Apache
/
activemq-artemis
mirror of https://github.com/apache/activemq-artemis.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
activemq-artemis/docs/user-manual/en/message-grouping.md

15 KiB
Raw Blame History

Message Grouping

Message groups are sets of messages that have the following characteristics:

  • Messages in a message group share the same group id, i.e. they have same group identifier property (JMSXGroupID for JMS, _AMQ_GROUP_ID for Apache ActiveMQ Artemis Core API).

  • Messages in a message group are always consumed by the same consumer, even if there are many consumers on a queue. They pin all messages with the same group id to the same consumer. If that consumer closes another consumer is chosen and will receive all messages with the same group id.

Message groups are useful when you want all messages for a certain value of the property to be processed serially by the same consumer.

An example might be orders for a certain stock. You may want orders for any particular stock to be processed serially by the same consumer. To do this you can create a pool of consumers (perhaps one for each stock, but less will work too), then set the stock name as the value of the _AMQ_GROUP_ID property.

This will ensure that all messages for a particular stock will always be processed by the same consumer.

Note:

Grouped messages can impact the concurrent processing of non-grouped messages due to the underlying FIFO semantics of a queue. For example, if there is a chunk of 100 grouped messages at the head of a queue followed by 1,000 non-grouped messages then all the grouped messages will need to be sent to the appropriate client (which is consuming those grouped messages serially) before any of the non-grouped messages can be consumed. The functional impact in this scenario is a temporary suspension of concurrent message processing while all the grouped messages are processed. This can be a performance bottleneck so keep it in mind when determining the size of your message groups, and consider whether or not you should isolate your grouped messages from your non-grouped messages.

Using Core API

The property name used to identify the message group is "_AMQ_GROUP_ID" (or the constant MessageImpl.HDR_GROUP_ID). Alternatively, you can set autogroup to true on the SessionFactory which will pick a random unique id.

Using JMS

The property name used to identify the message group is JMSXGroupID.

// send 2 messages in the same group to ensure the same
// consumer will receive both
Message message = ...
message.setStringProperty("JMSXGroupID", "Group-0");
producer.send(message);

message = ...
message.setStringProperty("JMSXGroupID", "Group-0");
producer.send(message);

Alternatively, you can set autogroup to true on the ActiveMQConnectonFactory which will pick a random unique id. This can also be set in the JNDI context environment, e.g. jndi.properties. Here's a simple example using the "ConnectionFactory" connection factory which is available in the context by default

java.naming.factory.initial=org.apache.activemq.artemis.jndi.ActiveMQInitialContextFactory
connectionFactory.myConnectionFactory=tcp://localhost:61616?autoGroup=true

Alternatively you can set the group id via the connection factory. All messages sent with producers created via this connection factory will set the JMSXGroupID to the specified value on all messages sent. This can also be set in the JNDI context environment, e.g. jndi.properties. Here's a simple example using the "ConnectionFactory" connection factory which is available in the context by default:

java.naming.factory.initial=org.apache.activemq.artemis.jndi.ActiveMQInitialContextFactory
connectionFactory.myConnectionFactory=tcp://localhost:61616?groupID=Group-0

Closing a Message Group

You generally don't need to close a message group, you just keep using it.

However if you really do want to close a group you can add a negative sequence number.

Example:

Mesasge message = session.createTextMessage("<foo>hey</foo>");
message.setStringProperty("JMSXGroupID", "Group-0");
message.setIntProperty("JMSXGroupSeq", -1);
...
producer.send(message);

This then closes the message group so if another message is sent in the future with the same message group ID it will be reassigned to a new consumer.

Notifying Consumer of Group Ownership change

ActiveMQ supports putting a boolean header, set on the first message sent to a consumer for a particular message group.

To enable this, you must set a header key that the broker will use to set the flag.

In the examples we use JMSXGroupFirstForConsumer but it can be any header key value you want.

By setting group-first-key to JMSXGroupFirstForConsumer at the queue level, every time a new group is assigned a consumer the header JMSXGroupFirstForConsumer will be set to true on the first message.

