Possible Mitigation for system performance after 'meltdown' bug patching

This is an interesting read on database performance (not specific to MySQL) after patching to secure data on multi tenant system.

https://blog.appoptics.com/visualizing-meltdown-aws/

Copied from the above article:

Applications that make frequent systems calls to read/write data either over network sockets or from disk systems will need to be better tuned for batching. Incurring small I/O operations is now more costly, and engineers will need to optimize their code to reduce the frequency of such calls. Finding the sweet spot between larger batch sizes and latency is difficult and will require software that adapts for multiple variables simultaneously. It was promising to see that the Kafka consumer libraries were able to optimize for this dynamically as network call latency increased.

Using triggers to audit database changes

I cannot count on people on my team to inform me of changes they are making to our databases so I've had to create some very basic monitoring which checks the status of certain system variables at a certain interval and saves this into a few tables. I've added triggers to these tables to audit changes and deletions.

This is an old but useful (and free) way of keeping historical information on changes in a MySQL databases. The triggers copy the old and the new value to a generic change log table (sometimes called audit log). In this manner multiple tables can use the same change log table.

In this example I have two tables, one called MyCluster and one called Tag. The Tag table uses key/value to store data while the MyCluster has specific attributes stored in columns for each cluster. Here is an example table structure for the two tables (actual table has many more columns):

CREATE TABLE `MyCluster` (
  `Profile` varchar(100) NOT NULL COMMENT 'Account where data came from',
  `DBClusterIdentifier` varchar(255) NOT NULL DEFAULT '',
  `Endpoint` varchar(255) DEFAULT NULL COMMENT 'Cluster Writer End Point Address',
  `EndpointIPAddress` varchar(50) DEFAULT NULL COMMENT 'Cluster Writer End Point IP Address',
  `ReaderEndpoint` varchar(255) DEFAULT NULL COMMENT 'Cluster Reader End Point Address',
  `ReaderEndpointIPAddress` varchar(50) DEFAULT NULL COMMENT 'Cluster Reader End Point IP Address',
  `ClusterCreateTime` timestamp NULL DEFAULT NULL COMMENT 'UTC time',
  `CreateTime` timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
  `UpdateTime` timestamp NOT NULL DEFAULT '0000-00-00 00:00:00' ON UPDATE CURRENT_TIMESTAMP,
  PRIMARY KEY (`Profile`,`DBClusterIdentifier`),
  KEY `ix_Endpoint` (`Endpoint`),
  KEY `ix_EndpointIPAddress` (`EndpointIPAddress`),
  KEY `ix_ClusterCreateTime` (`ClusterCreateTime`),
  KEY `ix_ReaderEndpoint` (`ReaderEndpoint`),
  KEY `ix_ReaderEndpointIPAddress` (`ReaderEndpointIPAddress`)
) ENGINE=InnoDB DEFAULT CHARSET=utf8 COMMENT='Cluster Specific info';


CREATE TABLE `Tag` (
  `Profile` varchar(100) NOT NULL COMMENT 'Account where data came from',
  `Type` varchar(255) NOT NULL DEFAULT '',
  `Identifier` varchar(255) NOT NULL DEFAULT '',
  `Key` varchar(100) NOT NULL DEFAULT '',
  `Value` varchar(1000) DEFAULT NULL,
  `ResourceARN` varchar(255) DEFAULT NULL,
  `CreateTime` timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
  `UpdateTime` timestamp NOT NULL DEFAULT '0000-00-00 00:00:00' ON UPDATE CURRENT_TIMESTAMP,
  PRIMARY KEY (`Profile`,`Type`,`Identifier`,`Key`),
  KEY `ix_Identifier` (`Identifier`),
  KEY `ix_Type` (`Type`),
  KEY `ix_Key` (`Key`),
  KEY `ix_Value` (`Value`(255))
) ENGINE=InnoDB DEFAULT CHARSET=utf8 COMMENT='Instance Tag info.';

-- This is the audit table or log table:

CREATE TABLE `Audit` (
  `id` bigint(20) unsigned NOT NULL AUTO_INCREMENT,
  `Profile` varchar(100) DEFAULT NULL,
  `Identifier` varchar(100) NOT NULL DEFAULT '',
  `TableName` varchar(100) DEFAULT NULL,
  `FieldName` varchar(100) DEFAULT NULL,
  `OldValue` varchar(100) DEFAULT NULL,
  `NewValue` varchar(100) DEFAULT NULL,
  `Type` varchar(100) DEFAULT NULL,
  `timestamp` timestamp NULL DEFAULT NULL,
  PRIMARY KEY (`id`)
) ENGINE=InnoDB DEFAULT CHARSET=utf8 COMMENT='Data comes from triggers on other tables';

-- Here are the example triggers...

