MaxScale 21.06 MariaDB Monitor
MariaDB Monitor
- MariaDB Monitor
- Overview
- Required Grants
- Master selection
- Configuration
- Common Monitor Parameters
- MariaDB Monitor optional parameters
- assume_unique_hostnames
- detect_stale_master
- detect_stale_slave
- detect_standalone_master
- master_conditions
- slave_conditions
- ignore_external_masters
- failcount
- enforce_writable_master
- enforce_read_only_slaves
- enforce_read_only_servers
- maintenance_on_low_disk_space
- cooperative_monitoring_locks
- script_max_replication_lag
- Cluster manipulation operations
- Operation details
- Manual activation
- Automatic activation
- Limitations and requirements
- External master support
- Configuration parameters
- auto_failover
- auto_rejoin
- switchover_on_low_disk_space
- enforce_simple_topology
- replication_user and replication_password
- replication_master_ssl
- failover_timeout and switchover_timeout
- verify_master_failure and master_failure_timeout
- servers_no_promotion
- promotion_sql_file and demotion_sql_file
- handle_events
- Cooperative monitoring
- Troubleshooting
- Using the MariaDB Monitor With Binlogrouter
Overview
MariaDB Monitor monitors a Master-Slave replication cluster. It probes the state of the backends and assigns server roles such as master and slave, which are used by the routers when deciding where to route a query. It can also modify the replication cluster by performing failover, switchover and rejoin. Backend server versions older than MariaDB/MySQL 5.5 are not supported. Failover and other similar operations require MariaDB 10.0.2 or later.
Up until MariaDB MaxScale 2.2.0, this monitor was called MySQL Monitor.
Required Grants
The monitor user requires the following grant:
CREATE USER 'maxscale'@'maxscalehost' IDENTIFIED BY 'maxscale-password'; GRANT REPLICATION CLIENT ON *.* TO 'maxscale'@'maxscalehost';
In MariaDB Server versions 10.5.0 to 10.5.8, the monitor user instead requires REPLICATION SLAVE ADMIN:
GRANT REPLICATION SLAVE ADMIN ON *.* TO 'maxscale'@'maxscalehost';
In MariaDB Server 10.5.9 and later, REPLICA MONITOR is required:
GRANT REPLICA MONITOR ON *.* TO 'maxscale'@'maxscalehost';
If the monitor needs to query server disk space (i.e. disk_space_threshold
is
set), then the FILE-grant is required with MariaDB Server versions 10.4.7,
10.3.17, 10.2.26 and 10.1.41 and later.
GRANT FILE ON *.* TO 'maxscale'@'maxscalehost';
MariaDB Server 10.5.2 introduces CONNECTION ADMIN. This is recommended since it allows the monitor to log in even if server connection limit has been reached.
GRANT CONNECTION ADMIN ON *.* TO 'maxscale'@'maxscalehost';
Cluster Manipulation Grants
If cluster manipulation operations are used, the following additional grants are required:
GRANT SUPER, RELOAD, PROCESS, SHOW DATABASES, EVENT ON *.* TO 'maxscale'@'maxscalehost'; GRANT SELECT ON mysql.user TO 'maxscale'@'maxscalehost';
As of MariaDB Server 11.0.1, the SUPER-privilege no longer contains several of its former sub-privileges. These must be given separately.
GRANT RELOAD, PROCESS, SHOW DATABASES, EVENT, SET USER, READ_ONLY ADMIN ON *.* TO 'maxscale'@'maxscalehost'; GRANT REPLICATION SLAVE ADMIN, BINLOG ADMIN, CONNECTION ADMIN ON *.* TO 'maxscale'@'maxscalehost'; GRANT SELECT ON mysql.user TO 'maxscale'@'maxscalehost';
If a separate replication user is defined (with replication_user
and
replication_password
), it requires the following grant:
CREATE USER 'replication'@'replicationhost' IDENTIFIED BY 'replication-password'; GRANT REPLICATION SLAVE ON *.* TO 'replication'@'replicationhost';
Master selection
Only one backend can be master at any given time. A master must be running
(successfully connected to by the monitor) and its read_only-setting must be
off. A master may not be replicating from another server in the monitored
cluster unless the master is part of a multimaster group. Master selection
prefers to select the server with the most slaves, possibly in multiple
replication layers. Only slaves reachable by a chain of running relays or
directly connected to the master count. When multiple servers are tied for
master status, the server which appears earlier in the servers
-setting of the
monitor is selected.
Servers in a cyclical replication topology (multimaster group) are interpreted as having all the servers in the group as slaves. Even from a multimaster group only one server is selected as the overall master.
After a master has been selected, the monitor prefers to stick with the choice even if other potential masters with more slave servers are available. Only if the current master is clearly unsuitable does the monitor try to select another master. An existing master turns invalid if:
- It is unwritable (read_only is on).
- It has been down for more than failcount monitor passes and has no running slaves. Running slaves behind a downed relay count. A slave in this context is any server with at least a partially running replication connection (either io or sql thread is running). The slave servers must also be down for more than failcount monitor passes to allow new master selection.
- It did not previously replicate from another server in the cluster but it is now replicating.
- It was previously part of a multimaster group but is no longer, or the multimaster group is replicating from a server not in the group.
Cases 1 and 2 cover the situations in which the DBA, an external script or even another MaxScale has modified the cluster such that the old master can no longer act as master. Cases 3 and 4 are less severe. In these cases the topology has changed significantly and the master should be re-selected, although the old master may still be the best choice.
The master change described above is different from failover and switchover described in section Failover, switchover and auto-rejoin. A master change only modifies the server roles inside MaxScale but does not modify the cluster other than changing the targets of read and write queries. Failover and switchover perform a master change on their own.
