MySQL/MariaDB: run less queries! (thanks to performance_schema)

Sisyphus sculpture
Nowadays probably Sisyphus runs SQL queries that return infinite rows,
and loops over them.
This sculpture is in Highgate Cemetary, in London.

Do you want to improve your application’s and database server’s performance? Of course one of the most important things to do is to optimise the most impacting queries.

You have various ways to find them, depending on what exactly you mean by “most impacting”. You can have an approach oriented to your systems workload, or an approach oriented to your users experience. On the former a lot has been written in the MySQL community. On the latter, I’ll write something in the future. They are not mutually exclusive and they both are important.

But there is a third important approach: get rid of the queries that serve no purpose. Because they add work to the server, they probably make requests slower, they could lock rows, and they could consume enough CPU and memory to increase the cost of our servers.

MySQL/MariaDB performance_schema will help find those queries. It needs to be enabled.

Performance Schema and useless queries

I know, I know: you think you don’t have such queries. Well, chances are, you’re wrong. In complex environments, normally there are useless queries. Maybe they were useful when they were written, but at some point they lost their purpose and no one removed them. Maybe they used to return data that don’t exist anymore, because a feature was removed or rows changed. Maybe they stopped working correctly after a table was modified. Maybe they never worked but no one ever noticed that because testing is insufficient.

Whatever the reason… let’s find them.

Queries that always return an error

Many queries return an error, but… only in certain situations. For example, some queries use the IN (1, 2, 3...) syntax, which is fine. But sometimes the list of values is empty, and the IN () syntax is not valid. Other examples include unquoted strings: if the string is a number, the query will work. But if it’s alphanumeric, it will fail. Of course those queries should be found and fixed.

But some queries always return an error. This happens if their syntax is not valid, and the syntax error does not depend on parameters. Or maybe parameters always cause errors (in the example above, maybe strings are always alphanumeric).

To find such queries, run:

SELECT
        DIGEST_TEXT,
        COUNT_STAR,
        FIRST_SEEN,
        LAST_SEEN
    FROM performance_schema.events_statements_summary_by_digest
    WHERE
        DIGEST_TEXT IS NOT NULL
        AND SUM_ERRORS = COUNT_STAR
        AND COUNT_STAR > 20
        AND LAST_SEEN > (NOW() - INTERVAL 1 MONTH)
    ORDER BY COUNT_STAR DESC
    LIMIT 10
\G

Some notes:

  • We are looking for completely useless queries. But after that, you may want to see queries whose ratio between COUNT_STAR and SUM_ERRORS is bad. Or even queries that returned at least one error.
  • We put a lower limit to COUNT_STAR to avoid false positives – queries that were executed by mistake, perhaps manually, and will never run again.
  • The limit on LAST_SEEN is also important. We want to exclude queries that don’t run anymore, or don’t fail anymore because the conditions that caused the error have changed.
  • Queries executed more times are obviously more important to erase.

Queries that never return rows

Believe it or not, I often see queries that always return an empty resultset. If you find such queries, you should at least suspect that they are useless and can be removed.

All queries consume some resources. But queries that don’t return rows are much more important to remove than queries that fail. The reason is that these queries are executed. They return nothing, but they could examine a big amount of rows, materialise them in memory, or keep them locked.

Find them in this way:

SELECT *
    FROM performance_schema.events_statements_summary_by_digest
    WHERE
        (
               TRIM(DIGEST_TEXT) LIKE 'SELECT%'
            OR TRIM(DIGEST_TEXT) LIKE 'CREATE%TABLE%SELECT%'
            OR TRIM(DIGEST_TEXT) LIKE 'DELETE%'
            OR TRIM(DIGEST_TEXT) LIKE 'UPDATE%'
            OR TRIM(DIGEST_TEXT) LIKE 'REPLACE%'
        )
        AND SUM_ROWS_SENT = 0
        AND SUM_ROWS_AFFECTED = 0
        AND COUNT_STAR > 20
        AND LAST_SEEN > (NOW() - INTERVAL 1 MONTH)
    ORDER BY SUM_ROWS_EXAMINED DESC
    LIMIT 10
\G

Notes:

  • See notes about the previous query.
  • We need to filter by DIGEST_TEXT because we are not interested in CREATE, DROP, SET, commands, and many others.
  • Writes (UPDATE, etc) do not return rows, but they do affect some rows (usually). We exclude them with SUM_ROWS_AFFECTED = 0.
  • Here I ordered the results by the number of rows examined because I consider it very significant about the amount of work cause by the query. Depending on which problems you have, you may want to filter by something else, for example SUM_CREATED_TMP_TABLES if you have memory consumption problems.

Empty queries

TL;DR: you hardly care about this.

A last type of useless queries is… empty queries, or queries with no text. A typical way to run them is to add a semicolon (;) at the end of a query. by mistake:

SELECT * FROM information_schema.TABLES;;

You may think it only happens when running queries manually, but occasionally I saw many empty queries in production. I suspect it’s caused by certain ORMs bugs.

However, finding empty queries could be time consuming for developers, and it’s hardly worth the effort. You may try to ask them to do so if the number of such queries is high (like, they happen multiple times a second), but I’m mentioning empty queries mainly for the sake of completeness.

We can spot empty queries more easily with the user_statistics plugin, by Percona – but it’s not necessary:

SHOW GLOBAL STATUS LIKE 'Com_empty_query';
SET GLOBAL userstat := 1;
SELECT USER, EMPTY_QUERIES
    FROM USER_STATISTICS
    ORDER BY EMPTY_QUERIES DESC;

See also

Reference:

Related articles:

Related services:

What to do next

Talk to developers. There could be reasons why a query must not be eliminated. If the syntax is wrong, it can be fixed. If it returns no rows, that could be expected for some reason.

High level developers hate database problems. But this particular type of problems has more chances to attract their attention. Their code does something potentially useless or wrong, and they usually will want to know why and fix the problem.

Conclusions

When I put my hands on a new OLTP environment, I find these queries more often than not. Eliminating them helps reducing the resource usage, and there fore the costs, and possibly the costs of cloud services.

Do you have better or alternative ways to find such queries? Did you find any mistake? Please comment! Comments to my articles make this website not just one person’s texts, but also a more collective knowledge base.

Toodle pip,
Federico

Photo credit

Leave a Reply

Your email address will not be published. Required fields are marked *