MariaDB starting with 10.1.0

The ANALYZE statement command was introduced in MariaDB 10.1.0.

Command description

The ANALYZE statement command is similar to EXPLAIN statement. ANALYZE statement will invoke the optimizer, execute the statement, and then produce EXPLAIN output instead of the result set. The EXPLAIN output will be annotated with statistics from statement execution.

This lets one check how close the optimizer's estimates about the query plan are to the reality.

The syntax is

ANALYZE explainable_statement;

where the statement is any statement for which one can run EXPLAIN.

Command output

Consider an example:

MariaDB> analyze select * from tbl1 where key1 between 10 and 200 and col1 like 'foo%'\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: tbl1
         type: range
possible_keys: key1
          key: key1
      key_len: 5
          ref: NULL
         rows: 181
       r_rows: 181
     filtered: 100.00
   r_filtered: 10.50
        Extra: Using index condition; Using where

Compared to EXPLAIN, ANALYZE produces two extra columns:

  • r_rows is an observation-based counterpart of the rows column. It shows how many rows were actually read from the table.
  • r_filtered is an observation-based counterpart of the filtered column. It shows which fraction of rows was left after applying the WHERE condition.

Interpreting the output

Joins

Let's consider a more complicated example.

analyze select * 
from orders, customer 
where 
  customer.c_custkey=orders.o_custkey and 
  customer.c_acctbal < 0 and 
  orders.o_totalprice > 200*1000
+----+-------------+----------+------+---------------+-------------+---------+--------------------+--------+--------+----------+------------+-------------+
| id | select_type | table    | type | possible_keys | key         | key_len | ref                | rows   | r_rows | filtered | r_filtered | Extra       |
+----+-------------+----------+------+---------------+-------------+---------+--------------------+--------+--------+----------+------------+-------------+
|  1 | SIMPLE      | customer | ALL  | PRIMARY,...   | NULL        | NULL    | NULL               | 149095 | 150000 |    18.08 |       9.13 | Using where |
|  1 | SIMPLE      | orders   | ref  | i_o_custkey   | i_o_custkey | 5       | customer.c_custkey |      7 |     10 |   100.00 |      30.03 | Using where |
+----+-------------+----------+------+---------------+-------------+---------+--------------------+--------+--------+----------+------------+-------------+

Here, one can see that

  • For table customer, customer.rows=149095, customer.r_rows=150000. The estimate for number of rows we will read was fairly precise
  • customer.filtered=18.08, customer.r_filtered=9.13. The optimizer somewhat overestimated the number of records that will match selectivity of condition attached to `customer` table. (in general, when you have a full scan and r_filtered is less than 15%, it's time to consider adding an appropriate index)
  • For table orders, orders.rows=7, orders.r_rows=10. This means that on average, there are 7 orders for a given c_custkey, but in our case there were 10, which is close to the expectation (when this number is consistently far from the expectation, maybe it's time to run ANALYZE TABLE, or even edit the table statistics manually to get better query plans)
  • orders.filtered=100, orders.r_filtered=30.03. The optimizer didn't have any way to estimate which fraction of records will be left after it checks the condition that is attached to table orders (it's orders.o_totalprice > 200*1000). So, it used 100%. In reality, it is 30%. 30% is typically not selective enough to warrant adding new indexes. For joins with many tables, it might be worth to collect and use column statistics for columns in question, this may help the optimizer to pick a better query plan.

Meaning of NULL in r_rows and r_filtered

Let's modify the previous example slightly

analyze select * 
from orders, customer 
where 
  customer.c_custkey=orders.o_custkey and 
  customer.c_acctbal < -0 and customer.c_comment like '%foo%'
  orders.o_totalprice > 200*1000
+----+-------------+----------+------+---------------+-------------+---------+--------------------+--------+--------+----------+------------+-------------+
| id | select_type | table    | type | possible_keys | key         | key_len | ref                | rows   | r_rows | filtered | r_filtered | Extra       |
+----+-------------+----------+------+---------------+-------------+---------+--------------------+--------+--------+----------+------------+-------------+
|  1 | SIMPLE      | customer | ALL  | PRIMARY,...   | NULL        | NULL    | NULL               | 149095 | 150000 |    18.08 |       0.00 | Using where |
|  1 | SIMPLE      | orders   | ref  | i_o_custkey   | i_o_custkey | 5       | customer.c_custkey |      7 |   NULL |   100.00 |       NULL | Using where |
+----+-------------+----------+------+---------------+-------------+---------+--------------------+--------+--------+----------+------------+-------------+

Here, one can see that orders.r_rows=NULL and orders.r_filtered=NULL. This means that table orders was not scanned even once. Indeed, we can also see customer.r_filtered=0.00. This shows that a part of WHERE attached to table `customer` was never satisfied (or, satisfied in less than 0.01% of cases).

Notes

  • ANALYZE UPDATE or ANALYZE DELETE will actually make updates/deletes (ANALYZE SELECT will perform the select operation and then discard the resultset).
  • PostgreSQL has a similar command, EXPLAIN ANALYZE.
  • EXPLAIN in the slow query log feature allows MariaDB to have ANALYZE output of slow queries printed into the slow query log (see MDEV-6388).
  • ANALYZE FORMAT=JSON produces more information in JSON form.

See also

  • Jira task for ANALYZE statement, MDEV-406

Comments

Comments loading...