MySQL · myrocks · collation 限制

背景

MyRocks中的数据是按索引列以memcmp方式进行排序的。对于一些数字类型，需要进行转化才能直接通过memcmp进行比较， 例如有符号数在计算机中是用补码表示的，那么如果负数和正数直接按字节比较，结果负数会比正数大，实际存储时会将符号会反转存储，读取时再转化回来。对于字符类型，处理更加复杂，涉及到字符集的转换。 记录格式可以参考[1], [2]

MyRocks索引字段如果包含字符类型，默认只支持binary collation，binary、latin1_bin、 utf8_bin其中的一种

## Error
create table t1(c1 int primary key, c2 varchar(10) unique) engine =rocksdb character set gbk;
ERROR 3046 (HY000): Unsupported collation on string indexed column test.t1.c2 Use binary collation (binary, latin1_bin, utf8_bin).

## Error
create table t1(c1 int primary key, c2 varchar(10) unique) engine =rocksdb character set utf8 collate utf8_general_ci;
ERROR 3046 (HY000): Unsupported collation on string indexed column test.t1.c2 Use binary collation (binary, latin1_bin, utf8_bin).

## OK
create table t1(c1 int primary key, c2 varchar(10) unique) engine =rocksdb character set utf8 collate utf8_bin;
Query OK, 0 rows affected (0.00 sec)

通过设置rocksdb_strict_collation_check参数为OFF可以跳出binary collation的限制

set global rocksdb_strict_collation_check=OFF;

## OK
create table t1(c1 int primary key, c2 varchar(10) unique) engine =rocksdb character set gbk;
Query OK, 0 rows affected (0.01 sec)

问题

MyRocks和InnoDB一样支持covering index. MyRocks在使用二级索引查询的时候，应尽量使用covering index, 因为MyRocks回表通过主键随机查询数据的开销比较大。

例如以下场景，idx1作为convering index被使用

create table t1(c1 int primary key, c2 varchar(10) unique) engine =rocksdb character set utf8 collate utf8_bin;

insert into t1 values(1,'ab');
insert into t1 values(2,'cd');

## covering index
explain select c2 from t1 where c2='ab';
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------------+
| id | select_type | table | type  | possible_keys | key  | key_len | ref   | rows | Extra       |
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------------+
|  1 | SIMPLE      | t1    | const | c2            | c2   | 33      | const |    1 | Using index |
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------------+

然而设置set global rocksdb_strict_collation_check=OFF;在某些情况下会导致我们无法使用covering index.

set global rocksdb_strict_collation_check=ON;

## Error
create table t1(c1 int primary key, c2 int, c3 varchar(10), key idx1(c2,c3)) engine =rocksdb  character set utf8 collate utf8_general_ci;
ERROR 3046 (HY000): Unsupported collation on string indexed column test.t1.c3 Use binary collation (binary, latin1_bin, utf8_bin).

set global rocksdb_strict_collation_check=OFF;

## OK
create table t1(c1 int primary key, c2 varchar(10) unique) engine =rocksdb character set utf8 collate utf8_general_ci;
Query OK, 0 rows affected (0.00 sec)

insert into t1 values(1,1,'ab');
insert into t1 values(2,2,'cd');


insert into t1 values(1,'ab');
insert into t1 values(2,'cd');

## non-covering index
explain select c2 from t1 where c2='ab';
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------+
| id | select_type | table | type  | possible_keys | key  | key_len | ref   | rows | Extra |
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------+
|  1 | SIMPLE      | t1    | const | c2            | c2   | 33      | const |    1 | NULL  |
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------+
1 row in set (0.00 sec)

原因

MyRocks二级索引由于collation的关系导致查询没有走covering index. MyRocks中索引列需要转化为memcomparable的形式，转化分为以下三种情况

• 1) 直接转换，不需要存储额外信息

  例如 binary、latin1_bin、 utf8_bin三种collation可以直接转换

  这种情况二级索引列数据可以完整的从二级索引中取到，不影响covering index使用

• 2) 间接转换，需在value中增加unpack_info

  例如latin1_general_ci,latin2_general_ci, ascii_general_ci,greek_general_ci等collation，具体可以参考函数rdb_is_collation_supported