As a general rule, it's best to avoid situations where the cluster has multiple standalone servers, separate master-slave pairs or separate multimaster groups. Due to master invalidation rule 2, a standalone master can easily lose the master status to another valid master if it goes down. The new master probably does not have the same data as the previous one. Non-standalone masters are less vulnerable, as a single running slave or multimaster group member will keep the master valid even when down.
Configuration
A minimal configuration for a monitor requires a set of servers for monitoring and a username and a password to connect to these servers.
[MyMonitor] type=monitor module=mariadbmon servers=server1,server2,server3 user=myuser password=mypwd
From MaxScale 2.2.1 onwards, the module name is mariadbmon
instead of
mysqlmon
. The old name can still be used.
The grants required by user
depend on which monitor features are used. A full
list of the grants can be found in the Required Grants
section.
Common Monitor Parameters
For a list of optional parameters that all monitors support, read the Monitor Common document.
MariaDB Monitor optional parameters
These are optional parameters specific to the MariaDB Monitor. Failover, switchover and rejoin-specific parameters are listed in their own section.
assume_unique_hostnames
Boolean, default: ON. When active, the monitor assumes that server hostnames and
ports are consistent between the server definitions in the MaxScale
configuration file and the "SHOW ALL SLAVES STATUS" outputs of the servers
themselves. Specifically, the monitor assumes that if server A is replicating
from server B, then A must have a slave connection with Master_Host
and
Master_Port
equal to B's address and port in the configuration file. If this
is not the case, e.g. an IP is used in the server while a hostname is given in
the file, the monitor may misinterpret the topology. In MaxScale 2.4.1, the
monitor attempts name resolution on the addresses if a simple string comparison
does not find a match. Using exact matching addresses is, however, more
reliable.
This setting must be ON to use any cluster operation features such as failover or switchover, because MaxScale uses the addresses and ports in the configuration file when issuing "CHANGE MASTER TO"-commands.
If the network configuration is such that the addresses MaxScale uses to connect
to backends are different from the ones the servers use to connect to each
other, assume_unique_hostnames
should be set to OFF. In this mode, MaxScale
uses server id:s it queries from the servers and the Master_Server_Id
fields
of the slave connections to deduce which server is replicating from which. This
is not perfect though, since MaxScale doesn't know the id:s of servers it has
never connected to (e.g. server has been down since MaxScale was started). Also,
the Master_Server_Id
-field may have an incorrect value if the slave connection
has not been established. MaxScale will only trust the value if the monitor has
seen the slave connection IO thread connected at least once. If this is not the
case, the slave connection is ignored.
detect_stale_master
Boolean, default: ON. Deprecated. If set to OFF, running_slave is added to master_conditions. Both settings may not be defined simultaneously.
detect_stale_slave
Boolean, default: ON. Deprecated. If set to OFF, linked_master is added to slave_conditions. Both settings may not be defined simultaneously.
detect_standalone_master
Boolean, default: ON. Deprecated. If set to OFF, connecting_slave is added to master_conditions. Both settings may not be defined simultaneously.
master_conditions
Enum, default: primary_monitor_master. Designate additional conditions for Master-status, i.e qualified for read and write queries.
Normally, if a suitable master candidate server is found as described in Master selection, MaxScale designates it Master. master_conditions sets additional conditions for a master server. This setting is an enum, allowing multiple conditions to be set simultaneously. Conditions 2, 3 and 4 refer to slave servers. If combined, a single slave must fulfill all of the given conditions for the master to be viable.
If the master candidate fails master_conditions but fulfills slave_conditions, it may be designated Slave instead.
The available conditions are:
- none : No additional conditions
- connecting_slave : At least one immediate slave (not behind relay) is attempting to replicate or is replicating from the master (Slave_IO_Running is 'Yes' or 'Connecting', Slave_SQL_Running is 'Yes'). A slave with incorrect replication credentials does not count. If the slave is currently down, results from the last successful monitor tick are used.
- connected_slave : Same as above, with the difference that the replication connection must be up (Slave_IO_Running is 'Yes'). If the slave is currently down, results from the last successful monitor tick are used.
- running_slave : Same as connecting_slave, with the addition that the slave must also be Running.
- primary_monitor_master : If this MaxScale is cooperating with another MaxScale and this is the secondary MaxScale, require that the candidate master is selected also by the primary MaxScale.
The default value of this setting is
master_requirements=primary_monitor_master
to ensure that both monitors use
the same master server when cooperating.
For example, to require that the master must have a slave which is both connected and running, set
master_conditions=connected_slave,running_slave
slave_conditions
Enum, default: none. Designate additional conditions for Slave-status, i.e qualified for read queries.
Normally, a server is Slave if it is at least attempting to replicate from the master candidate or a relay (Slave_IO_Running is 'Yes' or 'Connecting', Slave_SQL_Running is 'Yes', valid replication credentials). The master candidate does not necessarily need to be writable, e.g. if it fails its master_conditions. slave_conditions sets additional conditions for a slave server. This setting is an enum, allowing multiple conditions to be set simultaneously.
The available conditions are:
- none : No additional conditions. This is the default value.
- linked_master : The slave must be connected to the master (Slave_IO_Running and Slave_SQL_Running are 'Yes') and the master must be Running. The same applies to any relays between the slave and the master.
- running_master : The master must be running. Relays may be down.
- writable_master : The master must be writable, i.e. labeled Master.
- primary_monitor_master : If this MaxScale is cooperating with another MaxScale and this is the secondary MaxScale, require that the candidate master is selected also by the primary MaxScale.
