Introduction

Note: Changelog: Feature was added in PECL/mysqlnd_ms 1.1.0-beta

The below description applies to PECL/mysqlnd_ms >= 1.1.0-beta. It is not valid for prior versions.

The plugin uses its own configuration file. The configuration file holds information about the MySQL replication master server, the MySQL replication slave servers, the server pick (load balancing) policy, the failover strategy, and the use of lazy connections.

The plugin loads its configuration file at the beginning of a web request. It is then cached in memory and used for the duration of the web request. This way, there is no need to restart PHP after deploying the configuration file. Configuration file changes will become active almost instantly.

The PHP configuration directive mysqlnd_ms.config_file is used to set the plugins configuration file. Please note, that the PHP configuration directive may not be evaluated for every web request. Therefore, changing the plugins configuration file name or location may require a PHP restart. However, no restart is required to read changes if an already existing plugin configuration file is updated.

Using and parsing JSON is efficient, and using JSON makes it easier to express hierarchical data structures than the standard php.ini format.

mysqlnd_ms.ini file created...
Dumping file contents...
--------------------------------------------------------------------------------
{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost",
                "socket": "\/tmp\/mysql.sock"
            }
        },
        "slave": [

        ]
    }
}
--------------------------------------------------------------------------------

A plugin configuration file consists of one or more sections. Sections are represented by the top-level object properties of the object encoded in the JSON file. Sections could also be called configuration names.

Applications reference sections by their name. Applications use section names as the host (server) parameter to the various connect methods of the mysqli, mysql and PDO_MYSQL extensions. Upon connect, the mysqlnd plugin compares the hostname with all of the section names from the plugin configuration file. If the hostname and section name match, then the plugin will load the settings for that section.

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.2.27"
            },
            "slave_1": {
                "host": "192.168.2.27",
                "port": 3306
            }
        }
    },
    "localhost": {
        "master": [
            {
                "host": "localhost",
                "socket": "\/path\/to\/mysql.sock"
            }
        ],
        "slave": [
            {
                "host": "192.168.3.24",
                "port": "3305"
            },
            {
                "host": "192.168.3.65",
                "port": "3309"
            }
        ]
    }
}

Section names are strings. It is valid to use a section name such as 192.168.2.1, 127.0.0.1 or localhost. If, for example, an application connects to localhost and a plugin configuration section localhost exists, the semantics of the connect operation are changed. The application will no longer only use the MySQL server running on the host localhost, but the plugin will start to load balance MySQL queries following the rules from the localhost configuration section. This way you can load balance queries from an application without changing the applications source code. Please keep in mind, that such a configuration may not contribute to overall readability of your applications source code. Using section names that can be mixed up with host names should be seen as a last resort.

Each configuration section contains, at a minimum, a list of master servers and a list of slave servers. The master list is configured with the keyword master, while the slave list is configured with the slave keyword. Failing to provide a slave list will result in a fatal E_ERROR level error, although a slave list may be empty. It is possible to allow no slaves. However, this is only recommended with synchronous clusters, please see also supported clusters. The main part of the documentation focusses on the use of asynchronous MySQL replication clusters.

The master and slave server lists can be optionally indexed by symbolic names for the servers they describe. Alternatively, an array of descriptions for slave and master servers may be used.

"slave": [
    {
        "host": "192.168.3.24",
        "port": "3305"
    },
    {
        "host": "192.168.3.65",
        "port": "3309"
    }
]

An anonymous server list is encoded by the JSON array type. Optionally, symbolic names may be used for indexing the slave or master servers of a server list, and done so using the JSON object type.

"master": {
    "master_0": {
        "host": "localhost"
    }
}

It is recommended to index the server lists with symbolic server names. The alias names will be shown in error messages.

The order of servers is preserved and taken into account by mysqlnd_ms. If, for example, you configure round robin load balancing strategy, the first SELECT statement will be executed on the slave that appears first in the slave server list.

A configured server can be described with the host, port, socket, db, user, password and connect_flags. It is mandatory to set the database server host using the host keyword. All other settings are optional.

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "db_server_host",
                "port": "db_server_port",
                "socket": "db_server_socket",
                "db": "database_resp_schema",
                "user": "user",
                "password": "password",
                "connect_flags": 0
            }
        },
        "slave": {
            "slave_0": {
                "host": "db_server_host",
                "port": "db_server_port",
                "socket": "db_server_socket"
            }
        }
    }
}

If a setting is omitted, the plugin will use the value provided by the user API call used to open a connection. Please, see the using section names example above.

The configuration file format has been changed in version 1.1.0-beta to allow for chained filters. Filters are responsible for filtering the configured list of servers to identify a server for execution of a given statement. Filters are configured with the filter keyword. Filters are executed by mysqlnd_ms in the order of their appearance. Defining filters is optional. A configuration section in the plugins configuration file does not need to have a filters entry.

Filters replace the pick[] setting from prior versions. The new random and roundrobin provide the same functionality.

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            },
            "slave_1": {
                "host": "192.168.78.137",
                "port": "3306"
            }
        },
        "filters": {
            "roundrobin": [

            ]
        }
    }
}

The function mysqlnd_ms_set_user_pick_server() has been removed. Setting a callback is now done with the user filter. Some filters accept parameters. The user filter requires and accepts a mandatory callback parameter to set the callback previously set through the function mysqlnd_ms_set_user_pick_server().

"filters": {
    "user": {
        "callback": "pick_server"
    }
}

The validity of the configuration file is checked both when reading the configuration file and later when establishing a connection. The configuration file is read during PHP request startup. At this early stage a PHP extension may not display error messages properly. In the worst case, no error is shown and a connection attempt fails without an adequate error message. This problem has been cured in version 1.5.0.

Warning: mysqli::mysqli(): (mysqlnd_ms) (mysqlnd_ms) Failed to parse config file [s1.json]. Please, verify the JSON in Command line code

Warning: mysqli::mysqli(): (HY000/2002): php_network_getaddresses: getaddrinfo failed: Name or service not known in Command line code on line 1

Warning: mysqli::query(): Couldn't fetch mysqli in Command line code on line 1

Fatal error: Call to a member function fetch_assoc() on a non-object in Command line code on line 1

Since version 1.5.0 startup errors are additionally buffered and emitted when a connection attempt is made. Use the configuration directive mysqlnd_ms.force_config_usage to set the error type used to display buffered errors. By default an error of type E_WARNING will be emitted.

Warning: mysqli::mysqli(): (mysqlnd_ms) (mysqlnd_ms) Failed to parse config file [s1.json]. Please, verify the JSON in Command line code on line 1

It can be useful to set mysqlnd_ms.force_config_usage = 1 when debugging potential configuration file errors. This will not only turn the type of buffered startup errors into E_RECOVERABLE_ERROR but also help detecting misspelled section names.

mysqlnd_ms.force_config_usage=1

Warning: mysqli::mysqli(): (mysqlnd_ms) Exclusive usage of configuration enforced but did not find the correct INI file section (invalid_section) in Command line code on line 1 line 1

Configuration Directives

Here is a short explanation of the configuration directives that can be used.

master array or object

List of MySQL replication master servers. The list of either of the JSON type array to declare an anonymous list of servers or of the JSON type object. Please, see above for examples.

Setting at least one master server is mandatory. The plugin will issue an error of type E_ERROR if the user has failed to provide a master server list for a configuration section. The fatal error may read (mysqlnd_ms) Section [master] doesn't exist for host [name_of_a_config_section] in %s on line %d.

A server is described with the host, port, socket, db, user, password and connect_flags. It is mandatory to provide at a value for host. If any of the other values is not given, it will be taken from the user API connect call, please, see also: using section names example.

Table of server configuration keywords.

Keyword	Description	Version
host	Database server host. This is a mandatory setting. Failing to provide, will cause an error of type E_RECOVERABLE_ERROR when the plugin tries to connect to the server. The error message may read (mysqlnd_ms) Cannot find [host] in [%s] section in config in %s on line %d.	Since 1.1.0.
port	Database server TCP/IP port.	Since 1.1.0.
socket	Database server Unix domain socket.	Since 1.1.0.
db	Database (schemata).	Since 1.1.0.
user	MySQL database user.	Since 1.1.0.
password	MySQL database user password.	Since 1.1.0.
connect_flags	Connection flags.	Since 1.1.0.

The plugin supports using only one master server. An experimental setting exists to enable multi-master support. The details are not documented. The setting is meant for development only.

slave array or object

List of one or more MySQL replication slave servers. The syntax is identical to setting master servers, please, see master above for details.

The plugin supports using one or more slave servers.

