High Availability and Failover

We define high availability as the ability for the system to continue functioning after failure of one or more of the servers.

A part of high availability is failover which we define as the ability for client connections to migrate from one server to another in event of server failure so client applications can continue to operate.

Live - Backup Groups

Apache ActiveMQ Artemis allows servers to be linked together as live - backup groups where each live server can have 1 or more backup servers. A backup server is owned by only one live server. Backup servers are not operational until failover occurs, however 1 chosen backup, which will be in passive mode, announces its status and waits to take over the live servers work

Before failover, only the live server is serving the Apache ActiveMQ Artemis clients while the backup servers remain passive or awaiting to become a backup server. When a live server crashes or is brought down in the correct mode, the backup server currently in passive mode will become live and another backup server will become passive. If a live server restarts after a failover then it will have priority and be the next server to become live when the current live server goes down, if the current live server is configured to allow automatic failback then it will detect the live server coming back up and automatically stop.

HA Policies

Apache ActiveMQ Artemis supports two different strategies for backing up a server shared store and replication. Which is configured via the ha-policy configuration element.

<ha-policy>
  <replication/>
</ha-policy>

or

<ha-policy>
   <shared-store/>
</ha-policy>

As well as these 2 strategies there is also a 3rd called live-only. This of course means there will be no Backup Strategy and is the default if none is provided, however this is used to configure scale-down which we will cover in a later chapter.

Note

The ha-policy configurations replaces any current HA configuration in the root of the broker.xml configuration. All old configuration is now deprecated although best efforts will be made to honour it if configured this way.

Note

Only persistent message data will survive failover. Any non persistent message data will not be available after failover.

The ha-policy type configures which strategy a cluster should use to provide the backing up of a servers data. Within this configuration element is configured how a server should behave within the cluster, either as a master (live), slave (backup) or colocated (both live and backup). This would look something like:

<ha-policy>
   <replication>
      <master/>
   </replication>
</ha-policy>

or

<ha-policy>
   <shared-store/>
      <slave/>
   </shared-store/>
</ha-policy>

or

<ha-policy>
   <replication>
      <colocated/>
   </replication>
</ha-policy>

Data Replication

Support for network-based data replication was added in version 2.3.

When using replication, the live and the backup servers do not share the same data directories, all data synchronization is done over the network. Therefore all (persistent) data received by the live server will be duplicated to the backup.

Notice that upon start-up the backup server will first need to synchronize all existing data from the live server before becoming capable of replacing the live server should it fail. So unlike when using shared storage, a replicating backup will not be a fully operational backup right after start-up, but only after it finishes synchronizing the data with its live server. The time it will take for this to happen will depend on the amount of data to be synchronized and the connection speed.

Note

In general, synchronization occurs in parallel with current network traffic so this won't cause any blocking on current clients. However, there is a critical moment at the end of this process where the replicating server must complete the synchronization and ensure the replica acknowledges this completion. This exchange between the replicating server and replica will block any journal related operations. The maximum length of time that this exchange will block is controlled by the initial-replication-sync-timeout configuration element.

Replication will create a copy of the data at the backup. One issue to be aware of is: in case of a successful fail-over, the backup's data will be newer than the one at the live's storage. If you configure your live server to perform a failback to live server when restarted, it will synchronize its data with the backup's. If both servers are shutdown, the administrator will have to determine which one has the latest data.

The replicating live and backup pair must be part of a cluster. The Cluster Connection also defines how backup servers will find the remote live servers to pair with. Refer to Clusters for details on how this is done, and how to configure a cluster connection. Notice that:

  • Both live and backup servers must be part of the same cluster. Notice that even a simple live/backup replicating pair will require a cluster configuration.

  • Their cluster user and password must match.