For example, to require that the master server of the cluster must be running and writable for any servers to have Slave-status, set
slave_conditions=running_master,writable_master
ignore_external_masters
Boolean, default: OFF. Deprecated. Ignore any servers that are not monitored by this monitor but are a part of the replication topology.
An external server is a server not monitored by this monitor. If a server is replicating from an external server, it typically gains the Slave of External Server-status. If this setting is enabled, the status is not set.
failcount
Number of consecutive monitor passes a master server must be down before it is
considered failed. If automatic failover is enabled (auto_failover=true
), it
may be performed at this time. A value of 0 or 1 enables immediate failover.
If automatic failover is not possible, the monitor will try to search for another server to fulfill the master role. See section Master selection for more details. Changing the master may break replication as queries could be routed to a server without previous events. To prevent this, avoid having multiple valid master servers in the cluster.
The default value is 5 failures.
The worst-case delay between the master failure and the start of the failover
can be estimated by summing up the timeout values and monitor_interval
and
multiplying that by failcount
:
(monitor_interval + backend_connect_timeout) * failcount
enforce_writable_master
This feature is disabled by default. If set to ON, the monitor attempts to disable the read_only-flag on the master when seen. The flag is checked every monitor tick. The monitor user requires the SUPER-privilege for this feature to work.
Typically, the master server should never be in read-only-mode. Such a situation may arise due to misconfiguration or accident, or perhaps if MaxScale crashed during switchover.
When this feature is enabled, setting the master manually to read_only will no longer cause the monitor to search for another master. The master will instead for a moment lose its [Master]-status (no writes), until the monitor again enables writes on the master. When starting from scratch, the monitor still prefers to select a writable server as master if possible.
enforce_read_only_slaves
This feature is disabled by default. If set to ON, the monitor attempts to enable the read_only-flag on any writable slave server. The flag is checked every monitor tick. The monitor user requires the SUPER-privilege (or READ_ONLY ADMIN) for this feature to work. While the read_only-flag is ON, only users with the SUPER-privilege (or READ_ONLY ADMIN) can write to the backend server. If temporary write access is required, this feature should be disabled before attempting to disable read_only manually. Otherwise, the monitor will quickly re-enable it.
read_only won't be enabled on the master server, even if it has lost [Master]-status due to master_conditions and is marked [Slave].
enforce_read_only_servers
Boolean, default: false. Works similar to enforce_read_only_slaves except will set read_only on any writable server that is not the primary and not in maintenance (a superset of the servers altered by enforce_read_only_slaves).
The monitor user requires the SUPER-privilege (or READ_ONLY ADMIN) for this feature to work. If the cluster has no valid primary or primary candidate, read_only is not set on any server as it is unclear which servers should be altered.
maintenance_on_low_disk_space
This feature is enabled by default. If a running server that is not the master or a relay master is out of disk space the server is set to maintenance mode. Such servers are not used for router sessions and are ignored when performing a failover or other cluster modification operation. See the general monitor parameters disk_space_threshold and disk_space_check_interval on how to enable disk space monitoring.
Once a server has been put to maintenance mode, the disk space situation of that server is no longer updated. The server will not be taken out of maintenance mode even if more disk space becomes available. The maintenance flag must be removed manually:
maxctrl clear server server2 Maint
cooperative_monitoring_locks
Using this setting is recommended when multiple MaxScales are monitoring the same backend cluster. When enabled, the monitor attempts to acquire exclusive locks on the backend servers. The monitor considers itself the primary monitor if it has a majority of locks. The majority can be either over all configured servers or just over running servers. See Cooperative monitoring for more details on how this feature works and which value to use.
Allowed values:
1. none
Default value, no locking.
2. majority_of_all
Primary monitor requires majority of locks, even counting
servers which are [Down].
3. majority_of_running
Primary monitor requires majority of locks over
[Running] servers.
This setting is separate from the global MaxScale setting passive. If passive is set, cluster operations are disabled even if monitor has acquired the locks. Generally, it's best not to mix cooperative monitoring with the passive-setting.
script_max_replication_lag
Integer, default: -1. Defines a replication lag limit in seconds for launching the monitor script configured in the script-parameter. If the replication lag of a server goes above this limit, the script is ran with the $EVENT-placeholder replaced by "rlag_above". If the lag goes back below the limit, the script is ran again with replacement "rlag_below".
Negative values disable this feature. For more information on monitor scripts, see general monitor documentation.
Cluster manipulation operations
Starting with MaxScale 2.2.1, MariaDB Monitor supports replication cluster modification. The operations implemented are: - failover, which replaces a failed master with a slave - switchover, which swaps a running master with a slave - async-switchover, which schedules a switchover and returns - rejoin, which directs servers to replicate from the master - reset-replication (added in MaxScale 2.3.0), which deletes binary logs and resets gtid:s
See operation details for more information on the implementation of the commands.
The cluster operations require that the monitor user (user
) has the following
privileges:
- SUPER, to modify slave connections, set globals such as read_only and kill connections from other super-users
- REPLICATION CLIENT (REPLICATION SLAVE ADMIN in MariaDB Server 10.5), to list slave connections
- RELOAD, to flush binary logs
- PROCESS, to check if the event_scheduler process is running
- SHOW DATABASES and EVENT, to list and modify server events
- SELECT on mysql.user, to see which users have SUPER
A list of the grants can be found in the Required Grants section.
The privilege system was changed in MariaDB Server 10.5. The effects of this on the MaxScale monitor user are minor, as the SUPER-privilege contains many of the required privileges and is still required to kill connections from other super-users.
