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. The former is what you typically do with tools like PMM, the latter is done with tools like New Relic. They are not mutually exclusive and they both are important, for different reasons.
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.
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. If query went to production, they must have a purpose. Well, chances are you’re wrong. In complex environments, normally in a production environment useless queries can be found. 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:
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
- We are looking for completely useless queries. But after that, you may want to see queries whose ratio between
SUM_ERRORSis bad. Or even queries that returned at least one error.
- We put a lower limit to
COUNT_STARto avoid false positives – queries that were executed by mistake, perhaps manually, and will never run again.
- The limit on
LAST_SEENis 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:
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
- See notes about the previous query.
- We need to filter by
DIGEST_TEXTbecause we are not interested in
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_TABLESif you have memory consumption problems.
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
ORDER BY EMPTY_QUERIES DESC;
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. If that’s the case, a comment could be added to indicate that the query is not useless, so the next time you check you won’t waste your time on it.
Many high level developers hate to optimise queries. 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.
- events_statements_summary_by_digest in MariaDB KnowledgeBase
- events_statements_summary_by_digest in MySQL documentation
- MySQL/MariaDB: use SQL properly to run less queries – About merging multiple queries into one
When we put our hands on a new OLTP environment, we find these queries more often than not. Eliminating them helps reducing the resource usage and increase the speed of certain.