Readwritesplit

Readwritesplit

This document provides a short overview of the readwritesplit router module and its intended use case scenarios. It also displays all router configuration parameters with their descriptions. A list of current limitations of the module is included and use examples are provided.

Overview

The readwritesplit router is designed to increase the read-only processing capability of a cluster while maintaining consistency. This is achieved by splitting the query load into read and write queries. Read queries, which do not modify data, are spread across multiple nodes while all write queries will be sent to a single node.

The router is designed to be used with a traditional Primary-Replica replication cluster. It automatically detects changes in the primary server and will use the current primary server of the cluster. With a Galera cluster, one can achieve a resilient setup and easy primary failover by using one of the Galera nodes as a Write-Primary node, where all write queries are routed, and spreading the read load over all the nodes.

Interaction with servers in Maintenance and Draining state

When a server that readwritesplit uses is put into maintenance mode, any ongoing requests are allowed to finish before the connection is closed. If the server that is put into maintenance mode is a primary, open transaction are allowed to complete before the connection is closed. Note that this means neither idle session nor long-running transactions will be closed by readwritesplit. To forcefully close the connections, use the following command:

maxctrl set server <server> maintenance --force

If a server is put into the Draining state while a connection is open, the connection will be used normally. Whenever a new connection needs to be created, whether that be due to a network error or when a new session being opened, only servers that are neither Draining nor Drained will be used.

Configuration

Readwritesplit router-specific settings are specified in the configuration file of MariaDB MaxScale in its specific section. The section can be freely named but the name is used later as a reference in a listener section.

For more details about the standard service parameters, refer to the Configuration Guide.

Starting with 2.3, all router parameters can be configured at runtime. Use maxctrl alter service to modify them. The changed configuration will only be taken into use by new sessions.

Parameters

max_slave_connections

  • Type: integer
  • Mandatory: No
  • Dynamic: Yes
  • Default: 255

max_slave_connections sets the maximum number of replicas a router session uses at any moment. The default is to use at most 255 replica connections per client connection. In older versions the default was to use all available replicas with no limit.

For MaxScale 2.5.12 and newer, the minimum value is 0.

For MaxScale versions 2.5.11 and older, the minimum value is 1. These versions suffer from a bug (MXS-3536) that causes the parameter to accept any values but only function when a value greater than one was given.

Starting with MaxScale 2.5.0, the use of percentage values in max_slave_connections is deprecated. The support for percentages will be removed in a future release.

For example, if you have configured MaxScale with one primary and three replicas and set max_slave_connections=2, for each client connection a connection to the primary and two replica connections would be opened. The read query load balancing is then done between these two replicas and writes are sent to the primary.

By tuning this parameter, you can control how dynamic the load balancing is at the cost of extra created connections. With a lower value of max_slave_connections, less connections per session are created and the set of possible replica servers is smaller. With a higher value in max_slave_connections, more connections are created which requires more resources but load balancing will almost always give the best single query response time and performance. Longer sessions are less affected by a high max_slave_connections as the relative cost of opening a connection is lower.

Behavior of max_slave_connections=0

When readwritesplit is configured with max_slave_connections=0, readwritesplit will behave slightly differently in that it will route all reads to the current master server. This is a convenient way to force all of the traffic to go to a single node while still being able to leverage the replay and reconnection features of readwritesplit.

In this mode, the behavior of master_failure_mode=fail_on_write also changes slightly. If the current Master server fails and a read is done when there's no other Master server available, the connection will be closed. This is done to prevent an extra slave connection from being opened that would not be closed if a new Master server would arrive.

slave_connections

  • Type: integer
  • Mandatory: No
  • Dynamic: Yes
  • Default: 255

This parameter controls how many replica connections each new session starts with. The default value is 255 which is the same as the default value of max_slave_connections.

In contrast to max_slave_connections, slave_connections serves as a soft limit on how many replica connections are created. The number of replica connections can exceed slave_connections if the load balancing algorithm finds an unconnected replica server better than all other replicas.

Setting this parameter to 1 allows faster connection creation and improved resource usage due to the smaller amount of initial backend connections. It is recommended to use slave_connections=1 when the lifetime of the client connections is short.

max_slave_replication_lag

  • Type: duration
  • Mandatory: No
  • Dynamic: Yes
  • Default: 0s

Specify how many seconds a replica is allowed to be behind the primary. The lag of a replica must be less than the configured value in order for it to be used for routing. If set to 0 (the default value), the feature is disabled.

In MaxScale 2.5.0, the replica lag must be less than max_slave_replication_lag whereas in older versions the replica lag had to be less than or equal to max_slave_replication_lag. This means that in MaxScale 2.5.0 it is possible to define, with max_slave_replication_lag=1, that all replicas must be up to date in order for them to be used for routing.

Note that this feature does not guarantee that writes done on the primary are visible for reads done on the replica. This is mainly due to the method of replication lag measurement. For a feature that guarantees this, refer to causal_reads.

The lag 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 lag is seconds, a lag specified in milliseconds will be rejected, even if the duration is longer than a second.

The Readwritesplit-router does not detect the replication lag itself. A monitor such as the MariaDB-monitor for a Primary-Replica cluster is required. This option only affects Primary-Replica clusters. Galera clusters do not have a concept of replica lag even if the application of write sets might have lag. When a server is disqualified from routing because of replication lag, a warning is logged. Similarly, when the server has caught up enough to be a valid routing target, another warning is logged. These messages are only logged when a query is being routed and the replication state changes.

use_sql_variables_in

  • Type: enum
  • Mandatory: No
  • Dynamic: Yes
  • Values: master, all
  • Default: all

This parameter controls how SELECT statements that use SQL user variables are handled. Here is an example of such a query that uses it to return an increasing row number for a resultset:

SET @rownum := 0;
SELECT @rownum := @rownum + 1 AS rownum, user, host FROM mysql.user;

By default MaxScale will route both the SET and SELECT statements to all nodes. Any future reads of the user variables can also be performed on any node.

