MySQL :: MySQL 5.7 Reference Manual :: 8.2.1.12 Condition Filtering (original) (raw)

8.2.1.12 Condition Filtering

In join processing, prefix rows are those rows passed from one table in a join to the next. In general, the optimizer attempts to put tables with low prefix counts early in the join order to keep the number of row combinations from increasing rapidly. To the extent that the optimizer can use information about conditions on rows selected from one table and passed to the next, the more accurately it can compute row estimates and choose the best execution plan.

Without condition filtering, the prefix row count for a table is based on the estimated number of rows selected by theWHERE clause according to whichever access method the optimizer chooses. Condition filtering enables the optimizer to use other relevant conditions in theWHERE clause not taken into account by the access method, and thus improve its prefix row count estimates. For example, even though there might be an index-based access method that can be used to select rows from the current table in a join, there might also be additional conditions for the table in the WHERE clause that can filter (further restrict) the estimate for qualifying rows passed to the next table.

A condition contributes to the filtering estimate only if:

In EXPLAIN output, therows column indicates the row estimate for the chosen access method, and the filtered column reflects the effect of condition filtering.filtered values are expressed as percentages. The maximum value is 100, which means no filtering of rows occurred. Values decreasing from 100 indicate increasing amounts of filtering.

The prefix row count (the number of rows estimated to be passed from the current table in a join to the next) is the product of the rows andfiltered values. That is, the prefix row count is the estimated row count, reduced by the estimated filtering effect. For example, if rows is 1000 and filtered is 20%, condition filtering reduces the estimated row count of 1000 to a prefix row count of 1000 × 20% = 1000 × .2 = 200.

Consider the following query:

SELECT *
  FROM employee JOIN department ON employee.dept_no = department.dept_no
  WHERE employee.first_name = 'John'
  AND employee.hire_date BETWEEN '2018-01-01' AND '2018-06-01';

Suppose that the data set has these characteristics:

employee.first_name = 'John'  
employee.hire_date BETWEEN '2018-01-01' AND '2018-06-01'  
employee.first_name = 'John'  
AND employee.hire_date BETWEEN '2018-01-01' AND '2018-06-01'

Without condition filtering,EXPLAIN produces output like this:

+----+------------+--------+------------------+---------+---------+------+----------+
| id | table      | type   | possible_keys    | key     | ref     | rows | filtered |
+----+------------+--------+------------------+---------+---------+------+----------+
| 1  | employee   | ref    | name,h_date,dept | name    | const   | 8    | 100.00   |
| 1  | department | eq_ref | PRIMARY          | PRIMARY | dept_no | 1    | 100.00   |
+----+------------+--------+------------------+---------+---------+------+----------+

For employee, the access method on thename index picks up the 8 rows that match a name of 'John'. No filtering is done (filtered is 100%), so all rows are prefix rows for the next table: The prefix row count isrows × filtered = 8 × 100% = 8.

With condition filtering, the optimizer additionally takes into account conditions from the WHERE clause not taken into account by the access method. In this case, the optimizer uses heuristics to estimate a filtering effect of 16.31% for the BETWEEN condition on employee.hire_date. As a result, EXPLAIN produces output like this:

+----+------------+--------+------------------+---------+---------+------+----------+
| id | table      | type   | possible_keys    | key     | ref     | rows | filtered |
+----+------------+--------+------------------+---------+---------+------+----------+
| 1  | employee   | ref    | name,h_date,dept | name    | const   | 8    | 16.31    |
| 1  | department | eq_ref | PRIMARY          | PRIMARY | dept_no | 1    | 100.00   |
+----+------------+--------+------------------+---------+---------+------+----------+

Now the prefix row count is rows ×filtered = 8 × 16.31% = 1.3, which more closely reflects actual data set.

Normally, the optimizer does not calculate the condition filtering effect (prefix row count reduction) for the last joined table because there is no next table to pass rows to. An exception occurs forEXPLAIN: To provide more information, the filtering effect is calculated for all joined tables, including the last one.

To control whether the optimizer considers additional filtering conditions, use thecondition_fanout_filter flag of the optimizer_switch system variable (seeSection 8.9.2, “Switchable Optimizations”). This flag is enabled by default but can be disabled to suppress condition filtering (for example, if a particular query is found to yield better performance without it).

If the optimizer overestimates the effect of condition filtering, performance may be worse than if condition filtering is not used. In such cases, these techniques may help:

SET optimizer_switch = 'condition_fanout_filter=off';