In MariaDB Server 11.0.1 and later, SUPER no longer contains all the required grants. The monitor requires:
- READ_ONLY ADMIN, to set read_only
- REPLICA MONITOR and REPLICATION SLAVE ADMIN, to view and manage replication connections
- RELOAD, to flush binary logs
- PROCESS, to check if the event_scheduler process is running
- SHOW DATABASES, EVENT and SET USER, to list and modify server events
- BINLOG ADMIN, to delete binary logs (during reset-replication)
- CONNECTION ADMIN, to kill connections
- SELECT on mysql.user, to see which users have SUPER
In addition, the monitor needs to know which username and password a
slave should use when starting replication. These are given in
replication_user
and replication_password
.
The user can define files with SQL statements which are executed on any server
being demoted or promoted by cluster manipulation commands. See the sections on
promotion_sql_file
and demotion_sql_file
for more information.
The monitor can manipulate scheduled server events when promoting or demoting a
server. See the section on handle_events
for more information.
All cluster operations can be activated manually through MaxCtrl. See section Manual activation for more details.
See Limitations and requirements for information on possible issues with failover and switchover.
Operation details
Failover replaces a failed master with a running slave. It does the following:
- Select the most up-to-date slave of the old master to be the new master. The
selection criteria is as follows in descending priority:
- gtid_IO_pos (latest event in relay log)
- gtid_current_pos (most processed events)
- log_slave_updates is on
- disk space is not low
- If the new master has unprocessed relay log items, cancel and try again later.
- Prepare the new master:
- Remove the slave connection the new master used to replicate from the old master.
- Disable the read_only-flag.
- Enable scheduled server events (if event handling is on). Only events that were enabled on the old master are enabled.
- Run the commands in
promotion_sql_file
. - Start replication from external master if one existed.
- Redirect all other slaves to replicate from the new master:
- STOP SLAVE and RESET SLAVE
- CHANGE MASTER TO
- START SLAVE
- Check that all slaves are replicating.
Failover is considered successful if steps 1 to 3 succeed, as the cluster then has at least a valid master server.
Switchover swaps a running master with a running slave. It does the following:
- Prepare the old master for demotion:
- Stop any external replication.
- Kill connections from super-users since read_only does not affect them.
- Enable the read_only-flag to stop writes.
- Disable scheduled server events (if event handling is on).
- Run the commands in
demotion_sql_file
. - Flush the binary log (FLUSH LOGS) so that all events are on disk.
- Wait for the new master to catch up with the old master.
- Promote new master and redirect slaves as in failover steps 3 and 4. Also redirect the demoted old master.
- Check that all slaves are replicating.
Similar to failover, switchover is considered successful if the new master was successfully promoted.
Rejoin joins a standalone server to the cluster or redirects a slave replicating from a server other than the master. A standalone server is joined by:
- Run the commands in
demotion_sql_file
. - Enable the read_only-flag.
- Disable scheduled server events (if event handling is on).
- Start replication: CHANGE MASTER TO and START SLAVE.
A server which is replicating from the wrong master is redirected simply with STOP SLAVE, RESET SLAVE, CHANGE MASTER TO and START SLAVE commands.
Reset-replication (added in MaxScale 2.3.0) deletes binary logs and resets gtid:s. This destructive command is meant for situations where the gtid:s in the cluster are out of sync while the actual data is known to be in sync. The operation proceeds as follows:
- Reset gtid:s and delete binary logs on all servers:
- Stop (STOP SLAVE) and delete (RESET SLAVE ALL) all slave connections.
- Enable the read_only-flag.
- Disable scheduled server events (if event handling is on).
- Delete binary logs (RESET MASTER).
- Set the sequence number of gtid_slave_pos to zero. This also affects gtid_current_pos.
- Prepare new master:
- Disable the read_only-flag.
- Enable scheduled server events (if event handling is on). Events are only enabled if the cluster had a master server when starting the reset-replication operation. Only events that were enabled on the previous master are enabled on the new.
- Direct other servers to replicate from the new master as in the other operations.
Manual activation
Cluster operations can be activated manually through the REST API or MaxCtrl. The commands are only performed when MaxScale is in active mode. The commands generally match their automatic versions. The exception is rejoin, in which the manual command allows rejoining even when the joining server has empty gtid:s. This rule allows the user to force a rejoin on a server without binary logs.
All commands require the monitor instance name as the first parameter. Failover selects the new master server automatically and does not require additional parameters. Rejoin requires the name of the joining server as second parameter. Replication reset accepts the name of the new master server as second parameter. If not given, the current master is selected.
Switchover takes one to three parameters. If only the monitor name is given, switchover will autoselect both the slave to promote and the current master as the server to be demoted. If two parameters are given, the second parameter is interpreted as the slave to promote. If three parameters are given, the third parameter is interpreted as the current master. The user-given current master is compared to the master server currently deduced by the monitor and if the two are unequal, an error is given.
Example commands are below:
call command mariadbmon failover MyMonitor call command mariadbmon rejoin MyMonitor OldMasterServ call command mariadbmon reset-replication MyMonitor call command mariadbmon reset-replication MyMonitor NewMasterServ call command mariadbmon switchover MyMonitor call command mariadbmon switchover MyMonitor NewMasterServ call command mariadbmon switchover MyMonitor NewMasterServ OldMasterServ
The commands follow the standard module command syntax. All require the monitor configuration name (MyMonitor) as the first parameter. For switchover, the last two parameters define the server to promote (NewMasterServ) and the server to demote (OldMasterServ). For rejoin, the server to join (OldMasterServ) is required. Replication reset requires the server to promote (NewMasterServ).