The possible values for this parameter are:

  • all (default)

  • Modifications to user variables inside SELECT statements as well as reads of user variables are routed to all servers.

    Versions before MaxScale 22.08 returned an error if a user variable was modified inside of a SELECT statement when use_sql_variables_in=all was used. MaxScale 22.08 will instead route the query to all servers and discard the extra results.

  • master

  • Modifications to user variables inside SELECT statements as well as reads of user variables are routed to the primary server. This forces more of the traffic onto the primary server but it reduces the amount of data that is discarded for any SELECT statement that also modifies a user variable. With this mode, the state of user variables is not deterministic if they are modified inside of a SELECT statement. SET statements that modify user variabels are still routed to all servers.

DML statements, such as INSERT, UPDATE or DELETE, that modify SQL user variables are still treated as writes and are only routed to the primary server. For example, after the following query the value of @myid is no longer the same on all servers and the SELECT statement can return different values depending where it ends up being executed:

SET @myid := 0;
INSERT INTO test.t1 VALUES (@myid := @myid + 1);
SELECT @myid; -- Might return 1 or 0

connection_keepalive

Note: This parameter has been moved into the MaxScale core. For the current documentation, read the connection_keepalive section in the configuration guide.

Send keepalive pings to backend servers. This feature was introduced in MaxScale 2.2.0. The default value is 300 seconds starting with 2.3.2 and for older versions the feature was disabled by default. This parameter was converted into a service parameter in MaxScale 2.5.0.

master_reconnection

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Allow the primary server to change mid-session. This feature was introduced in MaxScale 2.3.0 and is disabled by default. This feature requires that disable_sescmd_history is not used.

When a readwritesplit session starts, it will pick a primary server as the current primary server of that session. By default, when this primary server is lost or changes to another server, the connection will be closed.

When master_reconnection is enabled, readwritesplit can sometimes recover a lost connection to the primary server. This largely depends on the value of master_failure_mode.

With master_failure_mode=fail_instantly, the primary server is only allowed to change to another server. This change must happen without a loss of the primary server.

With master_failure_mode=fail_on_write, the loss of the primary server is no longer a fatal error: if a replacement primary server appears before any write queries are received, readwritesplit will transparently reconnect to the new primary server.

In both cases the change in the primary server can only take place if prune_sescmd_history is enabled or max_sescmd_history has not yet been exceeded and the session does not have an open transaction.

The recommended configuration is to use master_reconnection=true and master_failure_mode=fail_on_write. This provides improved fault tolerance without any risk to the consistency of the database.

slave_selection_criteria

  • Type: enum
  • Mandatory: No
  • Dynamic: Yes
  • Values: LEAST_CURRENT_OPERATIONS, ADAPTIVE_ROUTING, LEAST_BEHIND_MASTER, LEAST_ROUTER_CONNECTIONS, LEAST_GLOBAL_CONNECTIONS
  • Default: LEAST_CURRENT_OPERATIONS

This option controls how the readwritesplit router chooses the replicas it connects to and how the load balancing is done. The default behavior is to route read queries to the replica server with the lowest amount of ongoing queries i.e. LEAST_CURRENT_OPERATIONS.

The option syntax:

slave_selection_criteria=<criteria>

Where <criteria> is one of the following values.

  • LEAST_CURRENT_OPERATIONS (default), the replica with least active operations
  • ADAPTIVE_ROUTING, based on server average response times.
  • LEAST_BEHIND_MASTER, the replica with smallest replication lag
  • LEAST_GLOBAL_CONNECTIONS, the replica with least connections from MariaDB MaxScale
  • LEAST_ROUTER_CONNECTIONS, the replica with least connections from this service

LEAST_CURRENT_OPERATIONS uses the current number of active operations (i.e. SQL queries) as the load balancing metric and it optimizes for maximal query throughput. Each query gets routed to the server with the least active operations which results in faster servers processing more traffic.

ADAPTIVE_ROUTING uses the server response time and current estimated server load as the load balancing metric. The server that is estimated to finish an additional query first is chosen. A modified average response time for each server is continuously updated to allow slow servers at least some traffic and quickly react to changes in server load conditions. This selection criteria is designed for heterogeneous clusters: servers of differing hardware, differing network distances, or when other loads are running on the servers (including a backup). If the servers are queried by other clients than MaxScale, the load caused by them is indirectly taken into account.

LEAST_BEHIND_MASTER uses the measured replication lag as the load balancing metric. This means that servers that are more up-to-date are favored which increases the likelihood of the data being read being up-to-date. However, this is not as effective as causal_reads would be as there's no guarantee that writes done by the same connection will be routed to a server that has replicated those changes. The recommended approach is to use LEAST_CURRENT_OPERATIONS or ADAPTIVE_ROUTING in combination with causal_reads

NOTE: LEAST_GLOBAL_CONNECTIONS and LEAST_ROUTER_CONNECTIONS should not be used, they are legacy options that exist only for backwards compatibility. Using them will result in skewed load balancing as the algorithm uses a metric that's too coarse (number of connections) to load balance something that's finer (individual SQL queries).

The LEAST_GLOBAL_CONNECTIONS and LEAST_ROUTER_CONNECTIONS use the connections from MariaDB MaxScale to the server, not the amount of connections reported by the server itself.

