Distinct Optimization

This document introduces the distinct optimization in the TiDB query optimizer, including SELECT DISTINCT and DISTINCT in the aggregate functions.

DISTINCT modifier in SELECT statements

The DISTINCT modifier specifies removal of duplicate rows from the result set. SELECT DISTINCT is transformed to GROUP BY, for example:

mysql> explain SELECT DISTINCT a from t;
+--------------------------+---------+-----------+---------------+-------------------------------------------------------+
| id                       | estRows | task      | access object | operator info                                         |
+--------------------------+---------+-----------+---------------+-------------------------------------------------------+
| HashAgg_6                | 2.40    | root      |               | group by:test.t.a, funcs:firstrow(test.t.a)->test.t.a |
| └─TableReader_11         | 3.00    | root      |               | data:TableFullScan_10                                 |
|   └─TableFullScan_10     | 3.00    | cop[tikv] | table:t       | keep order:false, stats:pseudo                        |
+--------------------------+---------+-----------+---------------+-------------------------------------------------------+
3 rows in set (0.00 sec)

DISTINCT option in aggregate functions

Usually, aggregate functions with the DISTINCT option is executed in the TiDB layer in a single-threaded execution model.

The tidb_opt_distinct_agg_push_down system variable or the distinct-agg-push-down configuration item in TiDB controls whether to rewrite the distinct aggregate queries and push them to the TiKV or TiFlash Coprocessor.

The tidb_opt_distinct_agg_push_down system variable in TiDB controls whether to rewrite the distinct aggregate queries and push them to the TiKV or TiFlash Coprocessor.

Take the following queries as an example of this optimization. tidb_opt_distinct_agg_push_down is disabled by default, which means the aggregate functions are executed in the TiDB layer. After enabling this optimization by setting its value to 1, the distinct a part of count(distinct a) is pushed to TiKV or TiFlash Coprocessor: there is a HashAgg_5 to remove the duplicated values on column a in the TiKV Coprocessor. It might reduce the computation overhead of HashAgg_8 in the TiDB layer.

mysql> desc select count(distinct a) from test.t;
+-------------------------+----------+-----------+---------------+------------------------------------------+
| id                      | estRows  | task      | access object | operator info                            |
+-------------------------+----------+-----------+---------------+------------------------------------------+
| StreamAgg_6             | 1.00     | root      |               | funcs:count(distinct test.t.a)->Column#4 |
| └─TableReader_10        | 10000.00 | root      |               | data:TableFullScan_9                     |
|   └─TableFullScan_9     | 10000.00 | cop[tikv] | table:t       | keep order:false, stats:pseudo           |
+-------------------------+----------+-----------+---------------+------------------------------------------+
3 rows in set (0.01 sec)

mysql> set session tidb_opt_distinct_agg_push_down = 1;
Query OK, 0 rows affected (0.00 sec)

mysql> desc select count(distinct a) from test.t;
+---------------------------+----------+-----------+---------------+------------------------------------------+
| id                        | estRows  | task      | access object | operator info                            |
+---------------------------+----------+-----------+---------------+------------------------------------------+
| HashAgg_8                 | 1.00     | root      |               | funcs:count(distinct test.t.a)->Column#3 |
| └─TableReader_9           | 1.00     | root      |               | data:HashAgg_5                           |
|   └─HashAgg_5             | 1.00     | cop[tikv] |               | group by:test.t.a,                       |
|     └─TableFullScan_7     | 10000.00 | cop[tikv] | table:t       | keep order:false, stats:pseudo           |
+---------------------------+----------+-----------+---------------+------------------------------------------+
4 rows in set (0.00 sec)