Split Region

For each new table created in TiDB, one Region is segmented by default to store the data of this table. This default behavior is controlled by split-table in the configuration file. When the data in this Region exceeds the default Region size limit, the Region starts to split into two.

In the above case, because there is only one Region at the beginning, all write requests occur on the TiKV where the Region is located. If there are a large number of writes for the newly created table, hotspots are caused.

To solve the hotspot problem in the above scenario, TiDB introduces the pre-split function, which can pre-split multiple Regions for a certain table according to the specified parameters and scatter them to each TiKV node.

Usage of Split Region

There are two types of Split Region syntax:

SPLIT TABLE table_name [INDEX index_name] BETWEEN (lower_value) AND (upper_value) REGIONS region_num

BETWEEN lower_value AND upper_value REGIONS region_num defines the upper boundary, the lower boundary, and the Region amount. Then the current region will be evenly spilt into the number of regions (as specified in region_num) between the upper and lower boundaries.

SPLIT TABLE table_name [INDEX index_name] BY (value_list) [, (value_list)] ...

BY value_list… specifies a series of points manually, based on which the current Region is spilt. It is suitable for scenarios with unevenly distributed data.

The following example shows the result of the SPLIT statement:

+--------------------+----------------------+
| TOTAL_SPLIT_REGION | SCATTER_FINISH_RATIO |
+--------------------+----------------------+
| 4                  | 1.0                  |
+--------------------+----------------------+

• TOTAL_SPLIT_REGION: the number of newly split Regions.
• SCATTER_FINISH_RATIO: the completion rate of scattering for newly split Regions. 1.0 means that all Regions are scattered. 0.5 means that only half of the Regions are scattered and the rest are being scattered.

Split Table Region

The key of row data in each table is encoded by table_id and row_id. The format is as follows:

t[table_id]_r[row_id]

For example, when table_id is 22 and row_id is 11:

t22_r11

Row data in the same table have the same table_id, but each row has its unique row_id that can be used for Region split.

Even Split

Because row_id is an integer, the value of the key to be split can be calculated according to the specified lower_value, upper_value, and region_num. TiDB first calculates the step value (step = (upper_value - lower_value)/num). Then split will be done evenly per each "step" between lower_value and upper_value to generate the number of Regions as specified by num.

For example, if you want 16 evenly split Regions split from key rangeminInt64~maxInt64 for table t, you can use this statement:

SPLIT TABLE t BETWEEN (-9223372036854775808) AND (9223372036854775807) REGIONS 16;

This statement splits table t into 16 Regions between minInt64 and maxInt64. If the given primary key range is smaller than the specified one, for example, 0~1000000000, you can use 0 and 1000000000 take place of minInt64 and maxInt64 respectively to split Regions.

SPLIT TABLE t BETWEEN (0) AND (1000000000) REGIONS 16;

Uneven split

If the known data is unevenly distributed, and you want a Region to be split respectively in key ranges -inf ~ 10000, 10000 ~ 90000, and 90000 ~ +inf, you can achieve this by setting fixed points, as shown below:

SPLIT TABLE t BY (10000), (90000);

Split index Region

The key of the index data in the table is encoded by table_id, index_id, and the value of the index column. The format is as follows:

t[table_id]_i[index_id][index_value]

For example, when table_id is 22, index_id is 5, and index_value is abc:

t22_i5abc

The table_id and index_id of the same index data in one table is the same. To split index Regions, you need to split Regions based on index_value.

Even Spilt

The way to split index evenly works the same as splitting data evenly. However, calculating the value of step is more complicated, because index_value might not be an integer.

The values of upper and lower are encoded into a byte array firstly. After removing the longest common prefix of lower and upper byte array, the first 8 bytes of lower and upper are converted into the uint64 format. Then step = (upper - lower)/num is calculated. After that, the calculated step is encoded into a byte array, which is appended to the longest common prefix of the lower and upper byte array for index split. Here is an example:

If the column of the idx index is of the integer type, you can use the following SQL statement to split index data:

SPLIT TABLE t INDEX idx BETWEEN (-9223372036854775808) AND (9223372036854775807) REGIONS 16;

This statement splits the Region of index idx in table t into 16 Regions from minInt64 to maxInt64.

If the column of index idx1 is of varchar type, and you want to split index data by prefix letters.

SPLIT TABLE t INDEX idx1 BETWEEN ("a") AND ("z") REGIONS 25;

This statement splits index idx1 into 25 Regions from a~z. The range of Region 1 is [minIndexValue, b); the range of Region 2 is [b, c); … the range of Region 25 is [y, minIndexValue]. For the idx index, data with the a prefix is written into Region 1, and data with the b prefix is written into Region 2, and so on.

