title: MySQL · myrocks · myrocks index condition pushdown

author: 张远

index condition pushdown

Index condition pushdown(ICP)是直到mysql5.6才引入的特性，主要是为了减少通过二级索引查找主键索引的次数。目前ICP相关的文章也比较多，本文主要从源码角度介绍ICP的实现。讨论之前，我们先再温习下。

以下图片来自mariadb

• 引入ICP之前
  
• 引入ICP之后
  

再来看个例子

CREATE TABLE `t1` (
  `a` int(11) DEFAULT NULL,
  `b` char(8) DEFAULT NULL,
  `c` int(11) DEFAULT '0',
  `pk` int(11) NOT NULL AUTO_INCREMENT,
  PRIMARY KEY (`pk`),
  KEY `idx1` (`a`,`b`)
) ENGINE=ROCKSDB;
INSERT INTO t1 (a,b) VALUES (1,'a'),(2,'b'),(3,'c');
INSERT INTO t1 (a,b) VALUES (4,'a'),(4,'b'),(4,'c'),(4,'d'),(4,'e'),(4,'f');

set optimizer_switch='index_condition_pushdown=off';

# 关闭ICP(Using where)
explain select * from t1 where a=4 and b!='e';
+----+-------------+-------+-------+---------------+------+---------+------+------+-------------+
| id | select_type | table | type  | possible_keys | key  | key_len | ref  | rows | Extra       |
+----+-------------+-------+-------+---------------+------+---------+------+------+-------------+
|  1 | SIMPLE      | t1    | range | idx1          | idx1 | 14      | NULL |    2 | Using where |
+----+-------------+-------+-------+---------------+------+---------+------+------+-------------+

# 关闭ICP走cover index(Using where; Using index)
explain select a,b from t1 where a=4 and b!='e';
+----+-------------+-------+------+---------------+------+---------+-------+------+--------------------------+
| id | select_type | table | type | possible_keys | key  | key_len | ref   | rows | Extra                    |
+----+-------------+-------+------+---------------+------+---------+-------+------+--------------------------+
|  1 | SIMPLE      | t1    | ref  | idx1          | idx1 | 5       | const |    4 | Using where; Using index |
+----+-------------+-------+------+---------------+------+---------+-------+------+--------------------------+

set optimizer_switch='index_condition_pushdown=on';

# 开启ICP(Using index conditione)
explain select * from t1 where a=4 and b!='e';
+----+-------------+-------+-------+---------------+------+---------+------+------+-----------------------+
| id | select_type | table | type  | possible_keys | key  | key_len | ref  | rows | Extra                 |
+----+-------------+-------+-------+---------------+------+---------+------+------+-----------------------+
|  1 | SIMPLE      | t1    | range | idx1          | idx1 | 14      | NULL |    2 | Using index condition |
+----+-------------+-------+-------+---------------+------+---------+------+------+-----------------------+

# 开启ICP仍然是cover index(Using where; Using index)
explain select a,b from t1 where a=4 and b!='e';
+----+-------------+-------+------+---------------+------+---------+-------+------+--------------------------+
| id | select_type | table | type | possible_keys | key  | key_len | ref   | rows | Extra                    |
+----+-------------+-------+------+---------------+------+---------+-------+------+--------------------------+
|  1 | SIMPLE      | t1    | ref  | idx1          | idx1 | 5       | const |    4 | Using where; Using index |
+----+-------------+-------+------+---------------+------+---------+-------+------+--------------------------+

这里总结下ICP的条件

• 适用于以下类型，range, ref, eq_ref, and ref_or_null 的二级索引
• 不能是覆盖索引(cover index)

server层主要负责判断是否符合ICP的条件，符合ICP则把需要的condition push到engine层。
engine层通过二级索引查找数据时，用server层push的condition再做一次判断，如果符合条件才会去查找主索引。

目前mysql支持ICP的引擎有MyISAM和InnoDB,MyRocks引入rocksdb后，也支持了ICP。
server层实现是一样的，engine层我们主要介绍innodb和rocksdb的实现。

server层

关键代码片段如下

make_join_readinfo()
  
switch (tab->type) {
    case JT_EQ_REF:
    case JT_REF_OR_NULL:
    case JT_REF:
      if (tab->select)
        tab->select->set_quick(NULL);
      delete tab->quick;
      tab->quick=0;
      /* fall through */
    case JT_SYSTEM:
    case JT_CONST:
      /* Only happens with outer joins */
      if (setup_join_buffering(tab, join, options, no_jbuf_after,
                               &icp_other_tables_ok))
        DBUG_RETURN(true);
      if (tab->use_join_cache != JOIN_CACHE::ALG_NONE)
        tab[-1].next_select= sub_select_op;

      if (table->covering_keys.is_set(tab->ref.key) &&
          !table->no_keyread)
        table->set_keyread(TRUE);
      else
        push_index_cond(tab, tab->ref.key, icp_other_tables_ok,
                        &trace_refine_table);
      break;

从代码中看出只有符合的类型range, ref, eq_ref, and ref_or_null 二级索引才可能会push_index_cond。

而这里通过covering_keys来判断并排除使用了cover index的情况。covering_keys是一个bitmap，保存了所有可能用到的覆盖索引。在解析查询列以及条件列时会设置covering_keys，详细可以参考setup_fields，setup_wild，setup_conds。

engine层

innodb

innodb在扫描二级索引时会根据是否有push condition来检查记录是否符合条件(row_search_idx_cond_check)
逻辑如下：

row_search_for_mysql()
......
  if (prebuilt->idx_cond)
  {
      row_search_idx_cond_check //检查condition
      row_sel_get_clust_rec_for_mysql //检查通过了才会去取主索引数据
  }
....

典型的堆栈如下

handler::compare_key_icp
innobase_index_cond
row_search_idx_cond_check
row_search_for_mysql
ha_innobase::index_read
ha_innobase::index_first
ha_innobase::rnd_next
handler::ha_rnd_next
rr_sequential
join_init_read_record
sub_select
do_select

rocksdb

rocksdb在扫描二级索引时也会根据是否有push condition来检查记录是否符合条件

逻辑如下

read_row_from_secondary_key()
{
   find_icp_matching_index_rec//push了condition才会检查condition
   get_row_by_rowid//检查通过了才会去取主索引数据
}

典型的堆栈如下

handler::compare_key_icp
myrocks::ha_rocksdb::check_index_cond
myrocks::ha_rocksdb::find_icp_matching_index_rec 
myrocks::ha_rocksdb::read_row_from_secondary_key 
myrocks::ha_rocksdb::index_read_map_impl 
myrocks::ha_rocksdb::read_range_first
handler::multi_range_read_next 

other

ICP对cover index作出了严格的限制，而实际上应该可以放开此限制，这样可以减少enging层传第给server层的数据量，至少可以减少server层的内存使用。欢迎指正！