-- Example of saving information from a change to column called EndpointIPAddress:

DROP TRIGGER IF EXISTS MyCluster_BU;

DELIMITER $$
CREATE DEFINER=`root`@`%` TRIGGER MyCluster_BU BEFORE UPDATE ON MyCluster
  FOR EACH ROW BEGIN
    IF (OLD.EndpointIPAddress <> NEW.EndpointIPAddress AND OLD.EndpointIPAddress <> '' AND NEW.EndpointIPAddress <> '') THEN
        INSERT INTO Audit (
        `Profile`,
        `Identifier`,
        `TableName`,
        `FieldName`,
        `OldValue`,
        `NewValue`,
        `Type`,
        `timestamp`
        )
        VALUES (
        OLD.Profile,
        OLD.DBClusterIdentifier,
        'MyCluster',
        'EndpointIPAddress',
        OLD.EndpointIPAddress,
        NEW.EndpointIPAddress,
        'change',
        NOW()
        );
    END IF;

END$$
DELIMITER ;

-- Example of saving information after a delete occurs:

DROP TRIGGER IF EXISTS MyCluster_AD;

DELIMITER $$

CREATE DEFINER=`root`@`%` TRIGGER MyCluster_AD
AFTER DELETE
   ON MyCluster FOR EACH ROW

BEGIN

  INSERT INTO Audit (
  `Profile`,
  `Identifier`,
  `TableName`,
  `FieldName`,
  `OldValue`,
  `NewValue`,
  `Type`,
  `timestamp`
  )
  VALUES (
  OLD.Profile,
  OLD.DBClusterIdentifier,
  'MyCluster',
  'EndPoint',
  OLD.EndPoint,
  '',
  'delete',
  NOW()
  );

END; $$

DELIMITER ;


-- Example of saving information after a delete occurs on a key value table. The key is StackName and the value can be anything:


DROP TRIGGER IF EXISTS Tag_AD;

DELIMITER $$

CREATE DEFINER=`root`@`%` TRIGGER Tag_AD
AFTER DELETE
   ON Tag FOR EACH ROW

BEGIN

IF (OLD.Key = 'StackName') THEN
  INSERT INTO Audit (
  `Profile`,
  `Identifier`,
  `TableName`,
  `FieldName`,
  `OldValue`,
  `NewValue`,
  `Type`,
  `timestamp`
  )
  VALUES (
  OLD.Profile,
  OLD.Identifier,
  'Tag',
  'StackName',
  OLD.Value,
  '',
  'delete',
  NOW()
  );
END IF;

END; $$

DELIMITER ;


-- Example of saving information from a change to key value. The key is StackName and the Value can be anything.

DROP TRIGGER IF EXISTS Tag_BU;

DELIMITER $$
CREATE DEFINER=`root`@`%` TRIGGER Tag_BU BEFORE UPDATE ON Tag
  FOR EACH ROW BEGIN
  IF (OLD.Key = 'StackName' AND OLD.VALUE <> NEW.VALUE) THEN
        INSERT INTO Audit (
        `Profile`,
        `Identifier`,
        `TableName`,
        `FieldName`,
        `OldValue`,
        `NewValue`,
        `Type`,
        `timestamp`
        )
        VALUES (
        OLD.Profile,
        OLD.Identifier,
        'Tag',
        'StackName',
        OLD.VALUE,
        NEW.VALUE,
        'change',
        NOW()
        );
    END IF;

END$$
DELIMITER ;


-- Now add some data into the tables, make changes and delete some rows.

