Skip to content

Commit

Permalink
cherry pick #3068 to release-4.0 (#3279)
Browse files Browse the repository at this point in the history 
Signed-off-by: ti-srebot <ti-srebot@pingcap.com>

Co-authored-by: tiancaiamao <tiancaiamao@gmail.com>
Co-authored-by: Keke Yi <40977455+yikeke@users.noreply.github.com>
  • Loading branch information
@ti-srebot @tiancaiamao @yikeke
3 people authored Jul 15, 2020
1 parent 18dbea9 commit dd08c96
Show file tree
Hide file tree
Showing 3 changed files with 168 additions and 0 deletions.
2 changes: 2 additions & 0 deletions TOC.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -103,6 +103,8 @@
+ SQL Optimization
+ [SQL Optimization Process](/sql-optimization-concepts.md)
+ Logic Optimization
+ [Subquery Related Optimizations](/subquery-optimization.md)
+ [Decorrelation of Correlated Subquery](/correlated-subquery-optimization.md)
+ [Predicates Push Down](/predicates-push-down.md)
+ [Join Reorder](/join-reorder.md)
+ Physical Optimization
Expand Down
78 changes: 78 additions & 0 deletions correlated-subquery-optimization.md
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,78 @@
---
title: Decorrelation of Correlated Subquery
summary: Understand how to decorrelate correlated subqueries.
---

# Decorrelation of Correlated Subquery

[Subquery related optimizations](/subquery-optimization.md) describes how TiDB handles subqueries when there are no correlated columns. Because decorrelation of correlated subquery is complex, this article introduces some simple scenarios and the scope to which the optimization rule applies.

## Introduction

Take `select * from t1 where t1.a < (select sum(t2.a) from t2 where t2.b = t1.b)` as an example. The subquery `t1.a < (select sum(t2.a) from t2 where t2.b = t1.b)` here refers to the correlated column in the query condition `t2.b=t1.b`, this condition happens to be an equivalent condition, so the query can be rewritten as `select t1.* from t1, (select b, sum(a) sum_a from t2 group by b) t2 where t1.b = t2.b and t1.a < t2.sum_a;`. In this way, a correlated subquery is rewritten into `JOIN`.

The reason why TiDB needs to do this rewriting is that the correlated subquery is bound to its external query result every time the subquery is executed. In the above example, if `t1.a` has 10 million values, this subquery would repeat 10 million times, because the condition `t2.b=t1.b` varies with the value of `t1.a`. When the correlation is lifted somehow, this subquery would execute only once.

## Restrictions

The disadvantage of this rewriting is that when the correlation is not lifted, the optimizer can use the index on the correlated column. That is, although this subquery may repeat many times, the index can be used to filter data each time. After using the rewriting rule, the position of the correlated column usually changes. Although the subquery is only executed once, the single execution time would be longer than that without decorrelation.

Therefore, when there are few external values, do not perform decorrelation, because it may bring better execution performance. At present, this optimization can be disabled by setting `subquery decorrelation` optimization rules in [blocklist of optimization rules and expression pushdown](/blocklist-control-plan.md).

## Example

{{< copyable "sql" >}}

```sql
create table t1(a int, b int);
create table t2(a int, b int, index idx(b));
explain select * from t1 where t1.a < (select sum(t2.a) from t2 where t2.b = t1.b);
```

```sql
+----------------------------------+----------+-----------+---------------+-----------------------------------------------------------------------------------------+
| id | estRows | task | access object | operator info |
+----------------------------------+----------+-----------+---------------+-----------------------------------------------------------------------------------------+
| HashJoin_11 | 9990.00 | root | | inner join, equal:[eq(test.t1.b, test.t2.b)], other cond:lt(cast(test.t1.a), Column#7) |
| ├─HashAgg_23(Build) | 7992.00 | root | | group by:test.t2.b, funcs:sum(Column#8)->Column#7, funcs:firstrow(test.t2.b)->test.t2.b |
| │ └─TableReader_24 | 7992.00 | root | | data:HashAgg_16 |
| │ └─HashAgg_16 | 7992.00 | cop[tikv] | | group by:test.t2.b, funcs:sum(test.t2.a)->Column#8 |
| │ └─Selection_22 | 9990.00 | cop[tikv] | | not(isnull(test.t2.b)) |
| │ └─TableFullScan_21 | 10000.00 | cop[tikv] | table:t2 | keep order:false, stats:pseudo |
| └─TableReader_15(Probe) | 9990.00 | root | | data:Selection_14 |
| └─Selection_14 | 9990.00 | cop[tikv] | | not(isnull(test.t1.b)) |
| └─TableFullScan_13 | 10000.00 | cop[tikv] | table:t1 | keep order:false, stats:pseudo |
+----------------------------------+----------+-----------+---------------+-----------------------------------------------------------------------------------------+

```

The above is an example where the optimization takes effect. `HashJoin_11` is a normal `inner join`.

