Best Practices of SQL Queries

Updated on 2024-05-07 GMT+08:00

View PDF

Based on the SQL execution mechanism and a large number of practices, SQL statements can be optimized by following certain rules to enable the database to execute SQL statements more quickly and obtain correct results.

Replace UNION with UNION ALL.
UNION eliminates duplicate rows while merging two result sets but UNION ALL merges the two result sets without deduplication. Therefore, replace UNION with UNION ALL if you are sure that the two result sets do not contain duplicate rows based on the service logic.
Add NOT NULL to the join columns.
If there are many NULL values in the JOIN columns, you can add the filter criterion IS NOT NULL to filter data in advance to improve the JOIN efficiency.

Convert NOT IN to NOT EXISTS.

nestloop anti join must be used to implement NOT IN, and hash anti join is required for NOT EXISTS. If no NULL value exists in the JOIN columns, NOT IN is equivalent to NOT EXISTS. Therefore, if you are sure that no NULL value exists, you can convert NOT IN to NOT EXISTS to generate hash join and to improve the query performance.

The statements for creating a foreign table are as follows:

DROP SCHEMA IF EXISTS no_in_to_no_exists_test CASCADE;
CREATE SCHEMA no_in_to_no_exists_test;
SET CURRENT_SCHEMA=no_in_to_no_exists_test;
CREATE TABLE t1(c1 int, c2 int, c3 int);
CREATE TABLE t2(d1 int, d2 int NOT NULL, d3 int);

The statement for implementing the query using NOT IN is as follows:

SELECT * FROM t1 WHERE  c1 NOT IN (SELECT d2 FROM t2);

The plan is as follows:

gaussdb=# EXPLAIN SELECT * FROM t1 WHERE  c1 NOT IN (SELECT d2 FROM t2);
                                   QUERY PLAN
----------------------------------------------------------------------------------
 Streaming (type: GATHER)  (cost=0.06..38.57 rows=3 width=12)
   Node/s: All datanodes
   ->  Nested Loop Anti Join  (cost=0.00..38.44 rows=3 width=12)
         Join Filter: ((t1.c1 = t2.d2) OR (t1.c1 IS NULL))
         ->  Seq Scan on t1  (cost=0.00..14.14 rows=30 width=12)
         ->  Materialize  (cost=0.00..18.08 rows=90 width=4)
               ->  Streaming(type: BROADCAST)  (cost=0.00..17.93 rows=90 width=4)
                     Spawn on: All datanodes
                     ->  Seq Scan on t2  (cost=0.00..14.14 rows=30 width=4)
(9 rows)

Because there is no null value in the t2.d2 column (the t2.d2 column is NOT NULL in the table definition), the query can be equivalently modified as follows:

     
        SELECT * FROM t1 WHERE  NOT EXISTS (SELECT * FROM t2 WHERE t1.c1=t2.d2);

The generated plan is as follows:

gaussdb=# EXPLAIN SELECT * FROM t1 WHERE  NOT EXISTS (SELECT * FROM t2 WHERE t1.c1=t2.d2);
                                  QUERY PLAN
-------------------------------------------------------------------------------
 Streaming (type: GATHER)  (cost=14.38..29.99 rows=3 width=12)
   Node/s: All datanodes
   ->  Hash Right Anti Join  (cost=14.32..29.86 rows=3 width=12)
         Hash Cond: (t2.d2 = t1.c1)
         ->  Streaming(type: REDISTRIBUTE)  (cost=0.00..15.49 rows=30 width=4)
               Spawn on: All datanodes
               ->  Seq Scan on t2  (cost=0.00..14.14 rows=30 width=4)
         ->  Hash  (cost=14.14..14.14 rows=29 width=12)
               ->  Seq Scan on t1  (cost=0.00..14.14 rows=30 width=12)
(9 rows)

Use hashagg.
If a plan involving groupAgg and SORT operations generated by the GROUP BY statement is poor in performance, you can set work_mem to a larger value to generate a hashagg plan, which does not require sorting and improves the performance.
Replace functions with CASE statements.
The GaussDB performance greatly deteriorates if a large number of functions are called. In this case, you can modify the pushdown functions to CASE statements.
Do not use functions or expressions for indexes.
Using functions or expressions for indexes stops indexing. Instead, it enables scanning on the full table.
Do not use != or <> operators, NULL, OR, or implicit parameter conversion in WHERE clauses.
Split complex SQL statements.
You can split an SQL statement into several ones and save the execution result to a temporary table if the SQL statement is too complex to be tuned using the solutions above, including but not limited to the following scenarios:
- The same subquery is involved in multiple SQL statements of a job and the subquery contains large amounts of data.
- Incorrect plan cost causes a small hash bucket of subquery. For example, the actual number of rows is 10 million, but only 1000 rows are in hash bucket.
- Functions such as substr and to_number cause incorrect measures for subqueries containing a large amount of data.
- BROADCAST subqueries are performed on large tables in multi-DN environment.

For details about optimization, see Typical SQL Optimization Methods.

Parent topic: Best Practices

Previous topic: Checking a Node Where a Table Resides

Next topic: Best Practices for Data Skew Query

Feedback

Was this page helpful?

Helpful Not helpful

Provide feedback

Thank you very much for your feedback. We will continue working to improve the documentation.See the reply and handling status in My Cloud VOC.

The system is busy. Please try again later.

Which of the following issues have you encountered?

Content is inconsistent with the product UI

Unclear descriptions

Lack of examples or code

Incorrect steps

Can't find what I need

Lack of best practices

Feedback (optional)

0/500

Select at least one type of issue, and enter your comments or suggestions.

Enter a maximum of 500 characters.

Submit Cancel

For any further questions, feel free to contact us through the chatbot.

Chatbot