Setting a list of slave servers is mandatory. The plugin will report an error of the type E_ERROR if slave is not given for a configuration section. The fatal error message may read (mysqlnd_ms) Section [slave] doesn't exist for host [%s] in %s on line %d. Note, that it is valid to use an empty slave server list. The error has been introduced to prevent accidentally setting no slaves by forgetting about the slave setting. A master-only setup is still possible using an empty slave server list.

If an empty slave list is configured and an attempt is made to execute a statement on a slave the plugin may emit a warning like mysqlnd_ms) Couldn't find the appropriate slave connection. 0 slaves to choose from. upon statement execution. It is possible that another warning follows such as (mysqlnd_ms) No connection selected by the last filter.

global_transaction_id_injection array or object

Global transaction identifier configuration related to both the use of the server built-in global transaction ID feature and the client-side emulation.

Keyword	Description	Version
fetch_last_gtid	SQL statement for accessing the latest global transaction identifier. The SQL statement is run if the plugin needs to know the most recent global transaction identifier. This can be the case, for example, when checking MySQL Replication slave status. Also used with mysqlnd_ms_get_last_gtid().	Since 1.2.0.
check_for_gtid	SQL statement for checking if a replica has replicated all transactions up to and including ones searched for. The SQL statement is run when searching for replicas which can offer a higher level of consistency than eventual consistency. The statement must contain a placeholder #GTID which is to be replaced with the global transaction identifier searched for by the plugin. Please, check the quickstart for examples.	Since 1.2.0.
report_errors	Whether to emit an error of type warning if an issue occurs while executing any of the configured SQL statements.	Since 1.2.0.
on_commit	Client-side global transaction ID emulation only. SQL statement to run when a transaction finished to update the global transaction identifier sequence number on the master. Please, see the quickstart for examples.	Since 1.2.0.
wait_for_gtid_timeout	Instructs the plugin to wait up to wait_for_gtid_timeout seconds for a slave to catch up when searching for slaves that can deliver session consistency. The setting limits the time spend for polling the slave status. If polling the status takes very long, the total clock time spend waiting may exceed wait_for_gtid_timeout. The plugin calls sleep(1) to sleep one second between each two polls. The setting can be used both with the plugins client-side emulation and the server-side global transaction identifier feature of MySQL 5.6. Waiting for a slave to replicate a certain GTID needed for session consistency also means throttling the client. By throttling the client the write load on the master is reduced indirectly. A primary copy based replication system, such as MySQL Replication, is given more time to reach a consistent state. This can be desired, for example, to increase the number of data copies for high availability considerations or to prevent the master from being overloaded.	Since 1.4.0.

fabric object

MySQL Fabric related settings. If the plugin is used together with MySQL Fabric, then the plugins configuration file no longer contains lists of MySQL servers. Instead, the plugin will ask MySQL Fabric which list of servers to use to perform a certain task.

A minimum plugin configuration for use with MySQL Fabric contains a list of one or more MySQL Fabric hosts that the plugin can query. If more than one MySQL Fabric host is configured, the plugin will use a roundrobin strategy to choose among them. Other strategies are currently not available.

Example #11 Minimum pluging configuration for use with MySQL Fabric

{
    "myapp": {
        "fabric": {
            "hosts": [
                {
                    "host" : "127.0.0.1",
                    "port" : 8080
                }
            ]
        }
    }
}

Each MySQL Fabric host is described using a JSON object with the following members.

Keyword	Description	Version
host	Host name of the MySQL Fabric host.	Since 1.6.0.
port	The TCP/IP port on which the MySQL Fabric host listens for remote procedure calls sent by clients such as the plugin.	Since 1.6.0.

The plugin is using PHP streams to communicate with MySQL Fabric through XML RPC over HTTP. By default no timeouts are set for the network communication. Thus, the plugin defaults to PHP stream default timeouts. Those defaults are out of control of the plugin itself.

An optional timeout value can be set to overrule the PHP streams default timeout setting. Setting the timeout in the plugins configuration file has the same effect as setting a timeout for a PHP user space HTTP connection established through PHP streams.

The plugins Fabric timeout value unit is seconds. The allowed value range is from 0 to 65535. The setting exists since version 1.6.

Example #12 Optional timeout for communication with Fabric

{
    "myapp": {
        "fabric": {
            "hosts": [
                {
                    "host" : "127.0.0.1",
                    "port" : 8080
                }
            ],
            "timeout": 2
        }
    }
}

Transaction stickiness and MySQL Fabric logic can collide. The stickiness option disables switching between servers for the duration of a transaction. When using Fabric and sharding the user may (erroneously) start a local transaction on one share and then attempt to switch to a different shard using either mysqlnd_ms_fabric_select_shard() or mysqlnd_ms_fabric_select_global(). In this case, the plugin will not reject the request to switch servers in the middle of a transaction but allow the user to switch to another server regardless of the transaction stickiness setting used. It is clearly a user error to write such code.

If transaction stickiness is enabled and you would like to get an error of type warning when calling mysqlnd_ms_fabric_select_shard() or mysqlnd_ms_fabric_select_global(), set the boolean flag trx_warn_server_list_changes.

Example #13 Warnings about the violation of transaction boundaries

{
    "myapp": {
        "fabric": {
            "hosts": [
                {
                    "host" : "127.0.0.1",
                    "port" : 8080
                }
            ],
            "trx_warn_serverlist_changes": 1
        },
        "trx_stickiness": "on"
    }
}

<?php
$link = new mysqli("myapp", "root", "", "test");
/*
  For the demo the call may fail.
  Failed or not we get into the state
  needed for the example.
*/
@mysqlnd_ms_fabric_select_global($link, 1);
$link->begin_transaction();
@$link->query("DROP TABLE IF EXISTS test");
/*
  Switching servers/shards is a mistake due to open
  local transaction!
*/
mysqlnd_ms_select_global($link, 1);
?>

The above example will output:

PHP Warning: mysqlnd_ms_fabric_select_global(): (mysqlnd_ms) Fabric server exchange in the middle of a transaction in %s on line %d

Please, consider the feature experimental. Changes to syntax and semantics may happen.

filters object

List of filters. A filter is responsible to filter the list of available servers for executing a given statement. Filters can be chained. The random and roundrobin filter replace the pick[] directive used in prior version to select a load balancing policy. The user filter replaces the mysqlnd_ms_set_user_pick_server() function.

Filters may accept parameters to refine their actions.

If no load balancing policy is set, the plugin will default to random_once. The random_once policy picks a random slave server when running the first read-only statement. The slave server will be used for all read-only statements until the PHP script execution ends. No load balancing policy is set and thus, defaulting takes place, if neither the random nor the roundrobin are part of a configuration section.

If a filter chain is configured so that a filter which output no more than once server is used as input for a filter which should be given more than one server as input, the plugin may emit a warning upon opening a connection. The warning may read: (mysqlnd_ms) Error while creating filter '%s' . Non-multi filter '%s' already created. Stopping in %s on line %d. Furthermore, an error of the error code 2000, the sql state HY000 and an error message similar to the warning may be set on the connection handle.

Example #14 Invalid filter sequence

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "filters": [
            "roundrobin",
            "random"
        ]
    }
}

<?php
$link = new mysqli("myapp", "root", "", "test");
printf("[%d] %s\n", mysqli_connect_errno(), mysqli_connect_error());
$link->query("SELECT 1 FROM DUAL");
?>

The above example will output:

PHP Warning:  mysqli::mysqli(): (HY000/2000): (mysqlnd_ms) Error while creating filter 'random' . Non-multi filter 'roundrobin' already created. Stopping in filter_warning.php on line 1
[2000] (mysqlnd_ms) Error while creating filter 'random' . Non-multi filter 'roundrobin' already created. Stopping
PHP Warning:  mysqli::query(): Couldn't fetch mysqli in filter_warning.php on line 3

Filter: random object

The random filter features the random and random once load balancing policies, set through the pick[] directive in older versions.

The random policy will pick a random server whenever a read-only statement is to be executed. The random once strategy picks a random slave server once and continues using the slave for the rest of the PHP web request. Random once is a default, if load balancing is not configured through a filter.

If the random filter is not given any arguments, it stands for random load balancing policy.

Example #15 Random load balancing with random filter

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            },
            "slave_1": {
                "host": "192.168.78.137",
                "port": "3306"
            }
        },
        "filters": [
            "random"
        ]
    }
}

Optionally, the sticky argument can be passed to the filter. If the parameter sticky is set to the string 1, the filter follows the random once load balancing strategy.

Example #16 Random once load balancing with random filter

{
    "filters": {
        "random": {
            "sticky": "1"
        }
    }
}

Both the random and roundrobin filters support setting a priority, a weight for a server, since PECL/mysqlnd_ms 1.4.0. If the weight argument is passed to the filter, it must assign a weight for all servers. Servers must be given an alias name in the slave respectively master server lists. The alias must be used to reference servers for assigning a priority with weight.

