Explicit connection collation

[g105b]

When PDO binds data, it asks the schema for what type of data the binding should be done, so everything is done in a type-safe way. This is also the case for character sets and collations.

There are occasions when bindings are possible out-of-schema, such as when binding a value to a variable, e.g. select @example_variable := ?. Because this variable does not exist in the schema, PDO will infer the character set and collation from its connection information.

A bug arises when the connection is made in a different collation than the schema was created in. By default all schemas are created in utf8_general_ci, as this is loaded from the DefaultSettings object. However, when creating the PDO connection, the collation is never specified, so will default to whatever the OS deems as a correct default.

A recent update to MySQL/MariaDB has changed this default to move from utf8_general_ci to utf8_unicode_ci. This causes the issue in question when binding to out-of-schema values. When there is a mismatch of collations, MySQL throws an “Illegal mix of collations” error.

This can be solved by specifying the connection's collation using the MYSQL_ATTR_INIT_COMMAND attribute and setting to a value of SET SESSION collation_connection='$collation', where $collation can be loaded from the same consistent place that the migration script does.