At the core of Coherence is the concept of clustered data management. This implies the following goals:

• A fully coherent, single system image (SSI)

• Scalability for both read and write access

• Fast, transparent failover and failback

• Linear scalability for storage and processing

• No Single-Points-of-Failure (SPOFs)

• Cluster-wide locking and transactions

Built on top of this foundation are the various services that Coherence provides, including database caching, HTTP session management, grid agent invocation and distributed queries. Before going into detail about these features, some basic aspects of Coherence should be discussed.

Coherence supports many topologies for clustered data management. Each of these topologies has a trade-off in terms of performance and fault-tolerance. By using a single API, the choice of topology can be deferred until deployment if desired. This allows developers to work with a consistent logical view of Coherence, while providing flexibility during tuning or as application needs change.

Coherence provides several cache implementations:

• Local Cache – Local on-heap caching for non-clustered caching. See Understanding Local Caches.

• Distributed Cache – True linear scalability for both read and write access. Data is automatically, dynamically and transparently partitioned across nodes. The distribution algorithm minimizes network traffic and avoids service pauses by incrementally shifting data. See Understanding Distributed Caches.

• Near Cache – Provides the performance of local caching with the scalability of distributed caching. Several different near-cache strategies are available and offer a trade-off between performance and synchronization guarantees. See Understanding Near Caches.

• View Cache – Perfect for small, read-heavy caches. See Understanding View Caches.

In-process caching provides the highest level of raw performance, because objects are managed within the local JVM. This benefit is most directly realized by the Local, View, and Near Cache implementations.

Out-of-process (client/server) caching provides the option of using dedicated cache servers. This can be helpful when you want to partition workloads (to avoid stressing the application servers). This is accomplished by using the Partitioned cache implementation and simply disabling local storage on client nodes through a single command-line option or a one-line entry in the XML configuration.

Tiered caching (using the Near Cache functionality) enables you to couple local caches on the application server with larger, partitioned caches on the cache servers, combining the raw performance of local caching with the scalability of partitioned caching. This is useful for both dedicated cache servers and co-located caching (cache partitions stored within the application server JVMs).

See Using Caches.

While most customers use on-heap storage combined with dedicated cache servers, Coherence has several options for data storage:

• On-heap—The fastest option, though it can affect JVM garbage collection times.

• Journal—A combination of RAM storage and disk storage, optimized for solid state disks, that uses a journaling technique. Journal-based storage requires serialization/deserialization.

• File-based—Uses a Berkeley Database JE storage system.

Coherence storage is transient: the disk-based storage options are for managing cached data only. For persistent storage, Coherence offers backing maps coupled with a CacheLoader/CacheStore.

See Implementing Storage and Backing Maps.

Because serialization is often the most expensive part of clustered data management, Coherence provides the following options for serializing/deserializing data:

• com.tangosol.io.pof.PofSerializer – The Portable Object Format (also referred to as POF) is a language agnostic binary format. POF was designed to be incredibly efficient in both space and time and is the recommended serialization option in Coherence. See Using Portable Object Format.

• java.io.Serializable – The simplest, but slowest option.

• java.io.Externalizable – This requires developers to implement serialization manually, but can provide significant performance benefits. Compared to java.io.Serializable, this can cut serialized data size by a factor of two or more (especially helpful with Distributed caches, as they generally cache data in serialized form). Most importantly, CPU usage is dramatically reduced.

• com.tangosol.io.ExternalizableLite – This is very similar to java.io.Externalizable, but offers better performance and less memory usage by using a more efficient IO stream implementation.

• com.tangosol.run.xml.XmlBean – A default implementation of ExternalizableLite.

Coherence's API provides access to all Coherence functionality. The most commonly used subset of this API is exposed through simple XML options to minimize effort for typical use cases. There is no penalty for mixing direct configuration through the API with the easier XML configuration.

Coherence is designed to allow the replacement of its modules as needed. For example, the local "backing maps" (which provide the actual physical data storage on each node) can be easily replaced as needed. The vast majority of the time, this is not required, but it is there for the situations that require it. The general guideline is that 80% of tasks are easy, and the remaining 20% of tasks (the special cases) require a little more effort, but certainly can be done without significant hardship.