Starting with MaxScale versions 2.5.29, 6.4.11, 22.08.9, 23.02.5 and 23.08.1, lowercase versions of the values are also accepted. For example, slave_selection_criteria=LEAST_CURRENT_OPERATIONS and slave_selection_criteria=least_current_operations are both accepted as valid values.

max_sescmd_history

This parameter has been moved to the MaxScale core in MaxScale 6.0.

disable_sescmd_history

This parameter has been moved to the MaxScale core in MaxScale 6.0.

prune_sescmd_history

This parameter has been moved to the MaxScale core in MaxScale 6.0.

master_accept_reads

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

master_accept_reads allows the primary server to be used for reads. This is a useful option to enable if you are using a small number of servers and wish to use the primary for reads as well.

By default, no reads are sent to the primary as long as there is a valid replica server available. If no replicas are available, reads are sent to the primary regardless of the value of master_accept_reads.

# Use the primary for reads
master_accept_reads=true

strict_multi_stmt

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

This option is disabled by default since MaxScale 2.2.1. In older versions, this option was enabled by default.

When a client executes a multi-statement query, it will be treated as if it were a DML statement and routed to the primary. If the option is enabled, all queries after a multi-statement query will be routed to the primary to guarantee a consistent session state.

If the feature is disabled, queries are routed normally after a multi-statement query.

Warning: Enable the strict mode only if you know that the clients will send statements that cause inconsistencies in the session state.

# Enable strict multi-statement mode
strict_multi_stmt=true

strict_sp_calls

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Similar to strict_multi_stmt, this option allows all queries after a CALL operation on a stored procedure to be routed to the primary. This option is disabled by default and was added in MaxScale 2.1.9.

All warnings and restrictions that apply to strict_multi_stmt also apply to strict_sp_calls.

master_failure_mode

  • Type: enum
  • Mandatory: No
  • Dynamic: Yes
  • Values: fail_instantly, fail_on_write, error_on_write
  • Default: fail_instantly

This option controls how the failure of a primary server is handled. By default, the router will close the client connection as soon as the primary is lost.

The following table describes the values for this option and how they treat the loss of a primary server.

Value Description
fail_instantly When the failure of the primary server is detected, the connection will be closed immediately.
fail_on_write The client connection is closed if a write query is received when no primary is available.
error_on_write If no primary is available and a write query is received, an error is returned stating that the connection is in read-only mode.

These also apply to new sessions created after the primary has failed. This means that in fail_on_write or error_on_write mode, connections are accepted as long as replica servers are available.

When configured with fail_on_write or error_on_write, sessions that are idle will not be closed even if all backend connections for that session have failed. This is done in the hopes that before the next query from the idle session arrives, a reconnection to one of the replicas is made. However, this can leave idle connections around unless the client application actively closes them. To prevent this, use the connection_timeout parameter.

Note: If master_failure_mode is set to error_on_write and the connection to the primary is lost, by default, clients will not be able to execute write queries without reconnecting to MariaDB MaxScale once a new primary is available. If master_reconnection is enabled, the session can recover if one of the replicas is promoted as the primary.

retry_failed_reads

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: true

This option controls whether autocommit selects are retried in case of failure. This option is enabled by default.

When a simple autocommit select is being executed outside of a transaction and the replica server where the query is being executed fails, readwritesplit can retry the read on a replacement server. This makes the failure of a replica transparent to the client.

delayed_retry

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Retry queries over a period of time. This parameter takes a boolean value, was added in Maxscale 2.3.0 and is disabled by default.

When this feature is enabled, a failure to route a query due to a connection problem will not immediately result in an error. The routing of the query is delayed until either a valid candidate server is available or the retry timeout is reached. If a candidate server becomes available before the timeout is reached, the query is routed normally and no connection error is returned. If no candidates are found and the timeout is exceeded, the router returns to normal behavior and returns an error.

When combined with the master_reconnection parameter, failures of writes done outside of transactions can be hidden from the client connection. This allows a primary to be replaced while a write is in progress.

The delayed query retrying mode in readwritesplit does not do any sort of duplicate write detection. To prevent accidental data duplication, it is highly recommended to tune the monitor timeouts to values that produce accurate results.

Duplicate execution of a statement can occur if the connection to the server is lost or the server crashes but the server comes back up before the timeout for the retrying is exceeded. At this point, if the server managed to read the client's statement, it will be executed. For this reason, it is recommended to only enable delayed_retry when the possibility of duplicate statement execution is an acceptable risk.

delayed_retry_timeout

  • Type: duration
  • Mandatory: No
  • Dynamic: Yes
  • Default: 10s

The duration to wait until an error is returned to the client when delayed_retry is enabled. The default value is 10 seconds.

The 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.

transaction_replay

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Replay interrupted transactions. This parameter was added in MaxScale 2.3.0 and is disabled by default. Enabling this parameter enables both delayed_retry and master_reconnection and sets master_failure_mode to fail_on_write, thereby overriding any configured values for these parameters.

When the server where the transaction is in progress fails, readwritesplit can migrate the transaction to a replacement server. This can completely hide the failure of a primary node without any visible effects to the client.

If no replacement node becomes available, the client connection is closed.

To control how long a transaction replay can take, use transaction_replay_timeout.

Please refer to the Transaction Replay Limitations section for a more detailed explanation of what should and should not be done with transaction replay.

transaction_replay_max_size

  • Type: size
  • Mandatory: No
  • Dynamic: Yes
  • Default: 1 MiB

The limit on transaction size for transaction replay in bytes. Any transaction that exceeds this limit will not be replayed. The default value is 1 MiB. This limit applies at a session level which means that the total peak memory consumption can be transaction_replay_max_size times the number of client connections.

The amount of memory needed to store a particular transaction will be slightly larger than the length in bytes of the SQL used in the transaction. If the limit is ever exceeded, a message will be logged at the info level.