<address name="foo.bar">
   <multicast>
      <queue name="orders1" group-first-key="JMSXGroupFirstForConsumer"/>
   </multicast>
</address>

Or on auto-create when using the JMS Client by using address parameters when creating the destination used by the consumer.

Queue queue = session.createQueue("my.destination.name?group-first-key=JMSXGroupFirstForConsumer");
Topic topic = session.createTopic("my.destination.name?group-first-key=JMSXGroupFirstForConsumer");

Also the default for all queues under and address can be defaulted using the address-setting configuration:

<address-setting match="my.address">
   <default-group-first-key>JMSXGroupFirstForConsumer</default-group-first-key>
</address-setting>

By default this is null, and therefor OFF.

Rebalancing Message Groups

Sometimes after new consumers are added you can find that if you have long lived groups, that they have no groups assigned, and thus are not being utilised, this is because the long lived groups will already be assigned to existing consumers.

It is possibly to rebalance the groups.

note during the split moment of reset, a message to the original associated consumer could be in flight at the same time, a new message for the same group is dispatched to the new associated consumer.

Manually

via the management API or managment console by invoking resetAllGroups

Automatically

By setting group-rebalance to true at the queue level, every time a consumer is added it will trigger a rebalance/reset of the groups.

As noted above, when group rebalance is done, there is a risk you may have inflight messages being processed, by default the broker will continue to dispatch whilst rebalance is occuring. To ensure that inflight messages are processed before dispatch of new messages post rebalance, to different consumers, you can set group-rebalance-pause-dispatch to true which will cause the dispatch to pause whilst rebalance occurs, until all inflight messages are processed.

<address name="foo.bar">
   <multicast>
      <queue name="orders1" group-rebalance="true" group-rebalance-pause-dispatch="true"/>
   </multicast>
</address>

Or on auto-create when using the JMS Client by using address parameters when creating the destination used by the consumer.

Queue queue = session.createQueue("my.destination.name?group-rebalance=true&group-rebalance-pause-dispatch=true");
Topic topic = session.createTopic("my.destination.name?group-rebalance=true&group-rebalance-pause-dispatch=true");

Also the default for all queues under and address can be defaulted using the address-setting configuration:

<address-setting match="my.address">
   <default-group-rebalance>true</default-group-rebalance>
   <default-group-rebalance-pause-dispatch>true</default-group-rebalance-pause-dispatch>
</address-setting>

By default, default-group-rebalance is false meaning this is disabled/off. By default, default-group-rebalance-pause-dispatch is false meaning this is disabled/off.

Group Buckets

For handling groups in a queue with bounded memory allowing better scaling of groups, you can enable group buckets, essentially the group id is hashed into a bucket instead of keeping track of every single group id.

Setting group-buckets to -1 keeps default behaviour which means the queue keeps track of every group but suffers from unbounded memory use.

Setting group-buckets to 0 disables grouping (0 buckets), on a queue. This can be useful on a multicast address, where many queues exist but one queue you may not care for ordering and prefer to keep round robin behaviour.

There is a number of ways to set group-buckets.

<address name="foo.bar">
   <multicast>
      <queue name="orders1" group-buckets="1024"/>
   </multicast>
</address>

Specified on creating a Queue by using the CORE api specifying the parameter group-buckets to 20.

Or on auto-create when using the JMS Client by using address parameters when creating the destination used by the consumer.

Queue queue = session.createQueue("my.destination.name?group-buckets=1024");
Topic topic = session.createTopic("my.destination.name?group-buckets=1024");

Also the default for all queues under and address can be defaulted using the address-setting configuration:

<address-setting match="my.bucket.address">
   <default-group-buckets>1024</default-group-buckets>
</address-setting>

By default, default-group-buckets is -1 this is to keep compatibility with existing default behaviour.

Address wildcards can be used to configure group-buckets for a set of addresses.

Example

See the Message Group Example which shows how message groups are configured and used with JMS and via a connection factory.

Clustered Grouping

Before looking at the details for configuring clustered grouping support it is worth examing the idea of clustered grouping as a whole. In general, combining clustering and message grouping is a poor choice because the fundamental ideas of grouped (i.e. ordered) messages and horizontal scaling through clustering are essentially at odds with each other.

