Category Archives: General

MySQL 5.7 InnoDB Temporary Tablespace – but why?

So, recently we had a runaway query eat up all sorts of temporary table space on our machines. Several machines had several terabytes in their ibtmp1 file after this happened. So I set out to find out more about why the InnoDB temporary tablespace is used, why it is better than using regular files, which was what was used prior to MySQL 5.7, and how to make sure that runaway queries do not end up filling up disk space.

Unfortunately, the manual does not go into why ibtmp1 is better than one file per temporary query, which disappears once the query ends. There are a few sections to look into:

Temporary Table Undo Logs – has one paragraph that states that these are the undo logs for temporary tablespaces. Given that these are undo logs, my guess is that this makes MySQL more crash-safe. But that is just a guess.

There is also InnoDB Temporary Tablespace which is two paragraphs, with some more details, but again, no reasoning why.

And finally, the documentation for the innodb_temp_data_file_path system variable sheds a bit of light on the subject – It explains “Metadata about active InnoDB temporary tables is located in INFORMATION_SCHEMA.INNODB_TEMP_TABLE_INFO.”

There is a manual page on Innodb temp table info table as well, which shows both compressed and uncompressed tables – uncompressed tables live in the ibtmp tablespaces, and compressed temporary tables live in the .ibd tablespace of the compressed table in question – as shown in the “PER_TABLE_SPACE” and “IS_COMPRESSED” fields.

Sadly, the table does not give useful information such as which process ID or user is generating the data. And of course it is only the active temporary space usage at the time – if you have a large temporary tablespace but no active queries using the tablespace, INNODB_TEMP_TABLE_INFO is empty.

I can imagine a scenario with more than one long-running query using a lot of space in the temporary tablespace. But I do not see how the INNODB_TEMP_TABLE_INFO would help me determine any useful information as to which query I should kill. I guess it is useful to see if there is an active query currently using temporary tablespace, but when you have a large file with nothing in it, it is just that much more depressing.

# mysql INFORMATION_SCHEMA -p -e “SELECT * FROM INNODB_TEMP_TABLE_INFO”
Enter password:
# ls -rlth /var/lib/mysql/ibtmp1
-rw-r—– 1 mysql mysql 2.3T Oct 31 10:50 /var/lib/mysql/ibtmp1

What is an “unauthenticated user”?

Every so often we have a client worrying about unauthenticated users. For example, as part of the output of SHOW PROCESSLIST they will see:

+-----+----------------------+--------------------+------+---------+------+-------+------------------+
| Id  | User                 | Host               | db   | Command | Time | State | Info             |
+-----+----------------------+--------------------+------+---------+------+-------+------------------+
| 235 | unauthenticated user | 10.10.2.74:53216   | NULL | Connect | NULL | login | NULL             |
| 236 | unauthenticated user | 10.120.61.10:51721 | NULL | Connect | NULL | login | NULL             |
| 237 | user                 | localhost          | NULL | Query   | 0    | NULL  | show processlist |
+-----+----------------------+--------------------+------+---------+------+-------+------------------+

Who are these unauthenticated users, how do they get there, and why aren’t they authenticated?

The client-server handshake in MySQL is a 4-step process. Those familiar with mysql-proxy already know these steps, as there are four functions that a Lua script in mysql-proxy can override. The process is useful to know for figuring out exactly where a problem is when something breaks.

Step 1: Client sends connect request to server. There is no information here (as far as I can tell). However, it does mean that if you try to connect to a host and port of a mysqld server that is not available, you will get

ERROR 2003 (HY000): Can't connect to MySQL server on '[host]' (111)

Step 2: The server assigns a connection and sends back a handshake, which includes the server’s mysqld version, the thread id, the server host and port, the client host and port, and a “scramble buffer” (for salting authentication, I believe).

It is during Step 2 where the connections show up in SHOW PROCESSLIST. They have not been authenticated yet, but they are connected. If there are issues with authentication, connections will be stuck at this stage. Most often stuck connections are due to DNS not resolving properly, which the skip-name-resolve option will help with.

Step 3: Client sends authentication information, including the username, the password (salted and hashed) and default database to use. If the client sends an incorrect packet, or does not send authentication information within connect_timeout seconds, the server considers the connection aborted and increments its Aborted_connects status variable.

Step 4: Server sends back whether the authentication was successful or not. If the authentication was not successful, mysqld increments its Aborted_connects status variable and sends back an error message:

ERROR 1045 (28000): Access denied for user 'user'@'host' (using password: [YES/NO])

Hope this helps!