Starting with MaxScale 6.4.10, the number of times that this limit has been exceeded is shown in maxctrl show service as trx_max_size_exceeded.

Read the configuration guide for more details on size type parameters in MaxScale.

transaction_replay_attempts

  • Type: integer
  • Mandatory: No
  • Dynamic: Yes
  • Default: 5

The upper limit on how many times a transaction replay is attempted before giving up. The default value is 5.

A transaction replay failure can happen if the server where the transaction is being replayed fails while the replay is in progress. In practice this parameter controls how many server and network failures a single transaction replay tolerates. If a transaction is replayed successfully, the counter for failed attempts is reset.

transaction_replay_timeout

  • Type: duration
  • Mandatory: No
  • Dynamic: Yes
  • Default: 0s

The time how long transactions are attempted for. This feature is disabled by default and was added in MaxScale 6.2.1. To explicitly disable this feature, set the value to 0 seconds.

The timeout is a duration type and the value must include a unit for the duration.

When transaction_replay_timeout is enabled, the time a transaction replay can take is controlled solely by this parameter. This is a more convenient and predictable method of controlling how long a transaction replay can be attempted before the connection is closed.

If delayed_retry_timeout is less than transaction_replay_timeout, it is set to the same value.

By default the time how long a transaction can be retried is controlled by delayed_retry_timeout and transaction_replay_attempts. This can result in a maximum replay time limit of delayed_retry_timeout multiplied by transaction_replay_attempts, by default this is 50 seconds. The minimum replay time limit can be as low as transaction_replay_attempts seconds (5 seconds by default) in cases where the connection fails after it was created. Usually this happens due to problems like the max_connections limit being hit on the database server.

With the introduction of transaction_replay_timeout, these problems are avoided. Starting with MaxScale 6.2.1, this is the recommended method of controlling the timeouts for transaction replay.

transaction_replay_retry_on_deadlock

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Enable automatic retrying of transactions that end up in a deadlock. This parameter was added in MaxScale 2.4.6 and the feature is disabled by default. MaxScale versions from 2.4.0 to 2.4.5 always tried to replay deadlocked transactions.

If this feature is enabled and a transaction returns a deadlock error (e.g. SQLSTATE 40001: Deadlock found when trying to get lock; try restarting transaction), the transaction is automatically retried. If the retrying of the transaction results in another deadlock error, it is retried until it either succeeds or a transaction checksum error is encountered.

transaction_replay_retry_on_mismatch

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Retry transactions that end in checksum mismatch. This parameter was added in MaxScale 6.2.1 is disabled by default.

When enabled, any replayed transactions that end with a checksum mismatch are retried until they either succeeds or one of the transaction replay limits is reached (delayed_retry_timeout, transaction_replay_timeout or transaction_replay_attempts).

transaction_replay_checksum

  • Type: enum
  • Mandatory: No
  • Dynamic: Yes
  • Values: full, result_only, no_insert_id
  • Default: full

Selects which transaction checksum method is used to verify the result of the replayed transaction.

Note that only transaction_replay_checksum=full is guaranteed to retain the consistency of the replayed transaction.

Possible values are:

  • full (default)

  • All responses from the server are included in the checksum. This retains the full consistency guarantee of the replayed transaction as it must match exactly the one that was already returned to the client.

  • result_only

  • Only resultsets and errors are included in the checksum. OK packets (i.e. successful queries that do not return results) are ignored. This mode is intended to be used in cases where the extra information (auto-generated ID, warnings etc.) returned in the OK packet is not used by the application.

    This mode is safe to use only if the auto-generated ID is not actually used by any following queries. An example of such behavior would be a transaction that ends with an INSERT into a table with an AUTO_INCREMENT field.

  • no_insert_id

  • The same as result_only but results from queries that use LAST_INSERT_ID() are also ignored. This mode is safe to use only if the result of the query is not used by any subsequent statement in the transaction.

optimistic_trx

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Enable optimistic transaction execution. This parameter controls whether normal transactions (i.e. START TRANSACTION or BEGIN) are load balanced across replicas. This feature is disabled by default and enabling it implicitly enables transaction_replay, delayed_retry and master_reconnection parameters.

When this mode is enabled, all transactions are first attempted on replica servers. If the transaction contains no statements that modify data, it is completed on the replica. If the transaction contains statements that modify data, it is rolled back on the replica server and restarted on the primary. The rollback is initiated the moment a data modifying statement is intercepted by readwritesplit so only read-only statements are executed on replica servers.

As with transaction_replay and transactions that are replayed, if the results returned by the primary server are not identical to the ones returned by the replica up to the point where the first data modifying statement was executed, the connection is closed. If the execution of ROLLBACK statement on the replica fails, the connection to that replica is closed.

All limitations that apply to transaction_replay also apply to optimistic_trx.

causal_reads

  • Type: enum
  • Mandatory: No
  • Dynamic: Yes
  • Values: none, local, global, fast, fast_global, universal
  • Default: none

Enable causal reads. This parameter is disabled by default and was introduced in MaxScale 2.3.0.

If a client connection modifies the database and causal_reads is enabled, any subsequent reads performed on replica servers will be done in a manner that prevents replication lag from affecting the results.

The following table contains a comparison of the modes. Read the implementation of causal_reads for more information on what a sync consists of and why minimizing the number of them is important.

Mode Level of Causality Latency
local Session Low, one sync per write.
fast Session None, no sync at all.
global Service Medium, one sync per read.
fast_global Service None, no sync at all.
universal Cluster High, one sync per read plus a roundtrip to the primary.