INSERT INTO `MyCluster` (`Profile`, `DBClusterIdentifier`, `Endpoint`, `EndpointIPAddress`, `ReaderEndpoint`, `ReaderEndpointIPAddress`, `ClusterCreateTime`, `CreateTime`, `UpdateTime`) VALUES ('test', 'test', 'test', '123', NULL, NULL, NULL, CURRENT_TIMESTAMP, '0000-00-00 00:00:00');
UPDATE `MyCluster` SET `EndpointIPAddress` = '456' WHERE `Profile` = 'test' AND `DBClusterIdentifier` = 'test';
DELETE FROM `MyCluster` WHERE (`Profile` = 'test' AND `DBClusterIdentifier` = 'test');

INSERT INTO `Tag` (`Profile`, `Type`, `Identifier`, `Key`, `Value`, `ResourceARN`, `CreateTime`, `UpdateTime`) VALUES ('test', 'Cluster', '123', 'StackName', 'MyTest', NULL, CURRENT_TIMESTAMP, '0000-00-00 00:00:00');
UPDATE `Tag` SET `Value` = 'MyTestIsDone' WHERE `Profile` = 'test' AND `Type` = 'Cluster' AND `Identifier` = '123' AND `Key` = 'StackName';
DELETE FROM `Tag` WHERE (`Profile` = 'test' AND `Type` = 'Cluster' AND `Identifier` = '123' AND `Key` = 'StackName');

Values that were changed or Deleted for the columns that have triggers setup will now appear in the Audit table.


PROS:

1.     Easy to implement.
2.     Very simple triggers
3.     Only one table needed to keep history for any number of tables

CONS

1.     Data type for old and new values is very generic, all data no matter what type it originally was is stored as TEXT
2.     There are no foreign key constraints between the tables. The columns in the change log table can refer to anything. Without  constraints, there is nothing to stop accidental or intentional manipulating of the numbers to values that don’t exist in the source table.
3.     Triggers add additional overhead to the system which could slow performance
4.     Writing queries to revert data is not simple

Setting up consumers - events_statements_history for Aurora instances with performance schema

I am wanting to setup Percona Monitoring Manager (PMM) for a client which uses Aurora. To do so, the documentation says you need to turn on the consumer in the performance schema so that events_statements_history os enabled. There isn't an option in the AWS console to do this and there isn't a parameter group setting to modify this so it must be done directly on the instance.


Here I will show you from the command line:

MySQL [(none)]> show global variables like 'performance_schema';

+--------------------+-------+

| Variable_name      | Value |

+--------------------+-------+

| performance_schema | ON    |

+--------------------+-------+

1 row in set (0.00 sec)

MySQL [(none)]> use performance_schema

Reading table information for completion of table and column names

You can turn off this feature to get a quicker startup with -A

Database changed

MySQL [performance_schema]>

MySQL [performance_schema]> select * from setup_consumers WHERE name = 'events_statements_history';

+---------------------------+---------+

| NAME                      | ENABLED |

+---------------------------+---------+

| events_statements_history | NO      |

+---------------------------+---------+

1 row in set (0.00 sec)

MySQL [performance_schema]> update setup_consumers set enabled='yes' WHERE name = 'events_statements_history';

Query OK, 1 row affected (0.00 sec)

Rows matched: 1  Changed: 1  Warnings: 0

MySQL [performance_schema]> select * from setup_consumers WHERE name = 'events_statements_history';

+---------------------------+---------+

| NAME                      | ENABLED |

+---------------------------+---------+

| events_statements_history | YES     |

+---------------------------+---------+

1 row in set (0.00 sec)

However, after an instance restart, the changes performed to setup_consumers table will be reversed. 

MySQL [performance_schema]> show global variables like 'performance_schema';

+--------------------+-------+

| Variable_name      | Value |

+--------------------+-------+

| performance_schema | ON    |

+--------------------+-------+

1 row in set (0.00 sec)

MySQL [performance_schema]> select * from setup_consumers WHERE name = 'events_statements_history';

+---------------------------+---------+

| NAME                      | ENABLED |

+---------------------------+---------+

| events_statements_history | NO      |

+---------------------------+---------+

1 row in set (0.01 sec)

MySQL [performance_schema]> update setup_consumers set enabled='yes' WHERE name = 'events_statements_history';

Query OK, 1 row affected (0.01 sec)