Example #17 Referencing error

[E_RECOVERABLE_ERROR] mysqli_real_connect(): (mysqlnd_ms) Unknown server 'slave3' in 'random' filter configuration. Stopping in %s on line %d

Using a wrong alias name with weight may result in an error similar to the shown above.

If weight is omitted, the default weight of all servers is one.

Example #18 Assigning a weight for load balancing

{
   "myapp": {
       "master": {
           "master1":{
               "host":"localhost",
               "socket":"\/var\/run\/mysql\/mysql.sock"
           }
       },
       "slave": {
           "slave1": {
               "host":"192.168.2.28",
               "port":3306
           },
           "slave2": {
               "host":"192.168.2.29",
               "port":3306
           },
           "slave3": {
               "host":"192.0.43.10",
               "port":3306
           },
       },
       "filters": {
           "random": {
               "weights": {
                   "slave1":8,
                   "slave2":4,
                   "slave3":1,
                   "master1":1
               }
           }
       }
   }
}

At the average a server assigned a weight of two will be selected twice as often as a server assigned a weight of one. Different weights can be assigned to reflect differently sized machines, to prefer co-located slaves which have a low network latency or, to configure a standby failover server. In the latter case, you may want to assign the standby server a very low weight in relation to the other servers. For example, given the configuration above slave3 will get only some eight percent of the requests in the average. As long as slave1 and slave2 are running, it will be used sparsely, similar to a standby failover server. Upon failure of slave1 and slave2, the usage of slave3 increases. Please, check the notes on failover before using weight this way.

Valid weight values range from 1 to 65535.

Unknown arguments are ignored by the filter. No warning or error is given.

The filter expects one or more servers as input. Outputs one server. A filter sequence such as random, roundrobin may cause a warning and an error message to be set on the connection handle when executing a statement.

List of filter arguments.

Keyword	Description	Version
sticky	Enables or disabled random once load balancing policy. See above.	Since 1.2.0.
weight	Assigns a load balancing weight/priority to a server. Please, see above for a description.	Since 1.4.0.

Filter: roundrobin object

If using the roundrobin filter, the plugin iterates over the list of configured slave servers to pick a server for statement execution. If the plugin reaches the end of the list, it wraps around to the beginning of the list and picks the first configured slave server.

Example #19 roundrobin filter

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "filters": [
            "roundrobin"
        ]
    }
}

Expects one or more servers as input. Outputs one server. A filter sequence such as roundrobin, random may cause a warning and an error message to be set on the connection handle when executing a statement.

List of filter arguments.

Keyword	Description	Version
weight	Assigns a load balancing weight/priority to a server. Please, find a description above.	Since 1.4.0.

Filter: user object

The user replaces mysqlnd_ms_set_user_pick_server() function, which was removed in 1.1.0-beta. The filter sets a callback for user-defined read/write splitting and server selection.

The plugins built-in read/write query split mechanism decisions can be overwritten in two ways. The easiest way is to prepend a query string with the SQL hints MYSQLND_MS_MASTER_SWITCH, MYSQLND_MS_SLAVE_SWITCH or MYSQLND_MS_LAST_USED_SWITCH. Using SQL hints one can control, for example, whether a query shall be send to the MySQL replication master server or one of the slave servers. By help of SQL hints it is not possible to pick a certain slave server for query execution.

Full control on server selection can be gained using a callback function. Use of a callback is recommended to expert users only because the callback has to cover all cases otherwise handled by the plugin.

The plugin will invoke the callback function for selecting a server from the lists of configured master and slave servers. The callback function inspects the query to run and picks a server for query execution by returning the hosts URI, as found in the master and slave list.

If the lazy connections are enabled and the callback chooses a slave server for which no connection has been established so far and establishing the connection to the slave fails, the plugin will return an error upon the next action on the failed connection, for example, when running a query. It is the responsibility of the application developer to handle the error. For example, the application can re-run the query to trigger a new server selection and callback invocation. If so, the callback must make sure to select a different slave, or check slave availability, before returning to the plugin to prevent an endless loop.

Example #20 Setting a callback

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "filters": {
            "user": {
                "callback": "pick_server"
            }
        }
    }
}

The callback is supposed to return a host to run the query on. The host URI is to be taken from the master and slave connection lists passed to the callback function. If callback returns a value neither found in the master nor in the slave connection lists the plugin will emit an error of the type E_RECOVERABLE_ERROR The error may read like (mysqlnd_ms) User filter callback has returned an unknown server. The server 'server that is not in master or slave list' can neither be found in the master list nor in the slave list. If the application catches the error to ignore it, follow up errors may be set on the connection handle, for example, (mysqlnd_ms) No connection selected by the last filter with the error code 2000 and the sqlstate HY000. Furthermore a warning may be emitted.

Referencing a non-existing function as a callback will result in any error of the type E_RECOVERABLE_ERROR whenever the plugin tries to callback function. The error message may reads like: (mysqlnd_ms) Specified callback (pick_server) is not a valid callback. If the application catches the error to ignore it, follow up errors may be set on the connection handle, for example, (mysqlnd_ms) Specified callback (pick_server) is not a valid callback with the error code 2000 and the sqlstate HY000. Furthermore a warning may be emitted.

The following parameters are passed from the plugin to the callback.

Parameter	Description	Version
connected_host	URI of the currently connected database server.	Since 1.1.0.
query	Query string of the statement for which a server needs to be picked.	Since 1.1.0.
masters	List of master servers to choose from. Note, that the list of master servers may not be identical to the list of configured master servers if the filter is not the first in the filter chain. Previously run filters may have reduced the master list already.	Since 1.1.0.
slaves	List of slave servers to choose from. Note, that the list of slave servers may not be identical to the list of configured slave servers if the filter is not the first in the filter chain. Previously run filters may have reduced the slave list already.	Since 1.1.0.
last_used_connection	URI of the server of the connection used to execute the previous statement on.	Since 1.1.0.
in_transaction	Boolean flag indicating whether the statement is part of an open transaction. If autocommit mode is turned off, this will be set to true. Otherwise it is set to false. Transaction detection is based on monitoring the mysqlnd library call set_autocommit. Monitoring is not possible before PHP 5.4.0. Please, see connection pooling and switching concepts discussion for further details.	Since 1.1.0.

Example #21 Using a callback

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.2.27",
                "port": "3306"
            },
            "slave_1": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "filters": {
            "user": {
                "callback": "pick_server"
            }
        }
    }
}

<?php
function pick_server($connected, $query, $masters, $slaves, $last_used_connection, $in_transaction)
{
 static $slave_idx = 0;
 static $num_slaves = NULL;
 if (is_null($num_slaves))
  $num_slaves = count($slaves);

 /* default: fallback to the plugins build-in logic */
 $ret = NULL;

 printf("User has connected to '%s'...\n", $connected);
 printf("... deciding where to run '%s'\n", $query);

 $where = mysqlnd_ms_query_is_select($query);
 switch ($where)
 {
  case MYSQLND_MS_QUERY_USE_MASTER:
   printf("... using master\n");
   $ret = $masters[0];
   break;
  case MYSQLND_MS_QUERY_USE_SLAVE:
   /* SELECT or SQL hint for using slave */
   if (stristr($query, "FROM table_on_slave_a_only"))
   {
    /* a table which is only on the first configured slave  */
    printf("... access to table available only on slave A detected\n");
    $ret = $slaves[0];
   }
   else
   {
    /* round robin */
    printf("... some read-only query for a slave\n");
    $ret = $slaves[$slave_idx++ % $num_slaves];
   }
   break;
  case MYSQLND_MS_QUERY_LAST_USED:
   printf("... using last used server\n");
   $ret = $last_used_connection;
   break;
 }

 printf("... ret = '%s'\n", $ret);
 return $ret;
}

$mysqli = new mysqli("myapp", "root", "", "test");

if (!($res = $mysqli->query("SELECT 1 FROM DUAL")))
 printf("[%d] %s\n", $mysqli->errno, $mysqli->error);
else
 $res->close();

if (!($res = $mysqli->query("SELECT 2 FROM DUAL")))
 printf("[%d] %s\n", $mysqli->errno, $mysqli->error);
else
 $res->close();


if (!($res = $mysqli->query("SELECT * FROM table_on_slave_a_only")))
 printf("[%d] %s\n", $mysqli->errno, $mysqli->error);
else
 $res->close();

$mysqli->close();
?>

The above example will output:

User has connected to 'myapp'...
... deciding where to run 'SELECT 1 FROM DUAL'
... some read-only query for a slave
... ret = 'tcp://192.168.2.27:3306'
User has connected to 'myapp'...
... deciding where to run 'SELECT 2 FROM DUAL'
... some read-only query for a slave
... ret = 'tcp://192.168.78.136:3306'
User has connected to 'myapp'...
... deciding where to run 'SELECT * FROM table_on_slave_a_only'
... access to table available only on slave A detected
... ret = 'tcp://192.168.2.27:3306'

Filter: user_multi object

The user_multi differs from the user only in one aspect. Otherwise, their syntax is identical. The user filter must pick and return exactly one node for statement execution. A filter chain usually ends with a filter that emits only one node. The filter chain shall reduce the list of candidates for statement execution down to one. This, only one node left, is the case after the user filter has been run.