Then, turn off the subquery decorrelation rules:

{{< copyable "sql" >}}

```sql
insert into mysql.opt_rule_blacklist values("decorrelate");
admin reload opt_rule_blacklist;
explain select * from t1 where t1.a < (select sum(t2.a) from t2 where t2.b = t1.b);
```

```sql
+----------------------------------------+----------+-----------+------------------------+------------------------------------------------------------------------------+
| id | estRows | task | access object | operator info |
+----------------------------------------+----------+-----------+------------------------+------------------------------------------------------------------------------+
| Projection_10 | 10000.00 | root | | test.t1.a, test.t1.b |
| └─Apply_12 | 10000.00 | root | | CARTESIAN inner join, other cond:lt(cast(test.t1.a), Column#7) |
| ├─TableReader_14(Build) | 10000.00 | root | | data:TableFullScan_13 |
| │ └─TableFullScan_13 | 10000.00 | cop[tikv] | table:t1 | keep order:false, stats:pseudo |
| └─MaxOneRow_15(Probe) | 1.00 | root | | |
| └─HashAgg_27 | 1.00 | root | | funcs:sum(Column#10)->Column#7 |
| └─IndexLookUp_28 | 1.00 | root | | |
| ├─IndexRangeScan_25(Build) | 10.00 | cop[tikv] | table:t2, index:idx(b) | range: decided by [eq(test.t2.b, test.t1.b)], keep order:false, stats:pseudo |
| └─HashAgg_17(Probe) | 1.00 | cop[tikv] | | funcs:sum(test.t2.a)->Column#10 |
| └─TableRowIDScan_26 | 10.00 | cop[tikv] | table:t2 | keep order:false, stats:pseudo |
+----------------------------------------+----------+-----------+------------------------+------------------------------------------------------------------------------+
```

After disabling the subquery decorrelation rule, you can see `range: decided by [eq(test.t2.b, test.t1.b)]` in `operator info` of `IndexRangeScan_25(Build)`. It means that the decorrelation of correlated subquery is not performed and TiDB uses the index range query.
88 changes: 88 additions & 0 deletions subquery-optimization.md
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,88 @@
---
title: Subquery Related Optimizations
summary: Understand optimizations related to subqueries.
---

# Subquery Related Optimizations

This article mainly introduces subquery related optimizations.

Subqueries usually appear in the following situations:

- `NOT IN (SELECT ... FROM ...)`
- `NOT EXISTS (SELECT ... FROM ...)`
- `IN (SELECT ... FROM ..)`
- `EXISTS (SELECT ... FROM ...)`
- `... >/>=/</<=/=/!= (SELECT ... FROM ...)`

Sometimes a subquery contains non-subquery columns, such as `select * from t where t.a in (select * from t2 where t.b=t2.b)`. The `t.b` column in the subquery does not belong to the subquery, it is introduced from the outside of the subquery. This kind of subquery is usually called a "correlated subquery", and the externally introduced column is called a "correlated column". For optimizations about correlated subquery, see [Decorrelation of correlated subquery](/correlated-subquery-optimization.md). This article focuses on subqueries that do not involve correlated columns.

By default, subqueries use `semi join` mentioned in [Understanding TiDB Execution Plan](/query-execution-plan.md) as the execution method. For some special subqueries, TiDB do some logical rewrite to get better performance.