Rows matched: 1  Changed: 1  Warnings: 0

MySQL [performance_schema]> select * from setup_consumers WHERE name = 'events_statements_history';

+---------------------------+---------+

| NAME                      | ENABLED |

+---------------------------+---------+

| events_statements_history | YES     |

+---------------------------+---------+

1 row in set (0.00 sec)

I could setup a Zabbix trigger that turns it back whenever it detects a server restart or some other script. I think the easier route it to create an event that keeps turning it back on. Here is a simple example:

DROP EVENT IF EXISTS enable_statement_history;

CREATE

DEFINER=`root`@`localhost`

EVENT IF NOT EXISTS enable_statement_history

ON SCHEDULE EVERY 60 SECOND

    STARTS NOW()

DO

   update performance_schema.setup_consumers set enabled='yes' WHERE name = 'events_statements_history';

Encrypting the defaults file for logging into mysql

MySQL has an option to store credentials in a file so that you don't have to enter them at the command line when connecting to MySQL.

For example the "normal" way of connecting to MySQL from the command line would be like this:

# mysql -u<my_user> -p -h<server name>
Enter password:

OR

# mysql -u<my_user> -p<MYPassword>  -h<server name>

In the first example you have to enter in your password which won't work for scripts. In the second example you would have to hard code the password into your script or pull it out of a variable but it would get stored in the command line making it visible by anyone who can see what is running on the system. This is bad from a security perspective.

Instead you can use a defaults file and reference the file to logon to MySQL like this:

mysql --defaults-file=location_of_my_default_file.cnf -h<server name>

The defaults file only needs to contain these three lines:

[client]
user=my_user
password='123#_BLABLA'

You can also add a line for host if you want to limit the file to only be used by one server.

If your password is going to have special characters like # then make sure it is surrounded by single quotes like the above example.

The problem with this is now the password is stored in plain text and the security team at your company is not going to like it. This is better than having it in the command line history and visible in the process list but still too easy to discover. You could lock the permissions down so that only the root user can view it and only people with root access should theoretically ever be able to see it but that may still give several teams the possibility to view it and other indexing applications to easily discover it.

In MySQL 5.6, a new feature was added to encrypt this file with mysql_config_editor.

From the manual (https://dev.mysql.com/doc/refman/5.7/en/mysql-config-editor.html):

The encryption used by mysql_config_editor prevents passwords from appearing in .mylogin.cnf as cleartext and provides a measure of security by preventing inadvertent password exposure. For example, if you display a regular unencrypted my.cnf option file on the screen, any passwords it contains are visible for anyone to see. With .mylogin.cnf, that is not true. But the encryption used will not deter a determined attacker and you should not consider it unbreakable. A user who can gain system administration privileges on your machine to access your files could decrypt the .mylogin.cnf file with some effort.

Here is how you would create the encrypted file:

mysql_config_editor set --login-path=my_encrypted_defaults_file.cnf --host=localhost --user=root --password

And then to use it:

mysql --login-path=my_encrypted_defaults_file.cnf

OR

mysql --login-path=my_encrypted_defaults_file.cnf -h<server name>

You won't be able to view the file at my_encrypted_defaults_file.cnf. This may be "good enough" to satisfy your security team but we can do even better by using GPG. I do something similar to what is described in this Percona blog post using GPG on my laptop. I will leave that for another blog post:

https://www.percona.com/blog/2016/10/12/encrypt-defaults-file/

How to identify AWS Aurora instances with aurora_server_id

Many DBAs like to use "SELECT @@hostname" to identify a MySQL server. This can be problematic for a couple reasons. I've seen servers where the actual host name the result from "SELECT @@hostname" do not match. Someone had configured the MySQL hostname incorrectly  and the result was the hostname of a different server. In my scripts I have to check for this and make sure the resolved hostname matches the value from "SELECT @@hostname" and if it does not then do additional checking.

Furthermore, when using RDS/Aurora, you cannot rely on the value of "SELECT @@hostname" because it will give some other value, such as a private IP which can be the same on multiple systems and is not the hostname you are looking for.