The fast and fast_global modes should only be used when low latency is more important than proper distribution of reads. These modes should only be used when the workload is mostly read-only with only occasional writes. If used with a mixed or a write-heavy workload, the traffic will end up being routed almost exclusively to the primary server.

Note: This feature requires MariaDB 10.2.16 or newer to function. In addition to this, the session_track_system_variables parameter must include last_gtid in its list of tracked system variables.

The possible values for this parameter are:

  • none (default)

  • Read causality is disabled.

  • local

  • Writes are locally visible. Writes are guaranteed to be visible only to the connection that does it. Unrelated modifications done by other connections are not visible. This mode improves read scalability at the cost of latency and reduces the overall load placed on the primary server without breaking causality guarantees.

  • global

  • Writes are globally visible. If one connection writes a value, all connections to the same service will see it. In general this mode is slower than the local mode due to the extra synchronization it has to do. This guarantees global happens-before ordering of reads when all transactions are inside a single GTID domain.This mode gives similar benefits as the local mode in that it improves read scalability at the cost of latency.

    With MaxScale versions 2.5.14 and older, multi-domain use of causal_reads could cause non-causal reads to occur. Starting with MaxScale 2.5.15, this was fixed and all the GTID coordinates are passed alongside all requests which makes multi-domain GTIDs safe to use. However, this does mean that the GTID coordinates will never be reset: if replication is reset and and GTID coordinates go "backwards", readwritesplit will not consider these as being newer than the ones already stored. To reset the stored GTID coordinates in readwritesplit, MaxScale must be restarted.

    MaxScale 6.4.11 added the new reset-gtid module command to readwritesplit. This allows the global GTID state used by causal_reads=global to be reset without having to restart MaxScale.

  • fast

  • This mode is similar to the local mode where it will only affect the connection that does the write but where the local mode waits for a replica server to catch up, the fast mode will only use servers that are known to have replicated the write. This means that if no replica has replicated the write, the primary where the write was done will be used. The value of causal_reads_timeout is ignored in this mode. Currently the replication state is only updated by the mariadbmon monitor whenever the servers are monitored. This means that a smaller monitor_interval provides faster replication state updates and possibly better overall usage of servers.

    This mode is the inverse of the local mode in the sense that it improves read latency at the cost of read scalability while still retaining the causality guarantees for reads. This functionality can also be considered an improved version of the functionality that the CCRFilter module provides.

  • fast_global

  • This mode is identical to the fast mode except that it uses the global GTID instead of the session local one. This is similar to how local and global modes differ from each other. The value of causal_reads_timeout is ignored in this mode. Currently the replication state is only updated by the mariadbmon monitor whenever the servers are monitored. This means that a smaller monitor_interval provides faster replication state updates and possibly better overall usage of servers.

  • universal

  • The universal mode guarantees that all SELECT statements always see the latest observable transaction state on a database cluster. The basis of this is the @@gtid_current_pos variable which is read from the current primary server before each read. This guarantees that if a transaction was visible at the time the read is received by readwritesplit, the transaction is guaranteed to be complete on the replica server where the read is done.

    This mode is the most consistent of all the modes. It provides consistency regardless of where a write originated from but it comes at the cost of increased latency. For every read, a round trip to the current primary server is done. This means that the latency of any given SELECT statement increases by roughly twice the network latency between MaxScale and the database cluster. In addition, an extra SELECT statement is always executed on the primary which places some load on the server.

Before MaxScale 2.5.0, the causal_reads parameter was a boolean parameter. False values translated to none and true values translated to local. The use of boolean parameters is deprecated but still accepted in MaxScale 2.5.0.

Implementation of causal_reads

This feature is based on the MASTER_GTID_WAIT function and the tracking of server-side status variables. By tracking the latest GTID that each statement generates, readwritesplit can then perform a synchronization operation with the help of the MASTER_GTID_WAIT function.

If the replica has not caught up to the primary within the configured time, as specified by causal_reads_timeout, it will be retried on the primary. In MaxScale 2.3.0 an error was returned to the client when the replica timed out.

The exception to this rule is the fast mode which does not do any synchronization at all. This can be done as any reads that would go to out-of-date servers will be re-routed to the current primary.

Normal SQL

A practical example can be given by the following set of SQL commands executed with autocommit=1.

INSERT INTO test.t1 (id) VALUES (1);
SELECT * FROM test.t1 WHERE id = 1;

As the statements are not executed inside a transaction, from the load balancer's point of view, the latter statement can be routed to a replica server. The problem with this is that if the value that was inserted on the primary has not yet replicated to the server where the SELECT statement is being performed, it can appear as if the value we just inserted is not there.

By prefixing these types of SELECT statements with a command that guarantees consistent results for the reads, read scalability can be improved without sacrificing consistency.

The set of example SQL above will be translated by MaxScale into the following statements.

INSERT INTO test.t1 (id) VALUES (1);
SET @maxscale_secret_variable=(
    SELECT CASE
           WHEN MASTER_GTID_WAIT('0-3000-8', 10) = 0 THEN 1
           ELSE (SELECT 1 FROM INFORMATION_SCHEMA.ENGINES)
    END);
SELECT * FROM test.t1 WHERE id = 1;

The SET command will synchronize the replica to a certain logical point in the replication stream (see MASTER_GTID_WAIT for more details).

Prepared Statements

Binary protocol prepared statements are handled in a different manner. Instead of adding the synchronization SQL into the original SQL query, it is sent as a separate packet before the prepared statement is executed.