Within a cluster, there are two ways that a backup server will locate a live server to replicate from, these are:

  • specifying a node group. You can specify a group of live servers that a backup server can connect to. This is done by configuring group-name in either the master or the slave element of the broker.xml. A Backup server will only connect to a live server that shares the same node group name

  • connecting to any live. This will be the behaviour if group-name is not configured allowing a backup server to connect to any live server

Note

A group-name example: suppose you have 5 live servers and 6 backup servers:

  • live1, live2, live3: with group-name=fish

  • live4, live5: with group-name=bird

  • backup1, backup2, backup3, backup4: with group-name=fish

  • backup5, backup6: with group-name=bird

After joining the cluster the backups with group-name=fish will search for live servers with group-name=fish to pair with. Since there is one backup too many, the fish will remain with one spare backup.

The 2 backups with group-name=bird (backup5 and backup6) will pair with live servers live4 and live5.

The backup will search for any live server that it is configured to connect to. It then tries to replicate with each live server in turn until it finds a live server that has no current backup configured. If no live server is available it will wait until the cluster topology changes and repeats the process.

Note

This is an important distinction from a shared-store backup, if a backup starts and does not find a live server, the server will just activate and start to serve client requests. In the replication case, the backup just keeps waiting for a live server to pair with. Note that in replication the backup server does not know whether any data it might have is up to date, so it really cannot decide to activate automatically. To activate a replicating backup server using the data it has, the administrator must change its configuration to make it a live server by changing slave to master.

Much like in the shared-store case, when the live server stops or crashes, its replicating backup will become active and take over its duties. Specifically, the backup will become active when it loses connection to its live server. This can be problematic because this can also happen because of a temporary network problem. In order to address this issue, the backup will try to determine whether it still can connect to the other servers in the cluster. If it can connect to more than half the servers, it will become active, if more than half the servers also disappeared with the live, the backup will wait and try reconnecting with the live. This avoids a split brain situation.

Configuration

To configure the live and backup servers to be a replicating pair, configure the live server in ' broker.xml to have:

<ha-policy>
   <replication>
      <master/>
   </replication>
</ha-policy>
.
<cluster-connections>
   <cluster-connection name="my-cluster">
      ...
   </cluster-connection>
</cluster-connections>

The backup server must be similarly configured but as a slave

<ha-policy>
   <replication>
      <slave/>
   </replication>
</ha-policy>

All Replication Configuration

The following table lists all the ha-policy configuration elements for HA strategy Replication for master:

Name Description
`check-for-live-server` Whether to check the cluster for a (live) server using our own server ID when starting up. This option is only necessary for performing 'fail-back' on replicating servers.
`cluster-name` Name of the cluster configuration to use for replication. This setting is only necessary if you configure multiple cluster connections. If configured then the connector configuration of the cluster configuration with this name will be used when connecting to the cluster to discover if a live server is already running, see `check-for-live-server`. If unset then the default cluster connections configuration is used (the first one configured).
`group-name` If set, backup servers will only pair with live servers with matching group-name.
`initial-replication-sync-timeout` The amount of time the replicating server will wait at the completion of the initial replication process for the replica to acknowledge it has received all the necessary data. The default is 30,000 milliseconds. Note: during this interval any journal related operations will be blocked.

The following table lists all the ha-policy configuration elements for HA strategy Replication for slave:

Name Description
`cluster-name` Name of the cluster configuration to use for replication. This setting is only necessary if you configure multiple cluster connections. If configured then the connector configuration of the cluster configuration with this name will be used when connecting to the cluster to discover if a live server is already running, see `check-for-live-server`. If unset then the default cluster connections configuration is used (the first one configured)
`group-name` If set, backup servers will only pair with live servers with matching group-name
`max-saved-replicated-journals-size` This specifies how many times a replicated backup server can restart after moving its files on start. Once there are this number of backup journal files the server will stop permanently after if fails back.
`allow-failback` Whether a server will automatically stop when a another places a request to take over its place. The use case is when the backup has failed over
`initial-replication-sync-timeout` After failover and the slave has become live, this is set on the new live server. It represents the amount of time the replicating server will wait at the completion of the initial replication process for the replica to acknowledge it has received all the necessary data. The default is 30,000 milliseconds. Note: during this interval any journal related operations will be blocked.