The user_multi filter is a multi filter. It returns a list of slave and a list of master servers. This list needs further filtering to identify exactly one node for statement execution. A multi filter is typically placed at the top of the filter chain. The quality_of_service filter is another example of a multi filter.

The return value of the callback set for user_multi must be an array with two elements. The first element holds a list of selected master servers. The second element contains a list of selected slave servers. The lists shall contain the keys of the slave and master servers as found in the slave and master lists passed to the callback. The below example returns random master and slave lists extracted from the functions input.

Example #22 Returning random masters and slaves

<?php
function pick_server($connected, $query, $masters, $slaves, $last_used_connection, $in_transaction)
{
  $picked_masters = array()
  foreach ($masters as $key => $value) {
    if (mt_rand(0, 2) > 1)
      $picked_masters[] = $key;
  }
  $picked_slaves = array()
  foreach ($slaves as $key => $value) {
    if (mt_rand(0, 2) > 1)
      $picked_slaves[] = $key;
  }
  return array($picked_masters, $picked_slaves);
}
?>

The plugin will issue an error of type E_RECOVERABLE if the callback fails to return a server list. The error may read (mysqlnd_ms) User multi filter callback has not returned a list of servers to use. The callback must return an array in %s on line %d. In case the server list is not empty but has invalid servers key/ids in it, an error of type E_RECOVERABLE will the thrown with an error message like (mysqlnd_ms) User multi filter callback has returned an invalid list of servers to use. Server id is negative in %s on line %d, or similar.

Whether an error is emitted in case of an empty slave or master list depends on the configuration. If an empty master list is returned for a write operation, it is likely that the plugin will emit a warning that may read (mysqlnd_ms) Couldn't find the appropriate master connection. 0 masters to choose from. Something is wrong in %s on line %d. Typically a follow up error of type E_ERROR will happen. In case of a read operation and an empty slave list the behavior depends on the fail over configuration. If fail over to master is enabled, no error should appear. If fail over to master is deactivated the plugin will emit a warning that may read (mysqlnd_ms) Couldn't find the appropriate slave connection. 0 slaves to choose from. Something is wrong in %s on line %d.

Filter: node_groups object

The node_groups filter lets you group cluster nodes and query selected groups, for example, to support data partitioning. Data partitioning can be required for manual sharding, primary copy based clusters running multiple masters, or to avoid hot spots in update everywhere clusters that have no built-in partitioning. The filter is a multi filter which returns zero, one or multiple of its input servers. Thus, it must be followed by other filters to reduce the number of candidates down to one for statement execution.

Keyword	Description	Version
user defined node group name	One or more node groups must be defined. A node group can have an arbitrary user defined name. The name is used in combination with a SQL hint to restrict query execution to the nodes listed for the node group. To run a query on any of the servers of a node group, the query must begin with the SQL hint /*user defined node group name*/. Please note, no white space is allowed around user defined node group name. Because user defined node group name is used as-is as part of a SQL hint, you should choose the name that is compliant with the SQL language. Each node group entry must contain a list of master servers. Additional slave servers are allowed. Failing to provide a list of master for a node group name_of_group may cause an error of type E_RECOVERABLE_ERROR like (mysqlnd_ms) No masters configured in node group 'name_of_group' for 'node_groups' filter. The list of master and slave servers must reference corresponding entries in the global master respectively slave server list. Referencing an unknown server in either of the both server lists may cause an E_RECOVERABLE_ERROR error like (mysqlnd_ms) Unknown master 'server_alias_name' (section 'name_of_group') in 'node_groups' filter configuration. Example #23 Manual partitioning { "myapp": { "master": { "master_0": { "host": "localhost", "socket": "\/tmp\/mysql.sock" } }, "slave": { "slave_0": { "host": "192.168.2.28", "port": 3306 }, "slave_1": { "host": "127.0.0.1", "port": 3311 } }, "filters": { "node_groups": { "Partition_A" : { "master": ["master_0"], "slave": ["slave_0"] } }, "roundrobin": [] } } } Please note, if a filter chain generates an empty slave list and the PHP configuration directive mysqlnd_ms.multi_master=0 is used, the plugin may emit a warning.	Since 1.5.0.

Filter: quality_of_service object

The quality_of_service identifies cluster nodes capable of delivering a certain quality of service. It is a multi filter which returns zero, one or multiple of its input servers. Thus, it must be followed by other filters to reduce the number of candidates down to one for statement execution.

The quality_of_service filter has been introduced in 1.2.0-alpha. In the 1.2 series the filters focus is on the consistency aspect of service quality. Different types of clusters offer different default data consistencies. For example, an asynchronous MySQL replication slave offers eventual consistency. The slave may not be able to deliver requested data because it has not replicated the write, it may serve stale database because its lagging behind or it may serve current information. Often, this is acceptable. In some cases higher consistency levels are needed for the application to work correct. In those cases, the quality_of_service can filter out cluster nodes which cannot deliver the necessary quality of service.

The quality_of_service filter can be replaced or created at runtime. A successful call to mysqlnd_ms_set_qos() removes all existing qos filter entries from the filter list and installs a new one at the very beginning. All settings that can be made through mysqlnd_ms_set_qos() can also be in the plugins configuration file. However, use of the function is by far the most common use case. Instead of setting session consistency and strong consistency service levels in the plugins configuration file it is recommended to define only masters and no slaves. Both service levels will force the use of masters only. Using an empty slave list shortens the configuration file, thus improving readability. The only service level for which there is a case of defining in the plugins configuration file is the combination of eventual consistency and maximum slave lag.

Keyword	Description	Version
eventual_consistency	Request eventual consistency. Allows the use of all master and slave servers. Data returned may or may not be current. Eventual consistency accepts an optional age parameter. If age is given the plugin considers only slaves for reading for which MySQL replication reports a slave lag less or equal to age. The replication lag is measure using SHOW SLAVE STATUS. If the plugin fails to fetch the replication lag, the slave tested is skipped. Implementation details and tips are given in the quality of service concepts section. Please note, if a filter chain generates an empty slave list and the PHP configuration directive mysqlnd_ms.multi_master=0 is used, the plugin may emit a warning. Example #24 Global limit on slave lag { "myapp": { "master": { "master_0": { "host": "localhost" } }, "slave": { "slave_0": { "host": "192.168.2.27", "port": "3306" }, "slave_1": { "host": "192.168.78.136", "port": "3306" } }, "filters": { "quality_of_service": { "eventual_consistency": { "age":123 } } } } }	Since 1.2.0.
session_consistency	Request session consistency (read your writes). Allows use of all masters and all slaves which are in sync with the master. If no further parameters are given slaves are filtered out as there is no reliable way to test if a slave has caught up to the master or is lagging behind. Please note, if a filter chain generates an empty slave list and the PHP configuration directive mysqlnd_ms.multi_master=0 is used, the plugin may emit a warning. Session consistency temporarily requested using mysqlnd_ms_set_qos() is a valuable alternative to using master_on_write. master_on_write is likely to send more statements to the master than needed. The application may be able to continue operation at a lower consistency level after it has done some critical reads.	Since 1.1.0.
strong_consistency	Request strong consistency. Only masters will be used.	Since 1.2.0.

failover Up to and including 1.3.x: string. Since 1.4.0: object.

Failover policy. Supported policies: disabled (default), master, loop_before_master (Since 1.4.0).

If no failover policy is set, the plugin will not do any automatic failover (failover=disabled). Whenever the plugin fails to connect a server it will emit a warning and set the connections error code and message. Thereafter it is up to the application to handle the error and, for example, resent the last statement to trigger the selection of another server.

Please note, the automatic failover logic is applied when opening connections only. Once a connection has been opened no automatic attempts are made to reopen it in case of an error. If, for example, the server a connection is connected to is shut down and the user attempts to run a statement on the connection, no automatic failover will be tried. Instead, an error will be reported.

If using failover=master the plugin will implicitly failover to a master, if available. Please check the concepts documentation to learn about potential pitfalls and risks of using failover=master.