It is safe to perform manual operations even with automatic failover, switchover or rejoin enabled since automatic operations cannot happen simultaneously with manual ones.
When a cluster modification is initiated via the REST-API, the URL path is of the form:
/v1/maxscale/modules/mariadbmon/<operation>?<monitor-instance>&<server-param1>&<server-param2>
<operation>
is the name of the command: failover, switchover, rejoin or reset-replication.<monitor-instance>
is the monitor section name from the MaxScale configuration file.<server-param1>
and<server-param2>
are server parameters as described above for MaxCtrl. Only switchover accepts both, failover doesn't need any and both rejoin and reset-replication accept one.
Given a MaxScale configuration file like
[Cluster1] type=monitor module=mariadbmon servers=server1, server2, server3, server 4 ...
with the assumption that server2
is the current master, then the URL
path for making server4
the new master would be:
/v1/maxscale/modules/mariadbmon/switchover?Cluster1&server4&server2
Example REST-API paths for other commands are listed below.
/v1/maxscale/modules/mariadbmon/failover?Cluster1 /v1/maxscale/modules/mariadbmon/rejoin?Cluster1&server3 /v1/maxscale/modules/mariadbmon/reset-replication?Cluster1&server3
Queued switchover
Most cluster modification commands wait until the operation either succeeds or fails. async-switchover is an exception, as it returns immediately. Otherwise async-switchover works identical to a normal switchover command. Use the module command fetch-cmd-result to view the result of the queued command. fetch-cmd-result returns the status or result of the latest manual command, whether queued or not.
maxctrl call command mariadbmon async-switchover Cluster1 OK maxctrl call command mariadbmon fetch-cmd-result Cluster1 { "links": { "self": "http://localhost:8989/v1/maxscale/modules/mariadbmon/fetch-cmd-result" }, "meta": "switchover completed successfully." }
Automatic activation
Failover can activate automatically if auto_failover
is on. The activation
begins when the master has been down at least failcount
monitor iterations.
Before modifying the cluster, the monitor checks that all prerequisites for the
failover are fulfilled. If the cluster does not seem ready, an error is printed
and the cluster is rechecked during the next monitor iteration.
Switchover can also activate automatically with the
switchover_on_low_disk_space
-setting. The operation begins if the master
server is low on disk space but otherwise the operating logic is quite similar
to automatic failover.
Rejoin stands for starting replication on a standalone server or redirecting a slave replicating from the wrong master (any server that is not the cluster master). The rejoined servers are directed to replicate from the current cluster master server, forcing the replication topology to a 1-master-N-slaves configuration.
A server is categorized as standalone if the server has no slave connections, not even stopped ones. A server is replicating from the wrong master if the slave IO thread is connected but the master server id seen by the slave does not match the cluster master id. Alternatively, the IO thread may be stopped or connecting but the master server host or port information differs from the cluster master info. These criteria mean that a STOP SLAVE does not yet set a slave as standalone.
With auto_rejoin
active, the monitor will try to rejoin any servers matching
the above requirements. Rejoin does not obey failcount
and will attempt to
rejoin any valid servers immediately. When activating rejoin manually, the
user-designated server must fulfill the same requirements.
Limitations and requirements
Switchover and failover only understand simple topologies. They will not work if
the cluster has multiple masters, relay masters, or if the topology is circular.
The server cluster is assumed to be well-behaving with no significant
replication lag and all commands that modify the cluster complete in a few
seconds (faster than backend_read_timeout
and backend_write_timeout
).
The backends must all use GTID-based replication, and the domain id should not change during a switchover or failover. Master and slaves must have well-behaving GTIDs with no extra events on slave servers.
Failover cannot be performed if MaxScale was started only after the master server went down. This is because MaxScale needs reliable information on the gtid domain of the cluster and the replication topology in general to properly select the new master.
Failover may lose events. If a master goes down before sending new events to at least one slave, those events are lost when a new master is chosen. If the old master comes back online, the other servers have likely moved on with a diverging history and the old master can no longer join the replication cluster.
To reduce the chance of losing data, use semisynchronous replication. In semisynchronous mode, the master waits for a slave to receive an event before returning an acknowledgement to the client. This does not yet guarantee a clean failover. If the master fails after preparing a transaction but before receiving slave acknowledgement, it will still commit the prepared transaction as part of its crash recovery. Since the slaves may never have seen this transaction, the old master has diverged from the slaves. See Configuring the Master Wait Point for more information.
Even a controlled shutdown of the master may lose events. The server does not by default wait for all data to be replicated to the slaves when shutting down and instead simply closes all connections. Before shutting down the master with the intention of having a slave promoted, run switchover first to ensure that all data is replicated. For more information on server shutdown, see Binary Log Dump Threads and the Shutdown Process.
Switchover requires that the cluster is "frozen" for the duration of the operation. This means that no data modifying statements such as INSERT or UPDATE are executed and the GTID position of the master server is stable. When switchover begins, the monitor sets the global read_only flag on the old master backend to stop any updates. read_only does not affect users with the SUPER-privilege so any such user can issue writes during a switchover. These writes have a high chance of breaking replication, because the write may not be replicated to all slaves before they switch to the new master. To prevent this, any users who commonly do updates should not have the SUPER-privilege. For even more security, the only SUPER-user session during a switchover should be the MaxScale monitor user. This also applies to users running scheduled server events. Although the monitor by default disables events on the master, an event may already be executing. If the event definer has SUPER-privilege, the event can write to the database even through read_only.
When mixing rejoin with failover/switchover, the backends should have log_slave_updates on. The rejoining server is likely lagging behind the rest of the cluster. If the current cluster master does not have binary logs from the moment the rejoining server lost connection, the rejoining server cannot continue replication. This is an issue if the master has changed and the new master does not have log_slave_updates on.