Shared Store

When using a shared store, both live and backup servers share the same entire data directory using a shared file system. This means the paging directory, journal directory, large messages and binding journal.

When failover occurs and a backup server takes over, it will load the persistent storage from the shared file system and clients can connect to it.

This style of high availability differs from data replication in that it requires a shared file system which is accessible by both the live and backup nodes. Typically this will be some kind of high performance Storage Area Network (SAN). We do not recommend you use Network Attached Storage (NAS), e.g. NFS mounts to store any shared journal (NFS is slow).

The advantage of shared-store high availability is that no replication occurs between the live and backup nodes, this means it does not suffer any performance penalties due to the overhead of replication during normal operation.

The disadvantage of shared store replication is that it requires a shared file system, and when the backup server activates it needs to load the journal from the shared store which can take some time depending on the amount of data in the store.

If you require the highest performance during normal operation, have access to a fast SAN and live with a slightly slower failover (depending on amount of data).

ActiveMQ Artemis ha-shared-store.png

Configuration

To configure the live and backup servers to share their store, configure id via the ha-policy configuration in broker.xml:

<ha-policy>
   <shared-store>
      <master/>
   </shared-store>
</ha-policy>
.
<cluster-connections>
   <cluster-connection name="my-cluster">
...
   </cluster-connection>
</cluster-connections>

The backup server must also be configured as a backup.

<ha-policy>
   <shared-store>
      <slave/>
   </shared-store>
</ha-policy>

In order for live - backup groups to operate properly with a shared store, both servers must have configured the location of journal directory to point to the same shared location (as explained in Configuring the message journal)

Note

todo write something about GFS

Also each node, live and backups, will need to have a cluster connection defined even if not part of a cluster. The Cluster Connection info defines how backup servers announce there presence to its live server or any other nodes in the cluster. Refer to Clusters for details on how this is done.

Failing Back to live Server

After a live server has failed and a backup taken has taken over its duties, you may want to restart the live server and have clients fail back.

In case of "shared disk", simply restart the original live server and kill the new live server by can do this by killing the process itself. Alternatively you can set allow-fail-back to true on the slave config which will force the backup that has become live to automatically stop. This configuration would look like:

<ha-policy>
   <shared-store>
      <slave>
         <allow-failback>true</allow-failback>
      </slave>
   </shared-store>
</ha-policy>

In replication HA mode you need to set an extra property check-for-live-server to true in the master configuration. If set to true, during start-up a live server will first search the cluster for another server using its nodeID. If it finds one, it will contact this server and try to "fail-back". Since this is a remote replication scenario, the "starting live" will have to synchronize its data with the server running with its ID, once they are in sync, it will request the other server (which it assumes it is a back that has assumed its duties) to shutdown for it to take over. This is necessary because otherwise the live server has no means to know whether there was a fail-over or not, and if there was if the server that took its duties is still running or not. To configure this option at your broker.xml configuration file as follows:

<ha-policy>
   <replication>
      <master>
         <check-for-live-server>true</check-for-live-server>
      <master>
   </replication>
</ha-policy>

Warning

Be aware that if you restart a live server while after failover has occurred then check-for-live-server must be set to true. If not the live server will restart and server the same messages that the backup has already handled causing duplicates.

It is also possible, in the case of shared store, to cause failover to occur on normal server shutdown, to enable this set the following property to true in the ha-policy configuration on either the master or slave like so:

<ha-policy>
   <shared-store>
      <master>
         <failover-on-shutdown>true</failover-on-shutdown>
      </master>
   </shared-store>
</ha-policy>

By default this is set to false, if by some chance you have set this to false but still want to stop the server normally and cause failover then you can do this by using the management API as explained at Management

You can also force the running live server to shutdown when the old live server comes back up allowing the original live server to take over automatically by setting the following property in the broker.xml configuration file as follows:

<ha-policy>
   <shared-store>
      <slave>
         <allow-failback>true</allow-failback>
      </slave>
   </shared-store>
</ha-policy>

All Shared Store Configuration

The following table lists all the ha-policy configuration elements for HA strategy shared store for master:

Name Description
`failover-on-server-shutdown` If set to true then when this server is stopped normally the backup will become live assuming failover. If false then the backup server will remain passive. Note that if false you want failover to occur the you can use the the management API as explained at [Management](management.md)
`wait-for-activation` If set to true then server startup will wait until it is activated. If set to false then server startup will be done in the background. Default is true.