  这种情况二级索引列数据可以从key和unpack_info中解析取到，也不影响covering index使用

create table t1(c1 int primary key, c2 varchar(10) unique) engine =rocksdb character set latin1 collate latin1_general_ci;

insert into t1 values(1,'ab');
insert into t1 values(2,'cd');

## covering index
explain select c2 from t1 where c2='ab';
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------------+
| id | select_type | table | type  | possible_keys | key  | key_len | ref   | rows | Extra       |
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------------+
|  1 | SIMPLE      | t1    | const | c2            | c2   | 13      | const |    1 | Using index |
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------------+
1 row in set (0.00 sec)

• 3) 无法转换

  除1，2两种情况外的collation, 例如utf8_general_ci

  此时从二级索引中获取不到key的完整信息，需要从主键索引上获取，因此不能走covering index

create table t1(c1 int primary key, c2 varchar(10) unique) engine =rocksdb character set utf8 collate utf8_general_ci;

 insert into t1 values(1,'ab');
 insert into t1 values(2,'cd');

## non-covering index
explain select c2 from t1 where c2='ab';
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------+
| id | select_type | table | type  | possible_keys | key  | key_len | ref   | rows | Extra |
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------+
|  1 | SIMPLE      | t1    | const | c2            | c2   | 33      | const |    1 | NULL  |
+----+-------------+-------+-------+---------------+------+---------+-------+------+-------+
1 row in set (0.00 sec)

此时的数据获取路径如下

0  myrocks::ha_rocksdb::convert_record_from_storage_format 
1  myrocks::ha_rocksdb::get_row_by_rowid
2  myrocks::ha_rocksdb::get_row_by_rowid
3  myrocks::ha_rocksdb::read_row_from_secondary_key 
4  myrocks::ha_rocksdb::index_read_map_impl
5  myrocks::ha_rocksdb::index_read_map
6  handler::index_read_idx_map
7  handler::ha_index_read_idx_map
8  join_read_const
9  join_read_const_table
10 make_join_statistics
11 JOIN::optimize
12 mysql_execute_select
13 mysql_select
14 handle_select
15 execute_sqlcom_select
16 mysql_execute_command

转换的具体实现可以参考函数Rdb_field_packing::setup

MyRocks从索引读取数据时，不能仅通过keyread_only来判断是否可以使用covering index, 还需要判断是否存在collation列数据转换的问题，如果访问的列无法转换就不能使用covering index.

MyRocks会在value中存储covered_bitmap,表示索引列是否可以转换, read_row_from_secondary_key/secondary_index_read读取时会根据covered_bitmap来决定是否能使用covering index

bool Rdb_key_def::covers_lookup(TABLE *const table,
                                const rocksdb::Slice *const unpack_info,
                                const MY_BITMAP *const lookup_bitmap) const {
  DBUG_ASSERT(lookup_bitmap != nullptr);
  if (!use_covered_bitmap_format() || lookup_bitmap->bitmap == nullptr) {
    return false;
  }

  Rdb_string_reader unp_reader = Rdb_string_reader::read_or_empty(unpack_info);

  // Check if this unpack_info has a covered_bitmap
  const char *unpack_header = unp_reader.get_current_ptr();
  const bool has_covered_unpack_info =
      unp_reader.remaining_bytes() &&
      unpack_header[0] == RDB_UNPACK_COVERED_DATA_TAG;
  if (!has_covered_unpack_info ||
      !unp_reader.read(RDB_UNPACK_COVERED_HEADER_SIZE)) {
    return false;
  }

  MY_BITMAP covered_bitmap;
  my_bitmap_map covered_bits;
  bitmap_init(&covered_bitmap, &covered_bits, MAX_REF_PARTS, false);
  covered_bits = rdb_netbuf_to_uint16((const uchar *)unpack_header +
                                      sizeof(RDB_UNPACK_COVERED_DATA_TAG) +
                                      RDB_UNPACK_COVERED_DATA_LEN_SIZE);

  return bitmap_is_subset(lookup_bitmap, &covered_bitmap);
}

总结

MyRocks在字符collation上的限制需要引起我们关注，使用不当会影响查询效率。