In your scripts you can run:

show global variables like 'aurora_version';

This will let you know that you are on an Aurora instance. Then you can use this query to get the aurora_server_id which is the closest thing to hostname.

show global variables like 'aurora_server_id';

The aurora_server_id is not going to be resolvable unless you add an alias to it in your DNS but you can get the full resolvable server address from the AWS RDS CLI.

If you are using Python you can use the boto3 RDS docs.

Rewriting sub queries to use joins for better performance

For almost every client I have worked with, I have done some amount of profiling on their databases and recommend minor changes to improve performance. Almost every single time I see queries that use sub-selects when it is not necessary. When the data sets are small (in the thousands or less), sub selects performance is typically not an issue. However, when the number of rows in the tables get into the hundreds of thousands to millions, sub-select performance usually tanks. I've seen developers that treat MySQL tables like a queue (not a best practice) and use sub-selects. Performance will appear to be fine until the queue fills up and all of sudden the entire application is broken because they were not expecting 500,000 to a million rows in the table. Sometimes these poorly written sub-select queries will take 2~3 minutes and pile up on each other causing the MySQL server to be overwhelmed and grind to a halt.

A performance gain can almost always be gained by re-writing a query to use a join instead of a sub-select. Additionally, you need to make sure that the columns which are being used for the join have an index. If you read the "High Performance MySQL" book by the experts at Percona you will see this is one of their recommendations.

Here is a simple example I wrote to query a table that contains a list of database servers.

This query uses a sub-select. In the Explain plan's extra column, notice Using temporary; Using filesort and the the type column shows ALL meaning it is looking at all rows in the table:

This query uses a JOIN without the sub-select. If you know how to read MySQL explain plans, this one looks much better! The type columns changes to range meaning the query is no longer doing a full table scan and the filesort and temporary went away.

The data size on this table is pretty small (less than 2,000 rows). The query with a sub-select took about 65 ms while the query with the join took 50 ms.

Your mileage with re-writing queries to use a join will vary based on the data distribution, primary key, aggregation, indexes and such but the larger the table gets, a join will usually give you better performance than a sub-select. I've seen queries go from minutes to seconds or from 5 sec to 2 sec after re-writing them use a join. 

GNU Parallel for speeding up scripts

One of my favorite tools for speeding up work is GNU Parallel.

https://www.gnu.org/software/parallel/

When I write automation scripts, I try to write them to do "one simple task". For example, it will logon to a single database or a single server and so something.

In reality, I usually want the task to be repeated thousands of times and to be done in parallel to finish quickly.

Alternatively, I some times write the script to only print out the statement required to complete the task on the command line.

With GNU parallel I can use it to control my script and have it run in parallel and throttle the process. GNU parallel can limit the script to a certain number of instances of the script running at the same time.

In this example, my "createSomething.py" script only print out the commands to do the work. It will print out hundreds of commands which need to be run.

By piping it into parallel, it will automatically run 10 of those processes at a time and keep running them until they are all done.

 ./createSomething.py <input values> | parallel

In this example the file server_list.txt has multiple columns. Each column in the text file is separated by a tab. This is specified in --colsep '\t' as seen below. The values from each line will appear where is {}.

cat server_list.txt | parallel -j10 --colsep '\t' "my_shell_script.sh {}"

In this example the value which is read from server_list.txt is put into the input values after "my_python_script.py" where  {.} is:

cat server_list.txt | parallel -j10 python my_python_script.py {}

Here is another example running a query on a long list of servers:

contents of long_list_of_servers.txt would look like:
server1.com
server2.com
server3.com

Execute whatever code is in the "execute_this.sql" on every server in long_list_of_servers.txt. I use the -vvv option to log all the commands so they can be reviewed.

cat long_list_of_servers.txt  | parallel -j10 "mysql --defaults-file='my_defaults_file' -h '{}' -vvv < execute_this.sql >> outputfile.log"

If the code in "execute_this.sql" was a grant statement to create a new user or add more permissions on a list of Aurora instances, you could verify that it worked by running this next:

cat long_list_of_servers.txt | parallel -j10 "mysql --defaults-file='my_defaults_file.cnf' -h '{}' -e \"show global variables like 'aurora_server_id'; show grants for myUSER; '\" >> outputfile.log"

Now the log file will have the aurora_server_id for each servers and you can verify that each instance has the correct grants.