Example #25 Optional master failover when failing to connect to slave (PECL/mysqlnd_ms < 1.4.0)

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "failover": "master"
    }
}

Since PECL/mysqlnd_ms 1.4.0 the failover configuration keyword refers to an object.

Example #26 New syntax since 1.4.0

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "failover": {"strategy": "master" }
    }
}

Keyword	Description	Version
strategy	Failover policy. Possible values: disabled (default), master, loop_before_master A value of disabled disables automatic failover. Setting master instructs the plugin to try to connect to a master in case of a slave connection error. If the master connection attempt fails, the plugin exists the failover loop and returns an error to the user. If using loop_before_master and a slave request is made, the plugin tries to connect to other slaves before failing over to a master. If multiple master are given and multi master is enabled, the plugin also loops over the list of masters and attempts to connect before returning an error to the user.	Since 1.4.0.
remember_failed	Remember failures for the duration of a web request. Default: false. If set to true the plugin will remember failed hosts and skip the hosts in all future load balancing made for the duration of the current web request.	Since 1.4.0. The feature is only available together with the random and roundrobin load balancing filter. Use of the setting is recommended.
max_retries	Maximum number of connection attempts before skipping host. Default: 0 (no limit). The setting is used to prevent hosts from being dropped of the host list upon the first failure. If set to n > 0, the plugin will keep the node in the node list even after a failed connection attempt. The node will not be removed immediately from the slave respectively master lists after the first connection failure but instead be tried to connect to up to n times in future load balancing rounds before being removed.	Since 1.4.0. The feature is only available together with the random and roundrobin load balancing filter.

Setting failover to any other value but disabled, master or loop_before_master will not emit any warning or error.

lazy_connections bool

Controls the use of lazy connections. Lazy connections are connections which are not opened before the client sends the first connection. Lazy connections are a default.

It is strongly recommended to use lazy connections. Lazy connections help to keep the number of open connections low. If you disable lazy connections and, for example, configure one MySQL replication master server and two MySQL replication slaves, the plugin will open three connections upon the first call to a connect function although the application might use the master connection only.

Lazy connections bare a risk if you make heavy use of actions which change the state of a connection. The plugin does not dispatch all state changing actions to all connections from the connection pool. The few dispatched actions are applied to already opened connections only. Lazy connections opened in the future are not affected. Only some settings are "remembered" and applied when lazy connections are opened.

Example #27 Disabling lazy connection

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "lazy_connections": 0
    }
}

Please, see also server_charset to overcome potential problems with string escaping and servers using different default charsets.

server_charset string

The setting has been introduced in 1.4.0. It is recommended to set it if using lazy connections.

The server_charset setting serves two purposes. It acts as a fallback charset to be used for string escaping done before a connection has been established and it helps to avoid escaping pitfalls in heterogeneous environments which servers using different default charsets.

String escaping takes a connections charset into account. String escaping is not possible before a connection has been opened and the connections charset is known. The use of lazy connections delays the actual opening of connections until a statement is send.

An application using lazy connections may attempt to escape a string before sending a statement. In fact, this should be a common case as the statement string may contain the string that is to be escaped. However, due to the lazy connection feature no connection has been opened yet and escaping fails. The plugin may report an error of the type E_WARNING and a message like (mysqlnd_ms) string escaping doesn't work without established connection. Possible solution is to add server_charset to your configuration to inform you of the pitfall.

Setting server_charset makes the plugin use the given charset for string escaping done on lazy connection handles before establishing a network connection to MySQL. Furthermore, the plugin will enforce the use of the charset when the connection is established.

Enforcing the use of the configured charset used for escaping is done to prevent tapping into the pitfall of using a different charset for escaping than used later for the connection. This has the additional benefit of removing the need to align the charset configuration of all servers used. No matter what the default charset on any of the servers is, the plugin will set the configured one as a default.

The plugin does not stop the user from changing the charset at any time using the set_charset() call or corresponding SQL statements. Please, note that the use of SQL is not recommended as it cannot be monitored by the plugin. The user can, for example, change the charset on a lazy connection handle after escaping a string and before the actual connection is opened. The charset set by the user will be used for any subsequent escaping before the connection is established. The connection will be established using the configured charset, no matter what the server charset is or what the user has set before. Once a connection has been opened, set_charset is of no meaning anymore.

Example #28 String escaping on a lazy connection handle

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "lazy_connections": 1,
        "server_charset" : "utf8mb4"
    }
}

<?php
$mysqli = new mysqli("myapp", "username", "password", "database");
$mysqli->real_escape("this will be escaped using the server_charset setting - utf8mb4");
$mysqli->set_charset("latin1");
$mysqli->real_escape("this will be escaped using latin1");
/* server_charset implicitly set - utf8mb4 connection */
$mysqli->query("SELECT 'This connection will be set to server_charset upon establishing' AS _msg FROM DUAL");
/* latin1 used from now on */
$mysqli->set_charset("latin1");
?>

master_on_write bool

If set, the plugin will use the master server only after the first statement has been executed on the master. Applications can still send statements to the slaves using SQL hints to overrule the automatic decision.

The setting may help with replication lag. If an application runs an INSERT the plugin will, by default, use the master to execute all following statements, including SELECT statements. This helps to avoid problems with reads from slaves which have not replicated the INSERT yet.

Example #29 Master on write for consistent reads

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "master_on_write": 1
    }
}

Please, note the quality_of_service filter introduced in version 1.2.0-alpha. It gives finer control, for example, for achieving read-your-writes and, it offers additional functionality introducing service levels.

All transaction stickiness settings, including trx_stickiness=on, are overruled by master_on_write=1.

trx_stickiness string

Transaction stickiness policy. Supported policies: disabled (default), master.

The setting requires 5.4.0 or newer. If used with PHP older than 5.4.0, the plugin will emit a warning like (mysqlnd_ms) trx_stickiness strategy is not supported before PHP 5.3.99.

If no transaction stickiness policy is set or, if setting trx_stickiness=disabled, the plugin is not transaction aware. Thus, the plugin may load balance connections and switch connections in the middle of a transaction. The plugin is not transaction safe. SQL hints must be used avoid connection switches during a transaction.

As of PHP 5.4.0 the mysqlnd library allows the plugin to monitor the autocommit mode set by calls to the libraries set_autocommit() function. If setting set_stickiness=master and autocommit gets disabled by a PHP MySQL extension invoking the mysqlnd library internal function call set_autocommit(), the plugin is made aware of the begin of a transaction. Then, the plugin stops load balancing and directs all statements to the master server until autocommit is enabled. Thus, no SQL hints are required.

An example of a PHP MySQL API function calling the mysqlnd library internal function call set_autocommit() is mysqli_autocommit().

Although setting trx_stickiness=master, the plugin cannot be made aware of autocommit mode changes caused by SQL statements such as SET AUTOCOMMIT=0 or BEGIN.

As of PHP 5.5.0, the mysqlnd library features additional C API calls to control transactions. The level of control matches the one offered by SQL statements. The mysqli API has been modified to use these calls. Since version 1.5.0, PECL/mysqlnd_ms can monitor not only mysqli_autocommit(), but also mysqli_begin(), mysqli_commit() and mysqli_rollback() to detect transaction boundaries and stop load balancing for the duration of a transaction.

Example #30 Using master to execute transactions

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
        },
        "trx_stickiness": "master"
    }
}

Since version 1.5.0 automatic and silent failover is disabled for the duration of a transaction. If the boundaries of a transaction have been properly detected, transaction stickiness is enabled and a server fails, the plugin will not attempt to fail over to the next server, if any, regardless of the failover policy configured. The user must handle the error manually. Depending on the configuration, the plugin may emit an error of type E_WARNING reading like (mysqlnd_ms) Automatic failover is not permitted in the middle of a transaction. This error may then be overwritten by follow up errors such as (mysqlnd_ms) No connection selected by the last filter. Those errors will be generated by the failing query function.

Example #31 No automatic failover, error handling pitfall

<?php
/* assumption: automatic failover configured */
$mysqli = new mysqli("myapp", "username", "password", "database");

/* sets plugin internal state in_trx = 1 */
$mysqli->autocommit(false);

/* assumption: server fails */
if (!($res = $mysqli->query("SELECT 'Assume this query fails' AS _msg FROM DUAL"))) {
 /* handle failure of transaction, plugin internal state is still in_trx = 1 */
 printf("[%d] %s", $mysqli->errno, $mysqli->error);
 /*
  If using autocommit() based transaction detection it is a
  MUST to call autocommit(true). Otherwise the plugin assumes
  the current transaction continues and connection
  changes remain forbidden.
 */
 $mysqli->autocommit(true);
 /* Likewise, you'll want to start a new transaction */
 $mysqli->autocommit(false);
}
/* latin1 used from now on */
$mysqli->set_charset("latin1");
?>

If a server fails in the middle of a transaction the plugin continues to refuse to switch connections until the current transaction has been finished. Recall that the plugin monitors API calls to detect transaction boundaries. Thus, you have to, for example, enable auto commit mode to end the current transaction before the plugin continues load balancing and switches the server. Likewise, you will want to start a new transaction immediately thereafter and disable auto commit mode again.