The following table lists all the ha-policy configuration elements for HA strategy Shared Store for slave:

Name Description
`failover-on-server-shutdown` In the case of a backup that has become live. then when set to true then when this server is stopped normally the backup will become liveassuming failover. If false then the backup server will remain passive. Note that if false you want failover to occur the you can use the the management API as explained at [Management](management.md)
`allow-failback` Whether a server will automatically stop when a another places a request to take over its place. The use case is when the backup has failed over.

Colocated Backup Servers

It is also possible when running standalone to colocate backup servers in the same JVM as another live server. Live Servers can be configured to request another live server in the cluster to start a backup server in the same JVM either using shared store or replication. The new backup server will inherit its configuration from the live server creating it apart from its name, which will be set to colocated_backup_n where n is the number of backups the server has created, and any directories and its Connectors and Acceptors which are discussed later on in this chapter. A live server can also be configured to allow requests from backups and also how many backups a live server can start. this way you can evenly distribute backups around the cluster. This is configured via the ha-policy element in the broker.xml file like so:

<ha-policy>
   <replication>
      <colocated>
         <request-backup>true</request-backup>
         <max-backups>1</max-backups>
         <backup-request-retries>-1</backup-request-retries>
         <backup-request-retry-interval>5000</backup-request-retry-interval>
         <master/>
         <slave/>
      </colocated>
   <replication>
</ha-policy>

the above example is configured to use replication, in this case the master and slave configurations must match those for normal replication as in the previous chapter. shared-store is also supported

ActiveMQ Artemis ha-colocated.png

Configuring Connectors and Acceptors

If the HA Policy is colocated then connectors and acceptors will be inherited from the live server creating it and offset depending on the setting of backup-port-offset configuration element. If this is set to say 100 (which is the default) and a connector is using port 61616 then this will be set to 5545 for the first server created, 5645 for the second and so on.

Note

for INVM connectors and Acceptors the id will have colocated_backup_n appended, where n is the backup server number.

Remote Connectors

It may be that some of the Connectors configured are for external servers and hence should be excluded from the offset. for instance a Connector used by the cluster connection to do quorum voting for a replicated backup server, these can be omitted from being offset by adding them to the ha-policy configuration like so:

<ha-policy>
   <replication>
      <colocated>
         <excludes>
            <connector-ref>remote-connector</connector-ref>
         </excludes>
.........
</ha-policy>

Configuring Directories

Directories for the Journal, Large messages and Paging will be set according to what the HA strategy is. If shared store the the requesting server will notify the target server of which directories to use. If replication is configured then directories will be inherited from the creating server but have the new backups name appended.

The following table lists all the ha-policy configuration elements for colocated policy:

Name Description
`request-backup` If true then the server will request a backup on another node
`backup-request-retries` How many times the live server will try to request a backup, -1 means for ever.
`backup-request-retry-interval` How long to wait for retries between attempts to request a backup server.
`max-backups` How many backups a live server can create
`backup-port-offset` The offset to use for the Connectors and Acceptors when creating a new backup server.

Scaling Down

An alternative to using Live/Backup groups is to configure scaledown. when configured for scale down a server can copy all its messages and transaction state to another live server. The advantage of this is that you dont need full backups to provide some form of HA, however there are disadvantages with this approach the first being that it only deals with a server being stopped and not a server crash. The caveat here is if you configure a backup to scale down.

