Clustering and Cache Coordination

An application cluster is a set of middle tier server machines or VMs servicing requests for a single application, or set of applications. Multiple servers are used to increase the scalability of the application and/or to provide fault tolerance and high availability. Typically the same application will be deployed to all of the servers in the cluster and application requests will be load balanced across the set of servers. The application cluster will access a single database, or a database cluster. An application cluster may allow new servers to be added to increase scalability, and for servers to be removed such as for updates and servicing.

Application clusters can consist of Java EE servers, Web containers, or Java server applications.

EclipseLink can function in any clustered environment. The main issue in a clustered environment is utilizing a shared persistence unit (L2) cache. If you are using a shared cache (enabled by default in EclipseLink projects), then each server will maintain its own cache, and each caches data can get out of synchronization with the other servers and the database.

EclipseLink provides cache coordination in a clustered environment to ensure the servers caches are synchronized.

There are also many other solutions to caching in a clustered environment, including:

Disable the shared cache (through setting @Cacheable(false), or @Cache(isolation=ISOLATED)).
Use only cache read-only objects.
Set a cache invalidation timeout to reduce stale data.
Use refreshing on objects/queries when fresh data is required.
Use optimistic locking to ensure write consistency (writes on stale data will fail, and will automatically invalidate the cache).
Use database events to invalidate changed data in the cache (such as EclipseLink's support for Oracle Query Change Notification).

Cache coordination enables a set of persistence units deployed to different servers in the cluster (or on the same server) to synchronize their changes. Cache coordination works by each persistence unit on each server in the cluster being able to broadcast notification of transactional object changes to the other persistence units in the cluster. EclipseLink supports cache coordination over RMI and JMS. The cache coordination framework is also extensible so other options could be developed.

By default, EclipseLink optimizes concurrency to minimize cache locking during read or write operations. Use the default EclipseLink transaction isolation configuration unless you have a very specific reason to change it.

Cache coordination works by broadcasting changes for each transaction to the other servers in the cluster. Each other server will receive the change notification, and either invalidate the changed objects in their cache, or update the cached objects state with the changes. Cache coordination occurs after the database commit, so only committed changes are broadcast.

Cache coordination greatly reduces to chance of an application getting stale data, but does not eliminate the possibility. Optimistic locking should still be used to ensure data integrity. Even in a single server application stale data is still possible within a persistence context unless pessimistic locking is used. Optimistic (or pessimistic) locking is always required to ensure data integrity in any multi-user system.

For more information about cache coordination, including cache synchronization, see "Using Cache Coordination" in Solutions Guide for EclipseLink.

Coordinating JMS and RMI Caches

For a JMS coordinated caches, when a particular session's coordinated cache starts, it uses its JNDI naming service information to locate and create a connection to the JMS server. The coordinated cache is ready when all participating sessions are connected to the same topic on the same JMS server. At this point, sessions can start sending and receiving object change messages. You can then configure all sessions that are participating in the same coordinated cache with the same JMS and JNDI naming service information.

For an RMI coordinated cache, when a particular session's coordinated cache starts, the session binds its connection in its naming service (either an RMI registry or JNDI), creates an announcement message (that includes its own naming service information), and broadcasts the announcement to its multicast group. When a session that belongs to the same multicast group receives this announcement, it uses the naming service information in the announcement message to establish bidirectional connections with the newly announced session's coordinated cache. The coordinated cache is ready when all participating sessions are interconnected in this way, at which point sessions can start sending and receiving object change messages. You can then configure each session with naming information that identifies the host on which the session is deployed.

For more information on configuring JMS and RMI cache coordination, see "Configuring JMS Cache Coordination Using Persistence Properties" and "Configuring RMI Cache Coordination Using Persistence Properties" in Solutions Guide for EclipseLink.

Coordinating Custom Caches

You can define your own custom solutions for coordinated caches by using the classes in the EclipseLink org.eclipse.persistence.sessions.coordination package.