If an automatic cluster operation such as auto-failover or auto-rejoin fails,
all cluster modifying operations are disabled for failcount
monitor iterations,
after which the operation may be retried. Similar logic applies if the cluster is
unsuitable for such operations, e.g. replication is not using GTID.
External master support
The monitor detects if a server in the cluster is replicating from an external master (a server that is not monitored by the monitor). If the replicating server is the cluster master server, then the cluster itself is considered to have an external master.
If a failover/switchover happens, the new master server is set to replicate from
the cluster external master server. The username and password for the replication
are defined in replication_user
and replication_password
. The address and
port used are the ones shown by SHOW ALL SLAVES STATUS
on the old cluster
master server. In the case of switchover, the old master also stops replicating
from the external server to preserve the topology.
After failover the new master is replicating from the external master. If the failed old master comes back online, it is also replicating from the external server. To normalize the situation, either have auto_rejoin on or manually execute a rejoin. This will redirect the old master to the current cluster master.
Configuration parameters
auto_failover
Enable automated master failover. This parameter expects a boolean value and the default value is false.
When automatic failover is enabled, traditional MariaDB Master-Slave clusters
will automatically elect a new master if the old master goes down and stays down
a number of iterations given in failcount
. Failover will not take place when
MaxScale is configured as a passive instance. For details on how MaxScale
behaves in passive mode, see the documentation on failover_timeout
below.
The monitor user must have the SUPER and RELOAD privileges for failover to work.
auto_rejoin
Enable automatic joining of server to the cluster. This parameter expects a boolean value and the default value is false.
When enabled, the monitor will attempt to direct standalone servers and servers replicating from a relay master to the main cluster master server, enforcing a 1-master-N-slaves configuration.
For example, consider the following event series.
- Slave A goes down
- Master goes down and a failover is performed, promoting Slave B
- Slave A comes back
Slave A is still trying to replicate from the downed master, since it wasn't
online during failover. If auto_rejoin
is on, Slave A will quickly be
redirected to Slave B, the current master.
switchover_on_low_disk_space
This feature is disabled by default. If enabled, the monitor will attempt to
switchover a master server low on disk space with a slave. The switch is only
done if a slave without disk space issues is found. If
maintenance_on_low_disk_space
is also enabled, the old master (now a slave)
will be put to maintenance during the next monitor iteration.
For this parameter to have any effect, disk_space_threshold
must be specified
for the server
or the monitor.
Also, disk_space_check_interval
must be defined for the monitor.
switchover_on_low_disk_space=true
enforce_simple_topology
This setting tells the monitor to assume that the servers should be arranged in a
1-master-N-slaves topology and the monitor should try to keep it that way. If
enforce_simple_topology
is enabled, the settings assume_unique_hostnames
,
auto_failover
and auto_rejoin
are also activated regardless of their individual
settings.
By default, mariadbmon will not rejoin servers with more than one replication
stream configured into the cluster. Starting with MaxScale 6.2.0, when
enforce_simple_topology
is enabled, all servers will be rejoined into the
cluster and any extra replication sources will be removed. This is done to make
automated failover with multi-source external replication possible.
This setting also allows the monitor to perform a failover to a cluster where the master server has not been seen [Running]. This is usually the case when the master goes down before MaxScale is started. When using this feature, the monitor will guess the GTID domain id of the master from the slaves. For reliable results, the GTID:s of the cluster should be simple.
enforce_simple_topology=true
replication_user
and replication_password
The username and password of the replication user. These are given as the values
for MASTER_USER
and MASTER_PASSWORD
whenever a CHANGE MASTER TO
command is
executed.
Both replication_user
and replication_password
parameters must be defined if
a custom replication user is used. If neither of the parameters is defined, the
CHANGE MASTER TO
-command will use the monitor credentials for the replication
user.
The credentials used for replication must have the REPLICATION SLAVE
privilege.
replication_password
uses the same encryption scheme as other password
parameters. If password encryption is in use, replication_password
must be
encrypted with the same key to avoid erroneous decryption.
replication_master_ssl
Type: bool Default: off
If set to ON, any CHANGE MASTER TO
-command generated will set MASTER_SSL=1
to enable
encryption for the replication stream. This setting should only be enabled if the backend
servers are configured for ssl. This typically means setting ssl_ca, ssl_cert and
ssl_key in the server configuration file. Additionally, credentials for the replication
user should require an encrypted connection (e.g. ALTER USER repl@'%' REQUIRE SSL;
).
If the setting is left OFF, MASTER_SSL
is not set at all, which will preserve existing
settings when redirecting a slave connection.
failover_timeout
and switchover_timeout
Time limit for failover and switchover operations. The default
values are 90 seconds for both. switchover_timeout
is also used as the time
limit for a rejoin operation. Rejoin should rarely time out, since it is a
faster operation than switchover.
The timeouts are specified as documented here. If no explicit unit is provided, the value is interpreted as seconds in MaxScale 2.4. In subsequent versions a value without a unit may be rejected. Note that since the granularity of the timeouts is seconds, a timeout specified in milliseconds will be rejected, even if the duration is longer than a second.
If no successful failover/switchover takes place within the configured time period, a message is logged and automatic failover is disabled. This prevents further automatic modifications to the misbehaving cluster.
verify_master_failure
and master_failure_timeout
Enable additional master failure verification for automatic failover.
verify_master_failure
is a boolean value (default: true) which enables this
feature and master_failure_timeout
defines the timeout (default: 10 seconds).