Another disadvantage is that it is possible to lose message ordering. This happens in the following scenario, say you have 2 live servers and messages are distributed evenly between the servers from a single producer, if one of the servers scales down then the messages sent back to the other server will be in the queue after the ones already there, so server 1 could have messages 1,3,5,7,9 and server 2 would have 2,4,6,8,10, if server 2 scales down the order in server 1 would be 1,3,5,7,9,2,4,6,8,10.

ActiveMQ Artemis ha-scaledown.png

The configuration for a live server to scale down would be something like:

<ha-policy>
   <live-only>
      <scale-down>
         <connectors>
            <connector-ref>server1-connector</connector-ref>
         </connectors>
      </scale-down>
   </live-only>
</ha-policy>

In this instance the server is configured to use a specific connector to scale down, if a connector is not specified then the first INVM connector is chosen, this is to make scale down fromm a backup server easy to configure. It is also possible to use discovery to scale down, this would look like:

<ha-policy>
   <live-only>
      <scale-down>
         <discovery-group-ref discovery-group-name="my-discovery-group"/>
      </scale-down>
   </live-only>
</ha-policy>

Scale Down with groups

It is also possible to configure servers to only scale down to servers that belong in the same group. This is done by configuring the group like so:

<ha-policy>
   <live-only>
      <scale-down>
         ...
         <group-name>my-group</group-name>
      </scale-down>
   </live-only>
</ha-policy>

In this scenario only servers that belong to the group my-group will be scaled down to

Scale Down and Backups

It is also possible to mix scale down with HA via backup servers. If a slave is configured to scale down then after failover has occurred, instead of starting fully the backup server will immediately scale down to another live server. The most appropriate configuration for this is using the colocated approach. it means as you bring up live server they will automatically be backed up by server and as live servers are shutdown, there messages are made available on another live server. A typical configuration would look like:

<ha-policy>
   <replication>
      <colocated>
         <backup-request-retries>44</backup-request-retries>
         <backup-request-retry-interval>33</backup-request-retry-interval>
         <max-backups>3</max-backups>
         <request-backup>false</request-backup>
         <backup-port-offset>33</backup-port-offset>
         <master>
            <group-name>purple</group-name>
            <check-for-live-server>true</check-for-live-server>
            <cluster-name>abcdefg</cluster-name>
         </master>
         <slave>
            <group-name>tiddles</group-name>
            <max-saved-replicated-journals-size>22</max-saved-replicated-journals-size>
            <cluster-name>33rrrrr</cluster-name>
            <restart-backup>false</restart-backup>
            <scale-down>
               <!--a grouping of servers that can be scaled down to-->
               <group-name>boo!</group-name>
               <!--either a discovery group-->
               <discovery-group-ref discovery-group-name="wahey"/>
            </scale-down>
         </slave>
      </colocated>
   </replication>
</ha-policy>

Scale Down and Clients

When a server is stopping and preparing to scale down it will send a message to all its clients informing them which server it is scaling down to before disconnecting them. At this point the client will reconnect however this will only succeed once the server has completed scaledown. This is to ensure that any state such as queues or transactions are there for the client when it reconnects. The normal reconnect settings apply when the client is reconnecting so these should be high enough to deal with the time needed to scale down.

Failover Modes

Apache ActiveMQ Artemis defines two types of client failover:

  • Automatic client failover

  • Application-level client failover

Apache ActiveMQ Artemis also provides 100% transparent automatic reattachment of connections to the same server (e.g. in case of transient network problems). This is similar to failover, except it is reconnecting to the same server and is discussed in Client Reconnection and Session Reattachment

During failover, if the client has consumers on any non persistent or temporary queues, those queues will be automatically recreated during failover on the backup node, since the backup node will not have any knowledge of non persistent queues.

Automatic Client Failover

Apache ActiveMQ Artemis clients can be configured to receive knowledge of all live and backup servers, so that in event of connection failure at the client - live server connection, the client will detect this and reconnect to the backup server. The backup server will then automatically recreate any sessions and consumers that existed on each connection before failover, thus saving the user from having to hand-code manual reconnection logic.