Not handling failed queries and not ending a failed transaction using API calls may cause all following commands emit errors such as Commands out of sync; you can't run this command now. Thus, it is important to handle all errors.

transient_error object

The setting has been introduced in 1.6.0.

A database cluster node may reply a transient error to a client. The client can then repeat the operation on the same node, fail over to a different node or abort the operation. Per definition is it safe for a client to retry the same operation on the same node before giving up.

PECL/mysqlnd_ms can perform the retry loop on behalf of the application. By configuring transient_error the plugin can be instructed to repeat operations failing with a certain error code for a certain maximum number of times with a pause between the retries. If the transient error disappears during loop execution, it is hidden from the application. Otherwise, the error is forwarded to the application by the end of the loop.

Example #32 Retry loop for transient errors

{
    "myapp": {
        "master": {
            "master_0": {
                "host": "localhost"
            }
        },
        "slave": {
            "slave_0": {
                "host": "192.168.78.136",
                "port": "3306"
            }
       },
       "transient_error": {
          "mysql_error_codes": [
            1297
          ],
          "max_retries": 2,
          "usleep_retry": 100
       }
    }
}

Keyword	Description	Version
mysql_error_codes	List of transient error codes. You may add any MySQL error code to the list. It is possible to consider any error as transient not only 1297 (HY000 (ER_GET_TEMPORARY_ERRMSG), Message: Got temporary error %d '%s' from %s). Before adding other codes but 1297 to the list, make sure your cluster supports a new attempt without impacting the state of your application.	Since 1.6.0.
max_retries	How often to retry an operation which fails with a transient error before forwarding the failure to the user. Default: 1	Since 1.6.0.
usleep_retry	Milliseconds to sleep between transient error retries. The value is passed to the C function usleep(), hence the name. Default: 100	Since 1.6.0.

xa object

The setting has been introduced in 1.6.0.

Note: Experimental

The feature is currently under development. There may be issues and/or feature limitations. Do not use in production environments.

state_store

record_participant_credentials

Whether to store the username and password of a global transaction participant in the participants table. If disabled, the garbage collection will use the default username and password when connecting to the participants. Unless you are using a different username and password for each of your MySQL servers, you can use the default and avoid storing the sensible information in state store.

Please note, username and password are stored in clear text when using the MySQL state store, which is the only one available. It is in your responsibility to protect this sensible information.

Default: false

participant_localhost_ip

During XA garbage collection the plugin may find a participant server for which the host localhost has been recorded. If the garbage collection takes place on another host but the host that has written the participant record to the state store, the host name localhost now resolves to a different host. Therefore, when recording a participant servers host name in the state store, a value of localhost must be replaced with the actual IP address of localhost.

Setting participant_localhost_ip should be considered only if using localhost cannot be avoided. From a garbage collection point of view only, it is preferrable not to configure any socket connection but to provide an IP address and port for a node.

mysql

The MySQL state store is the only state store available.

global_trx_table

Name of the MySQL table used to store the state of an ongoing or aborted global transaction. Use the below SQL statement to create the table. Make sure to edit the table name to match your configuration.

Default: mysqlnd_ms_xa_trx

Example #33 SQL definition for the MySQL state store transaction table

CREATE TABLE mysqlnd_ms_xa_trx (
  store_trx_id int(11) NOT NULL AUTO_INCREMENT,
  gtrid int(11) NOT NULL,
  format_id int(10) unsigned NOT NULL DEFAULT '1',
  state enum('XA_NON_EXISTING','XA_ACTIVE','XA_IDLE','XA_PREPARED','XA_COMMIT','XA_ROLLBACK') NOT NULL DEFAULT 'XA_NON_EXISTING',
  intend enum('XA_NON_EXISTING','XA_ACTIVE','XA_IDLE','XA_PREPARED','XA_COMMIT','XA_ROLLBACK') DEFAULT 'XA_NON_EXISTING',
  finished enum('NO','SUCCESS','FAILURE') NOT NULL DEFAULT 'NO',
  modified timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
  started datetime DEFAULT NULL,
  timeout datetime DEFAULT NULL,
  PRIMARY KEY (store_trx_id),
  KEY idx_xa_id (gtrid,format_id,finished),
  KEY idx_state (state)
) ENGINE=InnoDB

participant_table

Name of the MySQL table used to store participants of an ongoing or aborted global transaction. Use the below SQL statement to create the table. Make sure to edit the table name to match your configuration.

Storing credentials can be enabled and disabled using record_participant_credentials

Default: mysqlnd_ms_xa_participants

Example #34 SQL definition for the MySQL state store transaction table

CREATE TABLE mysqlnd_ms_xa_participants (
  fk_store_trx_id int(11) NOT NULL,
  bqual varbinary(64) NOT NULL DEFAULT '',
  participant_id int(10) unsigned NOT NULL AUTO_INCREMENT,
  server_uuid varchar(127) DEFAULT NULL,
  scheme varchar(1024) NOT NULL,
  host varchar(127) DEFAULT NULL,
  port smallint(5) unsigned DEFAULT NULL,
  socket varchar(127) DEFAULT NULL,
  user varchar(127) DEFAULT NULL,
  password varchar(127) DEFAULT NULL,
  state enum('XA_NON_EXISTING','XA_ACTIVE','XA_IDLE','XA_PREPARED','XA_COMMIT','XA_ROLLBACK')
   NOT NULL DEFAULT 'XA_NON_EXISTING',
  health enum('OK','GC_DONE','CLIENT ERROR','SERVER ERROR') NOT NULL DEFAULT 'OK',
  connection_id int(10) unsigned DEFAULT NULL,
  client_errno smallint(5) unsigned DEFAULT NULL,
  client_error varchar(1024) DEFAULT NULL,
  modified timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
  PRIMARY KEY (participant_id),
  KEY idx_xa_bqual (bqual),
  KEY idx_store_trx (fk_store_trx_id),
  CONSTRAINT mysqlnd_ms_xa_participants_ibfk_1 FOREIGN KEY (fk_store_trx_id)
    REFERENCES mysqlnd_ms_xa_trx (store_trx_id) ON DELETE CASCADE ON UPDATE CASCADE
) ENGINE=InnoDB

garbage_collection_table

Name of the MySQL table used to track and synchronize garbage collection runs. Use the below SQL statement to create the table. Make sure to edit the table name to match your configuration.

Default: mysqlnd_ms_xa_gc

Example #35 SQL definition for the MySQL state store garbage collection table

CREATE TABLE mysqlnd_ms_xa_gc (
  gc_id int(10) unsigned NOT NULL AUTO_INCREMENT,
  gtrid int(11) NOT NULL,
  format_id int(10) unsigned NOT NULL DEFAULT '1',
  fk_store_trx_id int(11) DEFAULT NULL,
  modified timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
  attempts smallint(5) unsigned NOT NULL DEFAULT '0',
  PRIMARY KEY (gc_id),
  KEY idx_store_trx (gtrid,format_id,fk_store_trx_id)
) ENGINE=InnoDB

host

Host name of the MySQL server.

user

Name of the user used to connect to the MySQL server.

password

Password for the MySQL server user.

db

Database that holds the garbage collection tables. Please note, you have to create the garbage collection tables prior to using the plugin. The tables will not be created implicitly during runtime but garbage collection will fail if the tables to not exist.

port

Port of the MySQL server.

socket

Unix domain socket of the MySQL server. Please note, if you have multiple PHP servers each of them will try to carry out garbage collection and need to be able to connect to the state store. In this case, you may prefer configuring an IP address and a port for the MySQL state store server to ensure all PHP servers can reach it.

rollback_on_close

Whether to automatically rollback an open global transaction when a connection is closed. If enabled, it mimics the default behaviour of local transactions. Should a client disconnect, the server rolls back any open and unfinished transactions.

Default: true

garbage_collection

max_retries

Maximum number of garbage collection runs before giving up. Allowed values are from 0 to 100. A setting of 0 means no limit, unless the state store enforces a limit. Should the state store enforce a limit, it can be supposed to be significantly higher than 100. Available since 1.6.0.