The master failure timeout is specified as documented here. If no explicit unit is provided, the value is interpreted as seconds in MaxScale 2.4. In subsequent versions a value without a unit may be rejected. Note that since the granularity of the timeout is seconds, a timeout specified in milliseconds will be rejected, even if the duration is longer than a second.
Failure verification is performed by checking whether the slave servers are
still connected to the master and receiving events. An event is either a change
in the Gtid_IO_Pos-field of the SHOW SLAVE STATUS
output or a heartbeat
event. Effectively, if a slave has received an event within
master_failure_timeout
duration, the master is not considered down when
deciding whether to failover, even if MaxScale cannot connect to the master.
master_failure_timeout
should be longer than the Slave_heartbeat_period
of
the slave connection to be effective.
If every slave loses its connection to the master (Slave_IO_Running is not "Yes"), master failure is considered verified regardless of timeout. This allows faster failover when the master properly disconnects.
For automatic failover to activate, the failcount
requirement must also be
met.
servers_no_promotion
This is a comma-separated list of server names that will not be chosen for master promotion during a failover or autoselected for switchover. This does not affect switchover if the user selects the server to promote. Using this setting can disrupt new master selection for failover such that an non-optimal server is chosen. At worst, this will cause replication to break. Alternatively, failover may fail if all valid promotion candidates are in the exclusion list.
servers_no_promotion=backup_dc_server1,backup_dc_server2
promotion_sql_file
and demotion_sql_file
These optional settings are paths to text files with SQL statements in them. During promotion or demotion, the contents are read line-by-line and executed on the backend. Use these settings to execute custom statements on the servers to complement the built-in operations.
Empty lines or lines starting with '#' are ignored. Any results returned by the statements are ignored. All statements must succeed for the failover, switchover or rejoin to continue. The monitor user may require additional privileges and grants for the custom commands to succeed.
When promoting a slave to master during switchover or failover, the
promotion_sql_file
is read and executed on the new master server after its
read-only flag is disabled. The commands are ran before starting replication
from an external master if any.
demotion_sql_file
is ran on an old master during demotion to slave, before the
old master starts replicating from the new master. The file is also ran before
rejoining a standalone server to the cluster, as the standalone server is
typically a former master server. When redirecting a slave replicating from a
wrong master, the sql-file is not executed.
Since the queries in the files are ran during operations which modify
replication topology, care is required. If promotion_sql_file
contains data
modification (DML) queries, the new master server may not be able to
successfully replicate from an external master. demotion_sql_file
should never
contain DML queries, as these may not replicate to the slave servers before
slave threads are stopped, breaking replication.
promotion_sql_file=/home/root/scripts/promotion.sql demotion_sql_file=/home/root/scripts/demotion.sql
handle_events
This setting is on by default. If enabled, the monitor continuously queries the servers for enabled scheduled events and uses this information when performing cluster operations, enabling and disabling events as appropriate.
When a server is being demoted, any events with "ENABLED" status are set to "SLAVESIDE_DISABLED". When a server is being promoted to master, events that are either "SLAVESIDE_DISABLED" or "DISABLED" are set to "ENABLED" if the same event was also enabled on the old master server last time it was successfully queried. Events are considered identical if they have the same schema and name. When a standalone server is rejoined to the cluster, its events are also disabled since it is now a slave.
The monitor does not check whether the same events were disabled and enabled during a switchover or failover/rejoin. All events that meet the criteria above are altered.
The monitor does not enable or disable the event scheduler itself. For the events to run on the new master server, the scheduler should be enabled by the admin. Enabling it in the server configuration file is recommended.
Events running at high frequency may cause replication to break in a failover scenario. If an old master which was failed over restarts, its event scheduler will be on if set in the server configuration file. Its events will also remember their "ENABLED"-status and run when scheduled. This may happen before the monitor rejoins the server and disables the events. This should only be an issue for events running more often than the monitor interval or events that run immediately after the server has restarted.
Cooperative monitoring
As of MaxScale 2.5, MariaDB-Monitor supports cooperative monitoring. This means that multiple monitors (typically in different MaxScale instances) can monitor the same backend server cluster and only one will be the primary monitor. Only the primary monitor may perform switchover, failover or rejoin operations. The primary also decides which server is the master. Cooperative monitoring is enabled with the cooperative_monitoring_locks-setting. Even with this setting, only one monitor per server per MaxScale is allowed. This limitation can be circumvented by defining multiple copies of a server in the configuration file.
Cooperative monitoring uses server locks for coordinating between monitors. When cooperating, the monitor regularly checks the status of a lock named maxscale_mariadbmonitor on every server and acquires it if free. If the monitor acquires a majority of locks, it is the primary. If a monitor cannot claim majority locks, it is a secondary monitor.
The primary monitor of a cluster also acquires the lock maxscale_mariadbmonitor_master on the master server. Secondary monitors check which server this lock is taken on and only accept that server as the master. This arrangement is required so that multiple monitors can agree on which server is the master regardless of replication topology. If a secondary monitor does not see the master-lock taken, then it won't mark any server as [Master], causing writes to fail.
The lock-setting defines how many locks are required for primary status. Setting
cooperative_monitoring_locks=majority_of_all
means that the primary monitor
needs n_servers/2 + 1 (rounded down) locks. For example, a cluster of three
servers needs two locks for majority, a cluster of four needs three, and a
cluster of five needs three.
This scheme is resistant against split-brain situations in the sense
that multiple monitors cannot be primary simultaneously. However, a split may
cause both monitors to consider themselves secondary, in which case a master
server won't be detected.