Apache ActiveMQ Artemis clients detect connection failure when it has not received packets from the server within the time given by client-failure-check-period as explained in section Detecting Dead Connections. If the client does not receive data in good time, it will assume the connection has failed and attempt failover. Also if the socket is closed by the OS, usually if the server process is killed rather than the machine itself crashing, then the client will failover straight away.

Apache ActiveMQ Artemis clients can be configured to discover the list of live-backup server groups in a number of different ways. They can be configured explicitly or probably the most common way of doing this is to use server discovery for the client to automatically discover the list. For full details on how to configure server discovery, please see Clusters. Alternatively, the clients can explicitly connect to a specific server and download the current servers and backups see Clusters.

To enable automatic client failover, the client must be configured to allow non-zero reconnection attempts (as explained in Client Reconnection and Session Reattachment).

By default failover will only occur after at least one connection has been made to the live server. In other words, by default, failover will not occur if the client fails to make an initial connection to the live server - in this case it will simply retry connecting to the live server according to the reconnect-attempts property and fail after this number of attempts.

Failing over on the Initial Connection

Since the client does not learn about the full topology until after the first connection is made there is a window where it does not know about the backup. If a failure happens at this point the client can only try reconnecting to the original live server. To configure how many attempts the client will make you can set the property initialConnectAttempts on the ClientSessionFactoryImpl or ActiveMQConnectionFactory or initial-connect-attempts in xml. The default for this is 0, that is try only once. Once the number of attempts has been made an exception will be thrown.

For examples of automatic failover with transacted and non-transacted JMS sessions, please see the examples chapter.

A Note on Server Replication

Apache ActiveMQ Artemis does not replicate full server state between live and backup servers. When the new session is automatically recreated on the backup it won't have any knowledge of messages already sent or acknowledged in that session. Any in-flight sends or acknowledgements at the time of failover might also be lost.

By replicating full server state, theoretically we could provide a 100% transparent seamless failover, which would avoid any lost messages or acknowledgements, however this comes at a great cost: replicating the full server state (including the queues, session, etc.). This would require replication of the entire server state machine; every operation on the live server would have to replicated on the replica server(s) in the exact same global order to ensure a consistent replica state. This is extremely hard to do in a performant and scalable way, especially when one considers that multiple threads are changing the live server state concurrently.

It is possible to provide full state machine replication using techniques such as virtual synchrony, but this does not scale well and effectively serializes all operations to a single thread, dramatically reducing concurrency.

Other techniques for multi-threaded active replication exist such as replicating lock states or replicating thread scheduling but this is very hard to achieve at a Java level.

Consequently it has decided it was not worth massively reducing performance and concurrency for the sake of 100% transparent failover. Even without 100% transparent failover, it is simple to guarantee once and only once delivery, even in the case of failure, by using a combination of duplicate detection and retrying of transactions. However this is not 100% transparent to the client code.

Handling Blocking Calls During Failover

If the client code is in a blocking call to the server, waiting for a response to continue its execution, when failover occurs, the new session will not have any knowledge of the call that was in progress. This call might otherwise hang for ever, waiting for a response that will never come.

To prevent this, Apache ActiveMQ Artemis will unblock any blocking calls that were in progress at the time of failover by making them throw a javax.jms.JMSException (if using JMS), or a ActiveMQException with error code ActiveMQException.UNBLOCKED. It is up to the client code to catch this exception and retry any operations if desired.

If the method being unblocked is a call to commit(), or prepare(), then the transaction will be automatically rolled back and Apache ActiveMQ Artemis will throw a javax.jms.TransactionRolledBackException (if using JMS), or a ActiveMQException with error code ActiveMQException.TRANSACTION_ROLLED_BACK if using the core API.

Handling Failover With Transactions

If the session is transactional and messages have already been sent or acknowledged in the current transaction, then the server cannot be sure that messages sent or acknowledgements have not been lost during the failover.

