What is clustering?
Magnolia can be configured to run in a clustered environment to provide high availability and load balancing:
- High-availability clusters are also known as fail-over clusters. Their purpose is to ensure that content is served at all times. They operate by having redundant instances which are used to provide service when a public instance fails. The most common size for a high-availability cluster is two public instances, the standard Magnolia setup. In such a setup the redundant instance may even be dormant (not actively serving content) until it is called to service.
- Load-balancing clusters connect many instances together to share the workload. From a technical standpoint there are multiple instances but from the website visitor's perspective they function as a single virtual instance. A load balancer distributes requests from visitors to instances in the cluster. The result is a balanced computational workload among different instances, improving the performance of the site as a whole.
Conditions for clustering
If you consider to use clustering - make sure the following conditions are true for your project and the environment you are running it:
- You are in full control of the virtual machines (VMs) on which the cluster is running.
- Your public instances are not treated as "disposable" with automatic up/down scaling based on load
Your public instances use the database (cluster) running on separate VM that is not disposable.
- The amount of writing (JCR) data is not too high.
Each write action generates a record in the journal table that is used to synchronize all cluster nodes - thus generating additional chatter and load on each node (also due to sync of index).
- You do not update Magnolia too frequently or do not require high availability.
(When updating clusters, all cluster nodes except one need to be shutdown, and restarted with new libs after the first node was updated successfully.)
When above conditions are not met, it is in all likelihood better to avoid using cluster and distribute data by other means instead.
(For instance use REST services to write data directly from client or using Magnolia Public instances as proxy writing to such externalized service and being accessible from/via all instances).
We use Jackrabbit's clustering feature to share content between Magnolia instances. Clustering in Jackrabbit works on the principle that content is shared between all cluster nodes. This means that all Jackrabbit cluster nodes need access to the same persistent storage (persistence manager and data store). The persistence manager must be clusterable. Any database is clusterable by its very nature as it stores content by unique hash IDs. However, each cluster node needs its own (private) file system and search index. For more details see Jackrabbit clustering documentation.
Individual workspaces can be mapped to different repositories. The repository that holds shared content can reside in clustered storage. This is useful for content that needs to be in sync across all instances.
In the diagram, each Magnolia instance uses its own persistent storage for storing the content of
config workspaces. However, the
comments workspace has shared content that is stored in a clustered storage, the database in the middle.
User generated content such as comments written by site visitors is a typical clustering case. Imagine that users John and Jane request the same web page. A load balancer redirects John to public instance A. When John leaves a comment on the page, his comment is stored in a workspace that resides in clustered storage. Now Jane requests the same page. The load balancer redirects her to public instance B. John's comment is immediately visible to Jane since both instances tap into the same clustered storage.
Other examples of shared content are user accounts, permissions of public users and forum posts. They need to be available regardless of the instance that serves the page.
Cleaning the Jackrabbit journal
Cluster nodes write their changes to a journal that can become very large over time. By default, old revisions are not removed automatically so that you could easily add new cluster nodes. Jackrabbit 1.5 introduced automatic cleaning of the database-based journal using a process called Janitor.
Note the following about using Janitor:
- When Janitor is enabled you can no longer easily add cluster nodes, but it is still possible.
- All cluster nodes need to have written their local revision to the database before the clean-up task runs for the first time.
- If a cluster node is removed permanently, its entry in the
LOCAL_REVISIONStable should be removed manually.