## `... < ALL (SELECT ... FROM ...)` or `... > ANY (SELECT ... FROM ...)`

In this case, `ALL` and `ANY` can be replaced by `MAX` and `MIN`. When the table is empty, the result of `MAX(EXPR)` and `MIN(EXPR)` is NULL. It works the same when the result of `EXPR` contains `NULL`. Whether the result of `EXPR` contains `NULL` may affect the final result of the expression, so the complete rewrite is given in the following form:

- `t.id < all (select s.id from s)` is rewritten as `t.id < min(s.id) and if(sum(s.id is null) != 0, null, true)`
- `t.id < any (select s.id from s)` is rewritten as `t.id < max(s.id) or if(sum(s.id is null) != 0, null, false)`

## `... != ANY (SELECT ... FROM ...)`

In this case, if all the values from the subquery are distinct, it is enough to compare the query with them. If the number of different values in the subquery is more than one, then there must be inequality. Therefore, such subqueries can be rewritten as follows:

- `select * from t where t.id != any (select s.id from s)` is rewritten as `select t.* from t, (select s.id, count(distinct s.id) as cnt_distinct from s) where (t.id != s.id or cnt_distinct > 1)`

## `... = ALL (SELECT ... FROM ...)`

In this case, when the number of different values in the subquery is more than one, then the result of this expression must be false. Therefore, such subquery is rewritten into the following form in TiDB:

- `select * from t where t.id = all (select s.id from s)` is rewritten as `select t.* from t, (select s.id, count(distinct s.id) as cnt_distinct from s ) where (t.id = s.id and cnt_distinct <= 1)`

## `... IN (SELECT ... FROM ...)`

In this case, the subquery of `IN` is rewritten into `SELECT ... FROM ... GROUP ...`, and then rewritten into the normal form of `JOIN`.

For example, `select * from t1 where t1.a in (select t2.a from t2)` is rewritten as `select t1.* from t1, (select distinct(a) a from t2) t2 where t1.a = t2. The form of a`. The `DISTINCT` attribute here can be eliminated automatically if `t2.a` has the `UNIQUE` attribute.

{{< copyable "sql" >}}

```sql
explain select * from t1 where t1.a in (select t2.a from t2);
```

```sql
+------------------------------+---------+-----------+------------------------+----------------------------------------------------------------------------+
| id | estRows | task | access object | operator info |
+------------------------------+---------+-----------+------------------------+----------------------------------------------------------------------------+
| IndexJoin_12 | 9990.00 | root | | inner join, inner:TableReader_11, outer key:test.t2.a, inner key:test.t1.a |
| ├─HashAgg_21(Build) | 7992.00 | root | | group by:test.t2.a, funcs:firstrow(test.t2.a)->test.t2.a |
| │ └─IndexReader_28 | 9990.00 | root | | index:IndexFullScan_27 |
| │ └─IndexFullScan_27 | 9990.00 | cop[tikv] | table:t2, index:idx(a) | keep order:false, stats:pseudo |
| └─TableReader_11(Probe) | 1.00 | root | | data:TableRangeScan_10 |
| └─TableRangeScan_10 | 1.00 | cop[tikv] | table:t1 | range: decided by [test.t2.a], keep order:false, stats:pseudo |
+------------------------------+---------+-----------+------------------------+----------------------------------------------------------------------------+
```

This rewrite gets better performance when the `IN` subquery is relatively small and the external query is relatively large, because without rewriting, using `index join` with t2 as the driving table is impossible. However, the disadvantage is that when the aggregation cannot be automatically eliminated during the rewrite and the `t2` table is relatively large, this rewrite affects the performance of the query. Currently, the variable [tidb\_opt\_insubq\_to\_join\_and\_agg](/system-variables.md#tidb_opt_insubq_to_join_and_agg) is used to control this optimization. When this optimization is not suitable, you can manually disable it.

## `EXISTS` subquery and `... >/>=/</<=/=/!= (SELECT ... FROM ...)`

At present, for a subquery in such scenarios, if the subquery is not a correlated subquery, TiDB evaluates it in advance in the optimization stage, and directly replaces it with a result set. As shown in the figure below, the `EXISTS` subquery is evaluated to `TRUE` in the optimization stage in advance, so it does not show in the final execution result.

{{< copyable "sql" >}}

```sql
create table t1(a int);
create table t2(a int);
insert into t2 values(1);
explain select * from t where exists (select * from t2);
```

```sql
+------------------------+----------+-----------+---------------+--------------------------------+
| id | estRows | task | access object | operator info |
+------------------------+----------+-----------+---------------+--------------------------------+
| TableReader_12 | 10000.00 | root | | data:TableFullScan_11 |
| └─TableFullScan_11 | 10000.00 | cop[tikv] | table:t | keep order:false, stats:pseudo |
+------------------------+----------+-----------+---------------+--------------------------------+
```

0 comments on commit dd08c96

Please sign in to comment.