Please note, it is important to end failed XA transactions within reasonable time to make participating servers free resources bound to the transaction. The built-in garbage collection is not expected to fail for a long period as long as crashed servers become available again quickly. Still, a situation may arise where a human is required to act because the built-in garbage collection stopped or failed. In this case, you may first want to check if the transaction still cannot be fixed by forcing mysqlnd_ms_xa_gc() to ignore the setting, prior to handling it manually.

Default: 5

probability

Garbage collection probability. Allowed values are from 0 to 1000. A setting of 0 disables automatic background garbage collection. Despite a setting of 0 it is still possible to trigger garbage collection by calling mysqlnd_ms_gc(). Available since 1.6.0.

The automatic garbage collection of stalled XA transaction is only available if a state store have been configured. The state store is responsible to keep track of XA transactions. Based on its recordings it can find blocked XA transactions where the client has crashed, connect to the participants and rollback the unfinished transactions.

The garbage collection is triggered as part of PHP's request shutdown procedure at the end of a web request. That is after your PHP script has finished working. Do decide whether to run the garbage collection a random value between 0 and 1000 is computed. If the probability value is higher or equal to the random value, the state stores garbage collection routines are invoked.

Default: 5

max_transactions_per_run

Maximum number of unfinished XA transactions considered by the garbage collection during one run. Allowed values are from 1 to 32768. Available since 1.6.0.

Cleaning up an unfinished XA transaction takes considerable amounts of time and resources. The garbage collection routine may have to connect to several participants of a failed global transaction to issue the SQL commands for rolling back the unfinished tranaction.

Default: 100

Plugin configuration file (<= 1.0.x)

Note:

The below description applies to PECL/mysqlnd_ms < 1.1.0-beta. It is not valid for later versions.

The plugin is using its own configuration file. The configuration file holds information on the MySQL replication master server, the MySQL replication slave servers, the server pick (load balancing) policy, the failover strategy and the use of lazy connections.

The PHP configuration directive mysqlnd_ms.ini_file is used to set the plugins configuration file.

The configuration file mimics standard the php.ini format. It consists of one or more sections. Every section defines its own unit of settings. There is no global section for setting defaults.

Applications reference sections by their name. Applications use section names as the host (server) parameter to the various connect methods of the mysqli, mysql and PDO_MYSQL extensions. Upon connect the mysqlnd plugin compares the hostname with all section names from the plugin configuration file. If hostname and section name match, the plugin will load the sections settings.

[myapp]
master[] = localhost
slave[] = 192.168.2.27
slave[] = 192.168.2.28:3306
[localhost]
master[] = localhost:/tmp/mysql/mysql.sock
slave[] = 192.168.3.24:3305
slave[] = 192.168.3.65:3309

Section names are strings. It is valid to use a section name such as 192.168.2.1, 127.0.0.1 or localhost. If, for example, an application connects to localhost and a plugin configuration section [localhost] exists, the semantics of the connect operation are changed. The application will no longer only use the MySQL server running on the host localhost but the plugin will start to load balance MySQL queries following the rules from the [localhost] configuration section. This way you can load balance queries from an application without changing the applications source code.

The master[], slave[] and pick[] configuration directives use a list-like syntax. Configuration directives supporting list-like syntax may appear multiple times in a configuration section. The plugin maintains the order in which entries appear when interpreting them. For example, the below example shows two slave[] configuration directives in the configuration section [myapp]. If doing round-robin load balancing for read-only queries, the plugin will send the first read-only query to the MySQL server mysql_slave_1 because it is the first in the list. The second read-only query will be send to the MySQL server mysql_slave_2 because it is the second in the list. Configuration directives supporting list-like syntax result are ordered from top to bottom in accordance to their appearance within a configuration section.

[myapp]
master[] = mysql_master_server
slave[] = mysql_slave_1
slave[] = mysql_slave_2

Here is a short explanation of the configuration directives that can be used.

master[] string

URI of a MySQL replication master server. The URI follows the syntax hostname[:port|unix_domain_socket].

The plugin supports using only one master server.

Setting a master server is mandatory. The plugin will report a warning upon connect if the user has failed to provide a master server for a configuration section. The warning may read (mysqlnd_ms) Cannot find master section in config. Furthermore the plugin may set an error code for the connection handle such as HY000/2000 (CR_UNKNOWN_ERROR). The corresponding error message depends on your language settings.

slave[] string

URI of one or more MySQL replication slave servers. The URI follows the syntax hostname[:port|unix_domain_socket].

The plugin supports using one or more slave servers.

Setting a slave server is mandatory. The plugin will report a warning upon connect if the user has failed to provide at least one slave server for a configuration section. The warning may read (mysqlnd_ms) Cannot find slaves section in config. Furthermore the plugin may set an error code for the connection handle such as HY000/2000 (CR_UNKNOWN_ERROR). The corresponding error message depends on your language settings.

pick[] string

Load balancing (server picking) policy. Supported policies: random, random_once (default), roundrobin, user.

If no load balancing policy is set, the plugin will default to random_once. The random_once policy picks a random slave server when running the first read-only statement. The slave server will be used for all read-only statements until the PHP script execution ends.

The random policy will pick a random server whenever a read-only statement is to be executed.

If using roundrobin the plugin iterates over the list of configured slave servers to pick a server for statement execution. If the plugin reaches the end of the list, it wraps around to the beginning of the list and picks the first configured slave server.

Setting more than one load balancing policy for a configuration section makes only sense in conjunction with user and mysqlnd_ms_set_user_pick_server(). If the user defined callback fails to pick a server, the plugin falls back to the second configured load balancing policy.

failover string

Failover policy. Supported policies: disabled (default), master.

If no failover policy is set, the plugin will not do any automatic failover (failover=disabled). Whenever the plugin fails to connect a server it will emit a warning and set the connections error code and message. Thereafter it is up to the application to handle the error and, for example, resent the last statement to trigger the selection of another server.

If using failover=master the plugin will implicitly failover to a slave, if available. Please check the concepts documentation to learn about potential pitfalls and risks of using failover=master.

lazy_connections bool

Controls the use of lazy connections. Lazy connections are connections which are not opened before the client sends the first connection.

It is strongly recommended to use lazy connections. Lazy connections help to keep the number of open connections low. If you disable lazy connections and, for example, configure one MySQL replication master server and two MySQL replication slaves, the plugin will open three connections upon the first call to a connect function although the application might use the master connection only.

Lazy connections bare a risk if you make heavy use of actions which change the state of a connection. The plugin does not dispatch all state changing actions to all connections from the connection pool. The few dispatched actions are applied to already opened connections only. Lazy connections opened in the future are not affected. If, for example, the connection character set is changed using a PHP MySQL API call, the plugin will change the character set of all currently opened connection. It will not remember the character set change to apply it on lazy connections opened in the future. As a result the internal connection pool would hold connections using different character sets. This is not desired. Remember that character sets are taken into account for escaping.

master_on_write bool

If set, the plugin will use the master server only after the first statement has been executed on the master. Applications can still send statements to the slaves using SQL hints to overrule the automatic decision.

The setting may help with replication lag. If an application runs an INSERT the plugin will, by default, use the master to execute all following statements, including SELECT statements. This helps to avoid problems with reads from slaves which have not replicated the INSERT yet.

trx_stickiness string

Transaction stickiness policy. Supported policies: disabled (default), master.

Experimental feature.

The setting requires 5.4.0 or newer. If used with PHP older than 5.4.0, the plugin will emit a warning like (mysqlnd_ms) trx_stickiness strategy is not supported before PHP 5.3.99.

If no transaction stickiness policy is set or, if setting trx_stickiness=disabled, the plugin is not transaction aware. Thus, the plugin may load balance connections and switch connections in the middle of a transaction. The plugin is not transaction safe. SQL hints must be used avoid connection switches during a transaction.

As of PHP 5.4.0 the mysqlnd library allows the plugin to monitor the autocommit mode set by calls to the libraries trx_autocommit() function. If setting trx_stickiness=master and autocommit gets disabled by a PHP MySQL extension invoking the mysqlnd library internal function call trx_autocommit(), the plugin is made aware of the begin of a transaction. Then, the plugin stops load balancing and directs all statements to the master server until autocommit is enabled. Thus, no SQL hints are required.

An example of a PHP MySQL API function calling the mysqlnd library internal function call trx_autocommit() is mysqli_autocommit().

Although setting trx_stickiness=master, the plugin cannot be made aware of autocommit mode changes caused by SQL statements such as SET AUTOCOMMIT=0.

Testing

Note:

The section applies to mysqlnd_ms 1.1.0 or newer, not the 1.0 series.

The PECL/mysqlnd_ms test suite is in the tests/ directory of the source distribution. The test suite consists of standard phpt tests, which are described on the PHP Quality Assurance Teams website.