Coherence is organized as set of services. At the root is the Cluster service. A cluster is defined as a set of Coherence instances (one instance per JVM, with one or more JVMs on each computer). See Introduction to Coherence Clusters. Under the cluster service are the various services that comprise the Coherence API. These include the various caching services (Distributed, Federated, and so on) and the Invocation Service (for deploying agents to various nodes of the cluster). Each instance of a service is named, and there is typically a default service instance for each type. The cache services contain named caches (com.tangosol.net.NamedCache), which are analogous to database tables—that is, they typically contain a set of related objects.

The following source code returns a reference to a NamedCache instance. The underlying cache service is started if necessary.

import com.tangosol.net.*;
...
NamedCache cache = CacheFactory.getCache("MyCache");

Coherence scans the cache configuration XML file for a name mapping for MyCache. This is similar to Servlet name mapping in a web container's web.xml file. Coherence's cache configuration file contains (in the simplest case) a set of mappings (from cache name to cache scheme) and a set of cache schemes.

By default, Coherence uses the coherence-cache-config.xml file found at the root of coherence.jar. This can be overridden on the JVM command-line with -Dcoherence.cacheconfig=file.xml. This argument can reference either a file system path, or a Java resource path.

The com.tangosol.net.NamedCache interface extends several other interfaces:

• java.util.Map—basic Map methods such as get(), put(), remove().

• com.tangosol.net.cache.CacheMap—methods for getting a collection of keys (as a Map) that are in the cache and for putting objects in the cache. Also supports adding an expiry value when putting an entry in a cache.

• com.tangosol.util.QueryMap—methods for querying the cache. See Querying Data in a Cache.

• com.tangosol.util.InvocableMap—methods for server-side processing of cache data. See Processing Data In a Cache.

• com.tangosol.util.ObservableMap—methods for listening to cache events. See Using Map Events.

• com.tangosol.util.ConcurrentMap—methods for concurrent access such as lock() and unlock(). See Performing Transactions.

Coherence was originally designed to perform these tasks using synchronized() blocks (intrinsic java locks) because multiple threads can perform the above simultaneously.

Because intrinsic locks are non-interruptible, this approach could lead to application threads blocking indefinitely, waiting for the lock when, for example, a network outage occurs.

Interruptible locks provide the ability to do a try-lock that times out. This allows an application to not be indefinitely blocked. With this improvement, you can now start a cluster or cache within a try-with-resources block using a thread local timeout. See Class Timeout.

try (Timeout t = Timeout.after(20, TimeUnit.SECONDS))
    { 
    CacheFactory.getCache("dist"); // or CacheFactory.ensureCluster();
    }
catch (Exception e)
    {  
    // doSomething else.
    }

The default lock timeout is Long.MAX_VALUE. If the operation (getCache, ensureCluster, and so on) does not return within the specified timeout, an InterruptedException is displayed. If the corresponding locks cannot be acquired within the specified timeout, a RequestTimeoutException is displayed. The service request timeout is also honored for the service lock, if specified. For example, when references to caches are maintained and member restarts occur, underlying resources must be obtained again on the next call performed on a NamedCache reference. In this case, a configured request timeout will prevent blocking indefinitely.

If both request timeout and thread local timeout are specified, the latter takes precedence.

There are two general approaches to using a NamedCache:

• As a clustered implementation of java.util.Map with several added features (queries, concurrency), but with no persistent backing (a "side" cache).

• As a means of decoupling access to external data sources (an "inline" cache). In this case, the application uses the NamedCache interface, and the NamedCache takes care of managing the underlying database (or other resource).

Typically, an inline cache is used to cache data from:

• a database—The most intuitive use of a cache—simply caching database tables (in the form of Java objects).

• a service—Mainframe, web service, service bureau—any service that represents an expensive resource to access (either due to computational cost or actual access fees).

• calculations—Financial calculations, aggregations, data transformations. Using an inline cache makes it very easy to avoid duplicating calculations. If the calculation is complete, the result is simply pulled from the cache. Since any serializable object can be used as a cache key, it is a simple matter to use an object containing calculation parameters as the cache key.

See Caching Data Sources.

Write-back options:

• write-through—Ensures that the external data source always contains up-to-date information. Used when data must be persisted immediately, or when sharing a data source with other applications.

• write-behind—Provides better performance by caching writes to the external data source. Not only can writes be buffered to even out the load on the data source, but multiple writes can be combined, further reducing I/O. The trade-off is that data is not immediately persisted to disk; however, it is immediately distributed across the cluster, so the data survives the loss of a server. Furthermore, if the entire data set is cached, this option means that the application can survive a complete failure of the data source temporarily as both cache reads and writes do not require synchronous access the data source.