We'll use the same example SQL but use a binary protocol prepared statement for the SELECT:

COM_QUERY:         INSERT INTO test.t1 (id) VALUES (1);
COM_STMT_PREPARE:  SELECT * FROM test.t1 WHERE id = ?;
COM_STMT_EXECUTE:  ? = 123

The SQL that MaxScale executes will be the following:

COM_QUERY:         INSERT INTO test.t1 (id) VALUES (1);
COM_STMT_PREPARE:  SELECT * FROM test.t1 WHERE id = ?;
COM_QUERY:         IF (MASTER_GTID_WAIT('0-3000-8', 10) <> 0) THEN KILL (SELECT CONNECTION_ID()); END IF
COM_STMT_EXECUTE:  ? = 123

Both the synchronization query and the execution of the prepared statement are sent at the same time. This is done to remove the need to wait for the result of the synchronization query before routing the execution of the prepared statement. This keeps the performance of causal_reads for prepared statements the same as it is for normal SQL queries.

As a result of this, each time the the synchronization query times out, the connection will be killed by the KILL statement and readwritesplit will retry the query on the primary. This is done to prevent the execution of the prepared statement that follows the synchronization query from being processed by the MariaDB server.

It is recommend that the session command history is enabled whenever prepared statements are used with causal_reads. This allows new connections to be created whenever a causal read times out.

Starting with MaxScale 2.5.17, a failed causal read inside of a read-only transaction started with START TRANSACTION READ ONLY will return the following error:

Error:    1792
SQLSTATE: 25006
Message:  Causal read timed out while in a read-only transaction, cannot retry command.

Older versions of MaxScale attempted to retry the command on the current primary server which would cause the connection to be closed and a warning to be logged.

Limitations of Causal Reads

  • This feature does not work with Galera or any other non-standard replication mechanisms. As Galera does not update the gtid_slave_pos variable when events are replicated via the Galera library, the MASTER_GTID_WAIT function used by MaxScale to synchronize reads will wait until the timeout. With Galera this is not a serious issue as it, by nature, is a mostly-synchronous replication mechanism.

  • If the combination of the original SQL statement and the modifications added to it by readwritesplit exceed the maximum packet size (16777213 bytes), the causal read will not be attempted and a non-causal read is done instead. This applies only to text protocol queries as the binary protocol queries use a different synchronization mechanism.

causal_reads_timeout

  • Type: duration
  • Mandatory: No
  • Dynamic: Yes
  • Default: 10s

The timeout for the replica synchronization done by causal_reads. The default value is 10 seconds.

The 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.

lazy_connect

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Lazy connection creation causes connections to backend servers to be opened only when they are needed. This reduces the load that is placed on the backend servers when the client connections are short. This parameter is a boolean type and is disabled by default.

By default readwritesplit opens as many connections as it can when the session is first opened. This makes the execution of the first query faster when all available connections are already created. When lazy_connect is enabled, this initial connection creation is skipped. If the client executes only read queries, no connection to the primary is made. If only write queries are made, only the primary connection is used.

reuse_prepared_statements

  • Type: boolean
  • Mandatory: No
  • Dynamic: Yes
  • Default: false

Reuse identical prepared statements inside the same client connection. This is a boolean parameter and is disabled by default. This feature only applies to binary protocol prepared statements.

When this parameter is enabled and the connection prepares an identical prepared statement multiple times, instead of preparing it on the server the existing prepared statement handle is reused. This also means that whenever prepared statements are closed by the client, they will be left open by readwritesplit.

Enabling this feature will increase memory usage of a session. The amount of memory stored per prepared statement is proportional to the length of the prepared SQL statement and the number of parameters the statement has.

Router Diagnostics

The router_diagnostics output for a readwritesplit service contains the following fields.

  • queries: Number of queries executed through this service.
  • route_master: Number of writes routed to primary.
  • route_slave: Number of reads routed to replicas.
  • route_all: Number of session commands routed to all servers.
  • rw_transactions: Number of explicit read-write transactions.
  • ro_transactions: Number of explicit read-only transactions.
  • replayed_transactions: Number of replayed transactions.

  • server_query_statistics: Statistics for each configured and used server consisting of the following fields.

  • id: Name of the server
  • total: Total number of queries.
  • read: Total number of reads.
  • write: Total number of writes.
  • avg_sess_duration: Average duration of a client session to this server.
  • avg_sess_active_pct: Average percentage of time client sessions were active. 0% means connections were opened but never used.
  • avg_selects_per_session: Average number of selects per session.

Server Ranks

The general rule with server ranks is that primary servers will be used before secondary servers. Readwritesplit is an exception to this rule. The following rules govern how readwritesplit behaves with servers that have different ranks.

  • Sessions will use the current primary server as long as possible. This means that sessions with a secondary primary will not use the primary primary as long as the secondary primary is available.

  • All replica connections will use the same rank as the primary connection. Any stale connections with a different rank than the primary will be discarded.

  • If no primary connection is available and master_reconnection is enabled, a connection to the best primary is created. If the new primary has a different priority than existing connections have, the connections with a different rank will be discarded.

  • If open connections exist, these will be used for routing. This means that if the primary is lost but the session still has replica servers with the same rank, they will remain in use.

  • If no open connections exist, the servers with the best rank will used.

Routing hints

The readwritesplit router supports routing hints. For a detailed guide on hint syntax and functionality, please read this document.

Note: Routing hints will always have the highest priority when a routing decision is made. This means that it is possible to cause inconsistencies in the session state and the actual data in the database by adding routing hints to DDL/DML statements which are then directed to replica servers. Only use routing hints when you are sure that they can cause no harm.

An exception to this rule is transaction_replay: when it is enabled, all routing hints inside transaction are ignored. This is done to prevent changes done inside a re-playable transaction from affecting servers outside of the transaction. This behavior was added in MaxScale 6.1.4. Older versions allowed routing hints to override the transaction logic.