Even without a network split, cooperative_monitoring_locks=majority_of_all
will lead to neither monitor claiming lock majority once too many servers go
down. This scenario is depicted in the image below. Only two out of four servers
are running when three are needed for majority. Although both MaxScales see both
running servers, neither is certain they have majority and the cluster stays in
read-only mode. If the primary server is down, no failover is performed either.
Setting cooperative_monitoring_locks=majority_of_running
changes the way
n_servers is calculated. Instead of using the total number of servers, only
servers currently [Running] are considered. This scheme adapts to multiple
servers going down, ensuring that claiming lock majority is always possible.
However, it can lead to multiple monitors claiming primary status in a
split-brain situation. As an example, consider a cluster with servers 1 to 4
with MaxScales A and B, as in the image below. MaxScale A can connect to
servers 1 and 2 (and claim their locks) but not to servers 3 and 4 due to
a network split. MaxScale A thus assumes servers 3 and 4 are down. MaxScale B
does the opposite, claiming servers 3 and 4 and assuming 1 and 2 are down.
Both MaxScales claim two locks out of two available and assume that they have
lock majority. Both MaxScales may then promote their own primaries and route
writes to different servers.
The recommended strategy depends on which failure scenario is more likely and/or more destructive. If it's unlikely that multiple servers are ever down simultaneously, then majority_of_all is likely the safer choice. On the other hand, if split-brain is unlikely but multiple servers may be down simultaneously, then majority_of_running would keep the cluster operational.
To check if a monitor is primary, fetch monitor diagnostics with maxctrl show
monitors
or the REST API. The boolean field primary indicates whether the
monitor has lock majority on the cluster. If cooperative monitoring is disabled,
the field value is null. Lock information for individual servers is listed in
the server-specific field lock_held. Again, null indicates that locks are
not in use or the lock status is unknown.
If a MaxScale instance tries to acquire the locks but fails to get majority (perhaps another MaxScale was acquiring locks simultaneously) it will release any acquired locks and try again after a random number of monitor ticks. This prevents multiple MaxScales from fighting over the locks continuously as one MaxScale will eventually wait less time than the others. Conflict probability can be further decreased by configuring each monitor with a different monitor_interval.
The flowchart below illustrates the lock handling logic.
Releasing locks
Monitor cooperation depends on the server locks. The locks are connection-specific. The owning connection can manually release a lock, allowing another connection to claim it. Also, if the owning connection closes, the MariaDB Server process releases the lock. How quickly a lost connection is detected affects how quickly the primary monitor status moves from one monitor and MaxScale to another.
If the primary MaxScale or its monitor is stopped normally, the monitor connections are properly closed, releasing the locks. This allows the secondary MaxScale to quickly claim the locks. However, if the primary simply vanishes (broken network), the connection may just look idle. In this case, the MariaDB Server may take a long time before it considers the monitor connection lost. This time ultimately depends on TCP keepalive settings on the machines running MariaDB Server.
On MariaDB Server 10.3.3 and later, the TCP keepalive settings can be configured for just the server process. See Server System Variables for information on settings tcp_keepalive_interval, tcp_keepalive_probes and tcp_keepalive_time. These settings can also be set on the operating system level, as described here.
As of MaxScale 6.4.16, 22.08.13, 23.02.10, 23.08.6 and 24.02.2, configuring TCP keepalive is no longer necessary as monitor sets the session wait_timeout variable when acquiring a lock. This causes the MariaDB Server to close the monitor connection if the connection appears idle for too long. The value of wait_timeout used depends on the monitor interval and connection timeout settings, and is logged at MaxScale startup.
A monitor can also be ordered to manually release its locks via the module command release-locks. This is useful for manually changing the primary monitor. After running the release-command, the monitor will not attempt to reacquire the locks for one minute, even if it wasn't the primary monitor to begin with. This command can cause the cluster to become temporarily unusable by MaxScale. Only use it when there is another monitor ready to claim the locks.
maxctrl call command mariadbmon release-locks MyMonitor1
Troubleshooting
Failover/switchover fails
Before performing failover or switchover, the MariaDB Monitor first checks that
prerequisites are fulfilled, printing any found errors. This should catch and
explain most issues with failover or switchover not working. If the operations
are attempted and still fail, then most likely one of the commands the monitor
issued to a server failed or timed out. The log should explain which query failed.
To print out all queries sent to the servers, start MaxScale with
--debug=enable-statement-logging
. This setting prints all queries sent to the
backends by monitors and authenticators. The printed queries may include
usernames and passwords.
A typical reason for failure is that a command such as STOP SLAVE
takes longer than the
backend_read_timeout
of the monitor, causing the connection to break. As of 2.3, the
monitor will retry most such queries if the failure was caused by a timeout. The retrying
continues until the total time for a failover or switchover has been spent. If the log
shows warnings or errors about commands timing out, increasing the backend timeout
settings of the monitor should help. Another settings to look at are query_retries
and
query_retry_timeout
. These are general MaxScale settings described in the
Configuration guide. Setting
query_retries
to 2 is a reasonable first try.
Slave detection shows external masters
If a slave is shown in maxctrl as "Slave of External Server" instead of
"Slave", the reason is likely that the "Master_Host"-setting of the replication connection
does not match the MaxScale server definition. As of 2.3.2, the MariaDB Monitor by default
assumes that the slave connections (as shown by SHOW ALL SLAVES STATUS
) use the exact
same "Master_Host" as used the MaxScale configuration file server definitions. This is
controlled by the setting assume_unique_hostnames.
Using the MariaDB Monitor With Binlogrouter
Since MaxScale 2.2 it's possible to detect a replication setup which includes Binlog Server: the required action is to add the binlog server to the list of servers only if master_id identity is set.