Message grouping enforces ordered message consumption. Ordered message consumption requires that each message be fully consumed and acknowledged before the next message in the group is consumed. This results in serial message processing (i.e. no concurrency).

However, the idea of clustering is to scale brokers horizontally in order to increase message throughput by adding consumers which can process messages concurrently. But since the message groups are ordered the messages in each group cannot be consumed concurrently which defeats the purpose of horizontal scaling.

Clustered grouping is not recommended for these reasons.

However, if you've evaluated your overall use-case with these design caveats in mind and have determined that clustered grouping is still viable then read on for all the configuration details and best practices.

Clustered Grouping Configuration

Using message groups in a cluster is a bit more complex. This is because messages with a particular group id can arrive on any node so each node needs to know about which group id's are bound to which consumer on which node. The consumer handling messages for a particular group id may be on a different node of the cluster, so each node needs to know this information so it can route the message correctly to the node which has that consumer.

To solve this there is the notion of a grouping handler. Each node will have its own grouping handler and when a messages is sent with a group id assigned, the handlers will decide between them which route the message should take.

Here is a sample config for each type of handler. This should be configured in broker.xml.

<grouping-handler name="my-grouping-handler">
   <type>LOCAL</type>
   <address>jms</address>
   <timeout>5000</timeout>
</grouping-handler>

<grouping-handler name="my-grouping-handler">
   <type>REMOTE</type>
   <address>jms</address>
   <timeout>5000</timeout>
</grouping-handler>

  • type two types of handlers are supported - LOCAL and REMOTE. Each cluster should choose 1 node to have a LOCAL grouping handler and all the other nodes should have REMOTE handlers. It's the LOCAL handler that actually makes the decision as to what route should be used, all the other REMOTE handlers converse with this.

  • address refers to a cluster connection and the address it uses. Refer to the clustering section on how to configure clusters.

  • timeout how long to wait for a decision to be made. An exception will be thrown during the send if this timeout is reached, this ensures that strict ordering is kept.

The decision as to where a message should be routed to is initially proposed by the node that receives the message. The node will pick a suitable route as per the normal clustered routing conditions, i.e. round robin available queues, use a local queue first and choose a queue that has a consumer. If the proposal is accepted by the grouping handlers the node will route messages to this queue from that point on, if rejected an alternative route will be offered and the node will again route to that queue indefinitely. All other nodes will also route to the queue chosen at proposal time. Once the message arrives at the queue then normal single server message group semantics take over and the message is pinned to a consumer on that queue.

You may have noticed that there is a single point of failure with the single local handler. If this node crashes then no decisions will be able to be made. Any messages sent will be not be delivered and an exception thrown. To avoid this happening Local Handlers can be replicated on another backup node. Simple create your back up node and configure it with the same Local handler.

Clustered Grouping Best Practices

Some best practices should be followed when using clustered grouping:

  1. Make sure your consumers are distributed evenly across the different nodes if possible. This is only an issue if you are creating and closing consumers regularly. Since messages are always routed to the same queue once pinned, removing a consumer from this queue may leave it with no consumers meaning the queue will just keep receiving the messages. Avoid closing consumers or make sure that you always have plenty of consumers, i.e., if you have 3 nodes have 3 consumers.

  2. Use durable queues if possible. If queues are removed once a group is bound to it, then it is possible that other nodes may still try to route messages to it. This can be avoided by making sure that the queue is deleted by the session that is sending the messages. This means that when the next message is sent it is sent to the node where the queue was deleted meaning a new proposal can successfully take place. Alternatively you could just start using a different group id.

  3. Always make sure that the node that has the Local Grouping Handler is replicated. These means that on failover grouping will still occur.

  4. In case you are using group-timeouts, the remote node should have a smaller group-timeout with at least half of the value on the main coordinator. This is because this will determine how often the last-time-use value should be updated with a round trip for a request to the group between the nodes.

Clustered Grouping Example

See the Clustered Grouping Example which shows how to configure message groups with a ActiveMQ Artemis Cluster.