Known Limitations of Routing Hints

  • If a SELECT statement with a maxscale route to slave hint is received while autocommit is disabled, the query will be routed to a replica server. This causes some metadata locks to be acquired on the database in question which will block DDL statements on the server until either the connection is closed or autocommit is enabled again.

Module Commands

The readwritesplit router implements the following module commands.

reset-gtid

The command resets the global GTID state in the router. It can be used with causal_reads=global to reset the state. This can be useful when the cluster is reverted to an earlier state and the GTIDs recorded in MaxScale are no longer valid.

The first and only argument to the command is the router name. For example, to reset the GTID state of a readwritesplit named My-RW-Router, the following MaxCtrl command should be used:

maxctrl call command readwritesplit reset-gtid My-RW-Router

Examples

Examples of the readwritesplit router in use can be found in the Tutorials folder.

Readwritesplit routing decisions

Here is a small explanation which shows what kinds of queries are routed to which type of server.

Routing to Primary

Routing to primary is important for data consistency and because majority of writes are written to binlog and thus become replicated to replicas.

The following operations are routed to primary:

  • DML statements (INSERT, UPDATE, DELETE etc.)
  • DDL statements (DROP, CREATE, ALTER etc.)
  • All statements within an open read-write transaction
  • Stored procedure calls
  • User-defined function calls
  • Statements that use LAST_INSERT_ID()

In addition to these, if the readwritesplit service is configured with the max_slave_replication_lag parameter, and if all replicas suffer from too much replication lag, then statements will be routed to the primary. (There might be other similar configuration parameters in the future which limit the number of statements that will be routed to replicas.)

Transaction Isolation Level Tracking

Use of the SERIALIZABLE transaction isolation level with readwritesplit is not recommended as it somewhat goes against the goals of load balancing.

If either session_track_transaction_info=CHARACTERISTICS or session_track_system_variables=tx_isolation is configured for the MariaDB server, readwritesplit will track the transaction isolation level and lock the session to the primary when the isolation level is set to serializable. This retains the correctness of the isolation level which can otherwise cause problems. Once a session is locked to the primary, it will not be unlocked. To reinstate the normal routing behavior, a new connection must be created.

For example, if transaction isolation level tracking cannot be done and an autocommit SELECT is routed to a replica, it no longer behaves in a serializable manner. This can also have an effect on the replication in the replica server.

Routing to Replicas

The ability to route some statements to replicas is important because it also decreases the load targeted to primary. Moreover, it is possible to have multiple replicas to share the load in contrast to single primary.

Queries which can be routed to replicas must be auto committed and belong to one of the following group:

  • Read-only statements (i.e. SELECT) that only use read-only built-in functions
  • All statements within an explicit read-only transaction (START TRANSACTION READ ONLY)
  • SHOW statements except SHOW MASTER STATUS

The list of supported built-in fuctions can be found here.

Routing to every session backend

A third class of statements includes those which modify session data, such as session system variables, user-defined variables, the default database, etc. We call them session commands, and they must be replicated as they affect the future results of read and write operations. They must be executed on all servers that could execute statements on behalf of this client.

Session commands include for example:

  • Commands that modify the session state (SET, USE, CHANGE USER)
  • Text protocol PREPARE statements
  • Binary protocol prepared statements
  • Other miscellaneous commands (COM_QUIT, COM_PING etc.)

NOTE: if variable assignment is embedded in a write statement it is routed to primary only. For example, INSERT INTO t1 values(@myvar:=5, 7) would be routed to primary only.

The router stores all of the executed session commands so that in case of a replica failure, a replacement replica can be chosen and the session command history can be repeated on that new replica. This means that the router stores each executed session command for the duration of the session. Applications that use long-running sessions might cause MariaDB MaxScale to consume a growing amount of memory unless the sessions are closed. This can be solved by adjusting the value of max_sescmd_history.

Routing to previous target

In the following cases, a query is routed to the same server where the previous query was executed. If no previous target is found, the query is routed to the current primary.

  • If a query uses the FOUND_ROWS() function, it will be routed to the server where the last query was executed. This is done with the assumption that a query with SQL_CALC_FOUND_ROWS was previously executed.

  • COM_STMT_FETCH_ROWS will always be routed to the same server where the COM_STMT_EXECUTE was routed.

Limitations

Read queries are routed to the primary server in the following situations:

  • Query is executed inside an open read-write transaction
  • Statement includes a stored procedure or an UDF call
  • If there are multiple statements inside one query e.g. INSERT INTO ... ; SELECT LAST_INSERT_ID();

Prepares Statement Limitations

If a prepared statement targets a temporary table on the primary, the replica servers will fail to execute it. This will cause all replica connections to be closed (MXS-1816).

Transaction Replay Limitations

If the results from the replacement server are not identical when the transaction is replayed, the client connection is closed. This means that any transaction with a server specific result (e.g. NOW(), @@server_id) cannot be replayed successfully but it will still be attempted.

If a transaction reads data before updating it, the rows should be locked by using SELECT ... FOR UPDATE. This will prevent overlapping transactions when multiple transactions are being replayed that modify the same set of rows.

If the connection to the server where the transaction is being executed is lost when the final COMMIT is being executed, it is impossible to know whether the transaction was successfully committed. This means that there is a possibility for duplicate transaction execution which can result in data duplication in certain cases. Data duplication can happen if the transaction consists of the following statement types:

  • INSERT of rows into a table that does not have an auto-increment primary key
  • A "blind update" of one or more rows e.g. UPDATE t SET c = c + 1 WHERE id = 123
  • A "blind delete" e.g. DELETE FROM t LIMIT 100

This is not an exhaustive list and any operations that do not check the row contents before performing the operation on them might face this problem.

