主表路径的代价 obclient> explain select/*+index(t1 primary)*/ * from t1 where b < 10; +------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Query Plan | +------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | =================================== |ID|OPERATOR |NAME|EST. ROWS|COST| ----------------------------------- |0 |TABLE SCAN|t1 |200 |622 | =================================== Outputs & filters: ------------------------------------- 0 - output([t1.a], [t1.b], [t1.c]), filter([t1.b < 10]), access([t1.b], [t1.a], [t1.c]), partitions(p0) -- 所以k1路径的代价 obclient> explain select/*+index(t1 k1)*/ * from t1 where b < 10; +--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Query Plan | +--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | ===================================== |ID|OPERATOR |NAME |EST. ROWS|COST| ------------------------------------- |0 |TABLE SCAN|t1(k1)|200 |1114| ===================================== Outputs & filters: ------------------------------------- 0 - output([t1.a], [t1.b], [t1.c]), filter(nil), access([t1.b], [t1.a], [t1.c]), partitions(p0) ``` 对于一个访问路径，它的代价主要由两部分组成，扫描访问路径的代价和回表的代价。如果一个访问路径不需要回表，那么就没有回表的代价。在 OceanBase 数据库中，访问路径的代价取决于很多因素，比如扫描的行数，回表的行数，投影的列数，谓词的个数等。但是对于访问路径来说，代价在很大程度上取决于行数，所以在下面的分析中，我们从行数这个维度来介绍这两部分的代价。 * 扫描访问路径的代价: 扫描访问路径的代价跟扫描的行数成正比，理论上来说扫描的行数越多，执行时间就会越久。对于一个访问路径，query range 决定了需要扫描的范围，从而决定了需要扫描的行数。Query range 的扫描是顺序 IO。 * 回表的代价: 回表的代价跟回表的行数也也是正相关的，回表的行数越多(回表的行数是指满足所有能在索引上执行的谓词的行数)，执行时间就会越久。回表的扫描是随机 IO，所以回表一行的代价会比 query range 扫描一行的代价高很多。 当分析一个访问路径的性能的时候，可以从上面两个因素入手，既定位出来通过 query range 扫描的行数以及回表的行数。这两个行数通常我们可以通过执行 sql 语句来获取，如下例所示，对于查询 `select * from t1 where c2 > 20 and c2 < 800 and c3 < 200`, 对于索引 k1 这个访问路径，首先通常计划展示来获取用来抽取 query range 的谓词，谓词 c2 \> 20 and c2 \< 800 用来抽取 query range 了，谓词 c3 \< 200 被当成回表前的谓词。那么可以使用如下例所示的两个查询来检查 query range 抽取的行数以及回表之后的行数。 ```javascript obclient> create table t1(c1 int primary key, c2 int, c3 int, c4 int, c5 int, index k1(c2,c3)); Query OK, 0 rows affected (0.26 sec) obclient> explain extended_noaddr select/*+index(t1 k1)*/ * from t1 where c2 > 20 and c2 < 800 and c3 < 200; +----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Query Plan | +----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | ===================================== |ID|OPERATOR |NAME |EST. ROWS|COST| ------------------------------------- |0 |TABLE SCAN|t1(k1)|156 |1216| ===================================== Outputs & filters: ------------------------------------- 0 - output([t1.c1], [t1.c2], [t1.c3], [t1.c4], [t1.c5]), filter([t1.c3 < 200]), access([t1.c2], [t1.c3], [t1.c1], [t1.c4], [t1.c5]), partitions(p0), is_index_back=true, filter_before_indexback[true], range_key([t1.c2], [t1.c3], [t1.c1]), range(20,MAX,MAX ; 800,MIN,MIN), range_cond([t1.c2 > 20], [t1.c2 < 800]) -- query range 扫描的行数 obclient> select/*+index(t1 k1)*/ count(*) from t1 where c2 > 20 and c2 < 800; +----------+ | count(*) | +----------+ | 779 | +----------+ 1 row in set (0.02 sec) -- 回表的行数 obclient> select/*+index(t1 k1)*/ count(*) from t1 where c2 > 20 and c2 < 800 and c3 < 200; +----------+ | count(*) | +----------+ | 179 | +----------+ 1 row in set (0.01 sec) ```