Tigase Advanced Clustering Strategy for PubSub (ACS-PubSub)

Tigase Team <team@tigase.com> v3.4.0-SNAPSHOT, 2023-11-05 :numbered:

Overview

ACS for PubSub allows seamless clustering of PubSub nodes across Tigase XMPP server cluster installation. ACS for PubSub is required for clustered PubSub deployments. If offers various strategies to handle distribution of traffic and nodes across the cluster, which allows fine-tune configuration of the deployment to individual needs.

Tigase ACS PubSub Configuration

In order to use ACS for PubSub, main Advance Clustering Strategy (ACS) is required. Once it’s enabled, clustered version of PubSub component will be selected by default during startup therefore it’s not required to configure it explicitly (make sure no other class is configured).

pubsub () {}

It’s also possible to explicitly configure the class with the following configuration:

pubsub (class: tigase.pubsub.cluster.PubSubComponentClustered) {}

With the above configuration default ACS PubSub clustering strategy will be used. In order to select different strategy you have to configure it’s class for strategy bean within pubsub component bean:

pubsub () {
    strategy (class: tigase.pubsub.cluster.PartitionedStrategy) {}
}

ACS PubSub Strategies

ACS-PubSub Partitioned Strategy

Warning

This strategy is now deprecated.

Short description

This is the simplest strategy that can be used by Tigase ACS - PubSub component in which particular PubSub node is handled by only one cluster node.

Description of processing

In this strategy all configuration of nodes of the same PubSub service JID is done by the same node of a cluster which is dynamically selected based on hash of service JID and number of connected nodes within cluster. After any change to node configuration is done, then node established for processing manipulation of this particular PubSub node is notified about details of the PubSub node and detailed changes made to it’s configuration. Remaining nodes do not require to know about changes.

Presences sent from users to PubSub service will be handled on the local node of the cluster and then distributed to every node of a cluster as events.

Messages sent from users to PubSub service will be handled on the local node of the cluster.

Presence and IQ stanzas will be processed according to rules outlined below.

Rules of processing packets:
  • presence - packets are processed on the local cluster node and then other cluster nodes are notified about changed presence as each cluster node is responsible for handling different PubSub nodes
  • message - is always processed locally
  • iq - cluster node which will process this packet is selected based on following rules:

    • CAPS query responses will be processed on the local node
    • non-PubSub packets (no pubsub subelement) are processed on local node
    • PubSub related packets without PubSub node name or packets that change PubSub node configuration (and contains node name and one of the following subelements: create, configure, default, delete) are processed on cluster node selected on hashcode derived from to attribute (service JID) and number of cluster nodes (this is done deal with concurrency issues between configuration changes)
    • remaining PubSub packets (non-configuration and containing node name) are processed on cluster node selected based on to attribute of packet and name of PubSub node

Note

This strategy for every packet should force processing of packet on only one cluster node

Note

This strategy will react on PubSub configuration node change and will send notification to cluster node responsible for processing items for PubSub node (selected based on to attribute of packet, name of PubSub node) to trigger cached PubSub node configuration refresh

Note

Result of packets generated on remote node should not be filtered, so if packet from one cluster node was forwarded to other cluster node for processing, then response packet should not be filtered when this PubSub clustering strategy is used.

ACS-PubSub Clustered Node Strategy

Short description

This strategy is used by default by Tigase ACS - PubSub component in which each PubSub node is handled on every cluster node but each cluster node will contain only partial information about user connections. This way strategy is better suited for deployments with PubSub nodes having a lot of subscribers. The benefit of using ClusterNodeStrategy in this case is reduced network traffic on cluster connections, as most of notifications and retrieval of items will be handled on the same cluster node to which user is connected.

Description of processing

In this strategy almost all packets are processed on the local cluster node. Only packets related to settings default options (such as pubsub packets containing subelements options or default) will be forwared to all cluster nodes and processed on each of them. If packets will result in the node configuration, subscriptions or affiliations changes other nodes will be notified to refresh node configuration/subscriptions/affiliations.

Presences sent from users to PubSub service will be handled on the local node of the cluster and then distributed to every node of a cluster as events.

Messages sent from users to PubSub service will be handled on the local node of the cluster.

If every other IQ stanza sent to PubSub service which is does not change configuration of a node and is item publication stanza then this stanza will always be processed on local node (as data retrieval/removal may be done only on one node as items are not cached).

IQ stanzas which are stanza responsible for publication are processed on the local cluster node. In this case PubSub will able to properly send notifications (as notification are generated always on local user node, if user is connected, which reduces cluster network traffic).

Rules of processing packets:
  • presence - packets are delivered to every cluster node as each cluster node is responsible for handling different PubSub nodes
  • message - is always processed locally
  • iq - cluster node which will process this packet is selected based on following rules:

    • CAPS query responses will be processed on the local node
    • non-PubSub packets (no pubsub subelement) are processed on local node
    • PubSub related packets without PubSub node name are processed on the local node. If packets will result in the node configuration changes, other nodes will be notified to refresh node configuration.
    • PubSub related packets containing subelements named options or default will be forwarded to all connected cluster nodes
    • remaining PubSub packets are processed on the local cluster node. If packets will result in the node configuration changes, other nodes will be notified to refresh node configuration.

Note

This strategy will react to PubSub configuration node change and will send notification to every cluster node (as every cluster node is responsible for processing items for every PubSub node) to refresh particular PubSub node configuration.

Note

Result of processing of packets generated on remote node should be filtered if packet was also processed on local node, so if packet from one cluster node was forwarded to other cluster node for processing but was not processed locally, then response packet should be filtered when this PubSub clustering strategy is used.

In a nutshell

Clustered Node Strategy distributes processing of nodes across cluster by processing all requests locally - each cluster node has complete knowledge of all PubSub nodes and processed requests for all of them, but generates notifications only for users connected to this particular node.

Partitioned Strategy assigns PubSub node to cluster node and handle all processing on that particular cluster node (configuration change, generating notification packets).