In all cases the problem of duplicate transaction execution can be avoided by including a SELECT ... FOR UPDATE in the statement. This will guarantee that in the case that the transaction fails when it is being committed, the row is only modified if it matches the expected contents.

Similarly, a connection loss during COMMIT can also result in transaction replay failure. This happens due to the same reason as duplicate transaction execution but the retried transaction will not be committed. This can be considered a success case as the transaction replay detected that the results of the two transactions are different. In these cases readwritesplit will abort the transaction and close the client connection.

Statements that result in an implicit commit do not reset the transaction when transaction_replay is enabled. This means that if the transaction is replayed, the transaction will be committed twice due to the implicit commit being present. The exception to this are the transaction management statements such as BEGIN and START TRANSACTION: they are detected and will cause the transaction to be correctly reset.

Any changes to the session state (e.g. autocommit state, SQL mode) done inside a transaction will remain in effect even if the connection to the server where the transaction is being executed fails. When readwritesplit creates a new connection to a server to replay the transaction, it will first restore the session state by executing all session commands that were executed. This means that if the session state is changed mid-transaction in a way that affects the results, transaction replay will fail.

The following partial transaction demonstrates the problem by using SQL_MODE inside a transaction.

SET SQL_MODE='';            -- A session command
BEGIN;
SELECT "hello world";       -- Returns the string "hello world"
SET SQL_MODE='ANSI_QUOTES'; -- A session command
SELECT 'hello world';       -- Returns the string "hello world"

If this transaction has to be replayed the actual SQL that gets executed is the following.

SET SQL_MODE='';            -- Replayed session command
SET SQL_MODE='ANSI_QUOTES'; -- Replayed session command
BEGIN;
SELECT "hello world";       -- Returns an error
SELECT 'hello world';       -- Returns the string "hello world"

First the session state is restored by executing all commands that changed the state after which the actual transaction is replayed. Due to the fact that the SQL_MODE was changed mid-transaction, one of the queries will now return an error instead of the result we expected leading to a transaction replay failure.

In a service-to-service configuration (i.e. a service using another service in its targets list ), if the topmost service starts a transaction, all lower-level readwritesplit services will also behave as if a transaction is open. If a connection to a backend database fails during this, it can result in unnecessary transaction replays which in turn can end up with checksum conflicts. The recommended approach is to not use any commands inside a transaction that would be routed to more than one node.

If the connection to the server where a transaction is being executed is lost while a ROLLBACK is being executed, readwritesplit will still attempt to replay the transaction in the hopes that the real response can be delivered to the client. However, this does mean that it is possible that a rolled back transaction which gets replayed ends up with a conflict and is reported as a replay failure when in reality a rolled back transaction could be safely ignored.

Legacy Configuration

In older versions of MaxScale, routers were configured via the router_options parameter. This functionality was deprecated in 2.2 and was removed in 2.3.

JDBC Batched Statements

Readwritesplit does not support pipelining of JDBC batched statements. This is caused by the fact that readwritesplit executes the statements one at a time to track the state of the response.

Limitations in multi-statement handling

When a multi-statement query is executed through the readwritesplit router, it will always be routed to the primary. See strict_multi_stmt for more details.

If the multi-statement query creates a temporary table, it will not be detected and reads to this table can be routed to replica servers. To prevent this, always execute the temporary table creation as an individual statement.

Limitations in client session handling

Some of the queries that a client sends are routed to all backends instead of just to one. These queries include USE <db name> and SET autocommit=0, among many others. Readwritesplit sends a copy of these queries to each backend server and forwards the primary's reply to the client. Below is a list of MySQL commands which are classified as session commands.

COM_INIT_DB (USE <db name> creates this)
COM_CHANGE_USER
COM_STMT_CLOSE
COM_STMT_SEND_LONG_DATA
COM_STMT_RESET
COM_STMT_PREPARE
COM_QUIT (no response, session is closed)
COM_REFRESH
COM_DEBUG
COM_PING
SQLCOM_CHANGE_DB (USE ... statements)
SQLCOM_DEALLOCATE_PREPARE
SQLCOM_PREPARE
SQLCOM_SET_OPTION
SELECT ..INTO variable|OUTFILE|DUMPFILE
SET autocommit=1|0

Prior to MaxScale 2.3.0, session commands that were 2²⁴ - 1 bytes or longer were not supported and caused the session to be closed.

There is a possibility for misbehavior. If USE mytable is executed in one of the replicas and fails, it may be due to replication lag rather than the database not existing. Thus, the same command may produce different result in different backend servers. The replicas which fail to execute a session command will be dropped from the active list of replicas for this session to guarantee a consistent session state across all the servers used by the session. In addition, the server will not be used again for routing for the duration of the session.

The above-mentioned behavior for user variables can be partially controlled with the configuration parameter use_sql_variables_in:

use_sql_variables_in=[master|all] (default: all)

WARNING

If a SELECT query modifies a user variable when the use_sql_variables_in parameter is set to all, it will not be routed and the client will receive an error. A log message is written into the log further explaining the reason for the error. Here is an example use of a SELECT query which modifies a user variable and how MariaDB MaxScale responds to it.

MySQL [(none)]> set @id=1;
Query OK, 0 rows affected (0.00 sec)

MySQL [(none)]> SELECT @id := @id + 1 FROM test.t1;
ERROR 1064 (42000): Routing query to backend failed. See the error log for further details.

Allow user variable modification in SELECT queries by setting use_sql_variables_in=master. This will route all queries that use user variables to the primary.

Comments

Comments loading...
Content reproduced on this site is the property of its respective owners, and this content is not reviewed in advance by MariaDB. The views, information and opinions expressed by this content do not necessarily represent those of MariaDB or any other party.