[
https://issues.apache.org/jira/browse/SLING-10077?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17494430#comment-17494430
]
Joerg Hoh commented on SLING-10077:
-----------------------------------
I personally do not think that using the DEA is a good idea to distribute
events with a potentially high frequency in the cluster, because it adds
latency to the event consumption. And even more problematic is the I/O behavior
because for every single event a new ResourceResolver is created and the event
is written to the repository. This adds load to the clustered environment.
Isn't it possible to use the existing approach we already have and determine
per distribution event if the distribution action has originated from the local
node? In case it was sent from another node, this event can be discarded. In
this case we would not add further load to the repo.
> Mode to raise events only locally
> ---------------------------------
>
> Key: SLING-10077
> URL: https://issues.apache.org/jira/browse/SLING-10077
> Project: Sling
> Issue Type: Improvement
> Affects Versions: Content Distribution Journal Core 0.1.18
> Reporter: Timothee Maret
> Assignee: Timothee Maret
> Priority: Major
>
> Currently the {{o/a/s/d/a/p/queued}} and {{o/a/s/d/a/p/distributed}} events
> are raised as local events in each author instance of a cluster. Handling
> those events in each author instance makes sense for use cases that require
> maintaining a distributed data structure like a cache. However, some use
> cases like content invalidation, would need to handle the events only once.
> Currently, event handlers can support the later category of use cases by
> deduplicating the events using the Apache Discovery API. This works but it
> adds complexity in the event handlers and in some edge case, some events may
> be lost.
> To solve both use cases better, we should move from raising local events in
> each author instance to raising the events only on the author leader instance
> and [distribute the
> events|https://sling.apache.org/documentation/bundles/apache-sling-eventing-and-job-handling.html#distributed-events]
> across the cluster. Use cases that need to de-duplicate will be able to do
> that by filtering the events with
> {code}
> if (org.apache.sling.event.EventUtil.isLocal(event)) {
> // Process the event on the instance that sent it and thus once per
> cluster
> }
> {code}
> The PackageDistributedNotifier would be run only on the cluster leader
> instance. The leader instance can be detected using the [Discovery
> API|https://sling.apache.org/documentation/bundles/discovery-api-and-impl.html#topology-changes].
> We'd need to introduce a new component that listen to topology change and in
> case it runs on the leader, register the PackageDistributedNotifier service
> using
> {code}
> org.osgi.framework.BundleContext.registerService(PackageDistributedNotifier.class.getName(),
> packageDistributedNotifier, props)
> {code}
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
