Simplify Driver::connect() signature

[morozov]

Q	A
Type	improvement
BC Break	yes

API design considerations:

1. Driver::connect() is an adapter between the DBAL Connection::connect() and the Driver\Connection::__construct() APIs. It should do all conversion of the parameters and not leak array $params into the connection.
2. Driver\Connection::__construct() is a tiny wrapper over the underlying driver's "connect" function. Its API should be as close to the "connect" function as possible but not too close to have to leak the underlying connection resource back to the driver like it happens in some PDO drivers:

   1. dbal/src/Driver/PDOPgSql/Driver.php

      Line 38 in ae8c720

      $connection->getWrappedConnection()->setAttribute(PDO::PGSQL_ATTR_DISABLE_PREPARES, true);

   2. dbal/src/Driver/PDOSqlite/Driver.php

      Lines 43 to 47 in ae8c720

      	$pdo = $connection->getWrappedConnection();
      	foreach ($this->_userDefinedFunctions as $fn => $data) {
      	$pdo->sqliteCreateFunction($fn, $data['callback'], $data['numArgs']);
      	}

      
      In order to avoid the leakage but not to have too many parameters in MySQLi connection constructor either, a concept of connection initializers has been introduced. It can be thought of as an implementation of the Builder or Functional Options pattern.

Additional changes:

1. The custom error handler in MySQLi connection has been removed in favor of an error suppression operator. The error handler would yield the same results but required maintenance and a dedicated unit test that doesn't pass on Windows. Additionally, it masked improper usage of $connection->sqlstate in case of a connection failure (should help fixing mysqli::$sqlstate is inaccessible on a failed/non-initialized connection since PHP 8 #3947).
2. $supportedDriverOptions have been removed from the MySQLi connection. Whether the option is supported will be determined by the call to mysqli_options(). The previous approach led to improper error handling around mysqli_options() because an error was impossible to reproduce.

Although the code implementing secure MySQLi connection has been modified, it's not tested as it never was. It should be possible to set up a build job using cyprien/mysql-tls on Travis later.

Fixes #4069

[greg0ire]

Are there parts of this that could be backported to 2.11 to make the upgrade path smoother?

[greg0ire]

Should we use list<> for those?

[greg0ire]

iterable seems to be enough, we don't really need those to be arrays

[morozov]

Are there parts of this that could be backported to 2.11 to make the upgrade path smoother?

There cannot be an FC-layer provided for the changes in the Driver::connect() signature. The very fact that the drivers start using the "user", "password" and "driver_options" keys of $params that they used to ignore, even as a fallback, is a BC break.

Do you have something else in mind?

[greg0ire]

Do you have something else in mind?

No I read this just before getting to work and did not have the time to figure something out. Not sure if that's a big deal though, I don't think Driver classes are directly used downstream, are they?

[greg0ire]

As a great man asked me recently,

Does this removal need an upgrade note?

[morozov]

Not sure if that's a big deal though, I don't think Driver classes are directly used downstream, are they?

It's absolutely not a big deal. Unless someone maintains their own driver, they shouldn't be concerned about this change at all.