Running the tests requires setting up one to four MySQL servers. Some tests don't connect to MySQL at all. Others require one server for testing. Some require two distinct servers. In some cases two servers are used to emulate a replication setup. In other cases a master and a slave of an existing MySQL replication setup are required for testing. The tests will try to detect how many servers and what kind of servers are given. If the required servers are not found, the test will be skipped automatically.

Before running the tests, edit tests/config.inc to configure the MySQL servers to be used for testing.

The most basic configuration is as follows.

 putenv("MYSQL_TEST_HOST=localhost");
 putenv("MYSQL_TEST_PORT=3306");
 putenv("MYSQL_TEST_USER=root");
 putenv("MYSQL_TEST_PASSWD=");
 putenv("MYSQL_TEST_DB=test");
 putenv("MYSQL_TEST_ENGINE=MyISAM");
 putenv("MYSQL_TEST_SOCKET=");

 putenv("MYSQL_TEST_SKIP_CONNECT_FAILURE=1");
 putenv("MYSQL_TEST_CONNECT_FLAGS=0");
 putenv("MYSQL_TEST_EXPERIMENTAL=0");

 /* replication cluster emulation */
 putenv("MYSQL_TEST_EMULATED_MASTER_HOST=". getenv("MYSQL_TEST_HOST"));
 putenv("MYSQL_TEST_EMULATED_SLAVE_HOST=". getenv("MYSQL_TEST_HOST"));

 /* real replication cluster */
 putenv("MYSQL_TEST_MASTER_HOST=". getenv("MYSQL_TEST_EMULATED_MASTER_HOST"));
 putenv("MYSQL_TEST_SLAVE_HOST=". getenv("MYSQL_TEST_EMULATED_SLAVE_HOST"));

MYSQL_TEST_HOST, MYSQL_TEST_PORT and MYSQL_TEST_SOCKET define the hostname, TCP/IP port and Unix domain socket of the default database server. MYSQL_TEST_USER and MYSQL_TEST_PASSWD contain the user and password needed to connect to the database/schema configured with MYSQL_TEST_DB. All configured servers must have the same database user configured to give access to the test database.

Using host, host:port or host:/path/to/socket syntax one can set an alternate host, host and port or host and socket for any of the servers.

putenv("MYSQL_TEST_SLAVE_HOST=192.168.78.136:3307"));
putenv("MYSQL_TEST_MASTER_HOST=myserver_hostname:/path/to/socket"));

Debugging and Tracing

The mysqlnd debug log can be used to debug and trace the actitivities of PECL/mysqlnd_ms. As a mysqlnd PECL/mysqlnd_ms adds trace information to the mysqlnd library debug file. Please, see the mysqlnd.debug PHP configuration directive documentation for a detailed description on how to configure the debug log.

Configuration setting example to activate the debug log:

mysqlnd.debug=d:t:x:O,/tmp/mysqlnd.trace

Note:

This feature is only available with a debug build of PHP. Works on Microsoft Windows if using a debug build of PHP and PHP was built using Microsoft Visual C version 9 and above.

The debug log shows mysqlnd library and PECL/mysqlnd_ms plugin function calls, similar to a trace log. Mysqlnd library calls are usually prefixed with mysqlnd_. PECL/mysqlnd internal calls begin with mysqlnd_ms.

Example excerpt from the debug log (connect):

[...]
>mysqlnd_connect
| info : host=myapp user=root db=test port=3306 flags=131072
| >mysqlnd_ms::connect
| | >mysqlnd_ms_config_json_section_exists
| | | info : section=[myapp] len=[5]
| | | >mysqlnd_ms_config_json_sub_section_exists
| | | | info : section=[myapp] len=[5]
| | | | info : ret=1
| | | <mysqlnd_ms_config_json_sub_section_exists
| | | info : ret=1
| | <mysqlnd_ms_config_json_section_exists
[...]

The debug log is not only useful for plugin developers but also to find the cause of user errors. For example, if your application does not do proper error handling and fails to record error messages, checking the debug and trace log may help finding the cause. Use of the debug log to debug application issues should be considered only if no other option is available. Writing the debug log to disk is a slow operation and may have negative impact on the application performance.

Example excerpt from the debug log (connection failure):

[...]
| | | | | | | info : adding error [Access denied for user 'root'@'localhost' (using password: YES)] to the list
| | | | | | | info : PACKET_FREE(0)
| | | | | | | info : PACKET_FREE(0x7f3ef6323f50)
| | | | | | | info : PACKET_FREE(0x7f3ef6324080)
| | | | | | <mysqlnd_auth_handshake
| | | | | | info : switch_to_auth_protocol=n/a
| | | | | | info : conn->error_info.error_no = 1045
| | | | | <mysqlnd_connect_run_authentication
| | | | | info : PACKET_FREE(0x7f3ef63236d8)
| | | | | >mysqlnd_conn::free_contents
| | | | | | >mysqlnd_net::free_contents
| | | | | | <mysqlnd_net::free_contents
| | | | | | info : Freeing memory of members
| | | | | | info : scheme=unix:///tmp/mysql.sock
| | | | | | >mysqlnd_error_list_pdtor
| | | | | | <mysqlnd_error_list_pdtor
| | | | | <mysqlnd_conn::free_contents
| | | | <mysqlnd_conn::connect
[...]

The trace log can also be used to verify correct behaviour of PECL/mysqlnd_ms itself, for example, to check which server has been selected for query execution and why.

Example excerpt from the debug log (plugin decision):

[...]
>mysqlnd_ms::query
| info : query=DROP TABLE IF EXISTS test
| >_mysqlnd_plugin_get_plugin_connection_data
| | info : plugin_id=5
| <_mysqlnd_plugin_get_plugin_connection_data
| >mysqlnd_ms_pick_server_ex
| | info : conn_data=0x7fb6a7d3e5a0 *conn_data=0x7fb6a7d410d0
| | >mysqlnd_ms_select_servers_all
| | <mysqlnd_ms_select_servers_all
| | >mysqlnd_ms_choose_connection_rr
| | | >mysqlnd_ms_query_is_select
[...]
| | | <mysqlnd_ms_query_is_select
[...]
| | | info : Init the master context
| | | info : list(0x7fb6a7d3f598) has 1
| | | info : Using master connection
| | | >mysqlnd_ms_advanced_connect
| | | | >mysqlnd_conn::connect
| | | | | info : host=localhost user=root db=test port=3306 flags=131072 persistent=0 state=0

In this case the statement DROP TABLE IF EXISTS test has been executed. Note that the statement string is shown in the log file. You may want to take measures to restrict access to the log for security considerations.

The statement has been load balanced using round robin policy, as you can easily guess from the functions name >mysqlnd_ms_choose_connection_rr. It has been sent to a master server running on host=localhost user=root db=test port=3306 flags=131072 persistent=0 state=0.

Monitoring

Plugin activity can be monitored using the mysqlnd trace log, mysqlnd statistics, mysqlnd_ms plugin statistics and external PHP debugging tools. Use of the trace log should be limited to debugging. It is recommended to use the plugins statistics for monitoring.

Writing a trace log is a slow operation. If using an external PHP debugging tool, please refer to the vendors manual about its performance impact and the type of information collected. In many cases, external debugging tools will provide call stacks. Often, a call stack or a trace log is more difficult to interpret than the statistics provided by the plugin.

Plugin statistics tell how often which kind of cluster node has been used (slave or master), why the node was used, if lazy connections have been used and if global transaction ID injection has been performed. The monitoring information provided enables user to verify plugin decisions and to plan their cluster resources based on usage pattern. The function mysqlnd_ms_get_stats() is used to access the statistics. Please, see the functions description for a list of available statistics.

Statistics are collected on a per PHP process basis. Their scope is a PHP process. Depending on the PHP deployment model a process may serve one or multiple web requests. If using CGI model, a PHP process serves one web request. If using FastCGI or pre-fork web server models, a PHP process usually serves multiple web requests. The same is the case with a threaded web server. Please, note that threads running in parallel can update the statistics in parallel. Thus, if using a threaded PHP deployment model, statistics can be changed by more than one script at a time. A script cannot rely on the fact that it sees only its own changes to statistics.

mysqlnd_ms.enable=1
mysqlnd_ms.collect_statistics=1

Statistics are aggregated for all plugin activities and all connections handled by the plugin. It is not possible to tell how much a certain connection handle has contributed to the overall statistics.

Utilizing PHPs register_shutdown_function() function or the auto_append_file PHP configuration directive it is easily possible to dump statistics into, for example, a log file when a script finishes. Instead of using a log file it is also possible to send the statistics to an external monitoring tool for recording and display.

mysqlnd_ms.enable=1
mysqlnd_ms.collect_statistics=1
error_log=/tmp/php_errors.log