Consequently the transaction will be marked as rollback-only, and any subsequent attempt to commit it will throw a javax.jms.TransactionRolledBackException (if using JMS), or a ActiveMQException with error code ActiveMQException.TRANSACTION_ROLLED_BACK if using the core API.

Warning

The caveat to this rule is when XA is used either via JMS or through the core API. If 2 phase commit is used and prepare has already been called then rolling back could cause a HeuristicMixedException. Because of this the commit will throw a XAException.XA_RETRY exception. This informs the Transaction Manager that it should retry the commit at some later point in time, a side effect of this is that any non persistent messages will be lost. To avoid this use persistent messages when using XA. With acknowledgements this is not an issue since they are flushed to the server before prepare gets called.

It is up to the user to catch the exception, and perform any client side local rollback code as necessary. There is no need to manually rollback the session - it is already rolled back. The user can then just retry the transactional operations again on the same session.

Apache ActiveMQ Artemis ships with a fully functioning example demonstrating how to do this, please see the examples chapter.

If failover occurs when a commit call is being executed, the server, as previously described, will unblock the call to prevent a hang, since no response will come back. In this case it is not easy for the client to determine whether the transaction commit was actually processed on the live server before failure occurred.

Note

If XA is being used either via JMS or through the core API then an XAException.XA_RETRY is thrown. This is to inform Transaction Managers that a retry should occur at some point. At some later point in time the Transaction Manager will retry the commit. If the original commit has not occurred then it will still exist and be committed, if it does not exist then it is assumed to have been committed although the transaction manager may log a warning.

To remedy this, the client can simply enable duplicate detection (Duplicate Message Detection) in the transaction, and retry the transaction operations again after the call is unblocked. If the transaction had indeed been committed on the live server successfully before failover, then when the transaction is retried, duplicate detection will ensure that any durable messages resent in the transaction will be ignored on the server to prevent them getting sent more than once.

Note

By catching the rollback exceptions and retrying, catching unblocked calls and enabling duplicate detection, once and only once delivery guarantees for messages can be provided in the case of failure, guaranteeing 100% no loss or duplication of messages.

Handling Failover With Non Transactional Sessions

If the session is non transactional, messages or acknowledgements can be lost in the event of failover.

If you wish to provide once and only once delivery guarantees for non transacted sessions too, enabled duplicate detection, and catch unblock exceptions as described in Handling Blocking Calls During Failover

Getting Notified of Connection Failure

JMS provides a standard mechanism for getting notified asynchronously of connection failure: java.jms.ExceptionListener. Please consult the JMS javadoc or any good JMS tutorial for more information on how to use this.

The Apache ActiveMQ Artemis core API also provides a similar feature in the form of the class org.apache.activemq.artemis.core.client.SessionFailureListener

Any ExceptionListener or SessionFailureListener instance will always be called by ActiveMQ Artemis on event of connection failure, irrespective of whether the connection was successfully failed over, reconnected or reattached, however you can find out if reconnect or reattach has happened by either the failedOver flag passed in on the connectionFailed on SessionfailureListener or by inspecting the error code on the javax.jms.JMSException which will be one of the following:

JMSException error codes

Error code Description
FAILOVER Failover has occurred and we have successfully reattached or reconnected.
DISCONNECT No failover has occurred and we are disconnected.

Application-Level Failover

In some cases you may not want automatic client failover, and prefer to handle any connection failure yourself, and code your own manually reconnection logic in your own failure handler. We define this as application-level failover, since the failover is handled at the user application level.

To implement application-level failover, if you're using JMS then you need to set an ExceptionListener class on the JMS connection. The ExceptionListener will be called by Apache ActiveMQ Artemis in the event that connection failure is detected. In your ExceptionListener, you would close your old JMS connections, potentially look up new connection factory instances from JNDI and creating new connections.

For a working example of application-level failover, please see the examples chapter.

If you are using the core API, then the procedure is very similar: you would set a FailureListener on the core ClientSession instances.

results matching ""

    No results matching ""