In the split method above, both data with the y and z prefixes are written into Region 25, because the upper bound is not z, but { (the character next to z in ASCII). Therefore, a more accurate split method is as follows:

SPLIT TABLE t INDEX idx1 BETWEEN ("a") AND ("{") REGIONS 26;

This statement splits index idx1 of the table t into 26 Regions from a~{. The range of Region 1 is [minIndexValue, b); the range of Region 2 is [b, c); … the range of Region 25 is [y, z), and the range of Region 26 is [z, maxIndexValue).

If the column of index idx2 is of time type like timestamp/datetime, and you want to split index Region by year:

SPLIT TABLE t INDEX idx2 BETWEEN ("2010-01-01 00:00:00") AND ("2020-01-01 00:00:00") REGIONS 10;

This statement splits the Region of index idx2 in table t into 10 Regions from 2010-01-01 00:00:00 to 2020-01-01 00:00:00. The range of Region 1 is [minIndexValue, 2011-01-01 00:00:00); the range of Region 2 is [2011-01-01 00:00:00, 2012-01-01 00:00:00) and so on.

If you want to split the index Region by day, see the following example:

SPLIT TABLE t INDEX idx2 BETWEEN ("2020-06-01 00:00:00") AND ("2020-07-01 00:00:00") REGIONS 30;

This statement splits the data of June 2020 of index idx2 in table t into 30 Regions, each Region representing 1 day.

Region split methods for other types of index columns are similar.

For data Region split of joint indexes, the only difference is that you can specify multiple columns values.

For example, index idx3 (a, b) contains 2 columns, with column a of timestamp type and column b int. If you just want to do a time range split according to column a, you can use the SQL statement for splitting time index of a single column. In this case, do not specify the value of column b in lower_value and upper_velue.

SPLIT TABLE t INDEX idx3 BETWEEN ("2010-01-01 00:00:00") AND ("2020-01-01 00:00:00") REGIONS 10;

Within the same range of time, if you want to do one more split according to column b column. Just specify the value for column b when splitting.

SPLIT TABLE t INDEX idx3 BETWEEN ("2010-01-01 00:00:00", "a") AND ("2010-01-01 00:00:00", "z") REGIONS 10;

This statement splits 10 Regions in the range of a~z according to the value of column b, with the same time prefix as column a. If the value specified for column a is different, the value of column b might not be used in this case.

Uneven Split

Index data can also be split by specified index values.

For example, there is idx4 (a,b), with column a of the varchar type and column b of the timestamp type.

SPLIT TABLE t1 INDEX idx4 BY ("a", "2000-01-01 00:00:01"), ("b", "2019-04-17 14:26:19"), ("c", "");

This statement specifies 3 values to split 4 Regions. The range of each Region is as follows:

region1  [ minIndexValue               , ("a", "2000-01-01 00:00:01"))
region2  [("a", "2000-01-01 00:00:01") , ("b", "2019-04-17 14:26:19"))
region3  [("b", "2019-04-17 14:26:19") , ("c", "")                   )
region4  [("c", "")                    , maxIndexValue               )

pre_split_regions

To have evenly split Regions when a table is created, it is recommended you use SHARD_ROW_ID_BITS together with PRE_SPLIT_REGIONS. When a table is created successfully, PRE_SPLIT_REGIONS pre-spilts tables into the number of Regions as specified by 2^(PRE_SPLIT_REGIONS).

The tidb_scatter_region global variable affects the behavior of PRE_SPLIT_REGIONS. This variable controls whether to wait for Regions to be pre-split and scattered before returning results after the table creation. If there are intensive writes after creating the table, you need to set the value of this variable to 1, then TiDB will not return the results to the client until all the Regions are split and scattered. Otherwise, TiDB writes the data before the scattering is completed, which will have a significant impact on write performance.

Example

create table t (a int, b int,index idx1(a)) shard_row_id_bits = 4 pre_split_regions=2;

After building the table, this statement splits 4 + 1 Regions for table t. 4 (2^2) Regions are used to save table row data, and 1 Region is for saving the index data of idx1.

The ranges of the 4 table Regions are as follows:

region1:   [ -inf      ,  1<<61 )
region2:   [ 1<<61     ,  2<<61 )
region3:   [ 2<<61     ,  3<<61 )
region4:   [ 3<<61     ,  +inf  )

MySQL compatibility

This statement is a TiDB extension to MySQL syntax.

See also

There are two SPLIT REGION related session variables: tidb_scatter_region, tidb_wait_split_region_finish and tidb_wait_split_region_timeout. For details, see TiDB specific system variables and syntax.