Chapter 3. Clustered Caches

You can use Data Grid in invalidation mode to optimize systems that perform high volumes of read operations. A good example is to use invalidation to prevent lots of database writes when state changes occur.

This cache mode only makes sense if you have another, permanent store for your data such as a database and are only using Data Grid as an optimization in a read-heavy system, to prevent hitting the database for every read. If a cache is configured for invalidation, every time data is changed in a cache, other caches in the cluster receive a message informing them that their data is now stale and should be removed from memory and from any local store.

Sometimes the application reads a value from the external store and wants to write it to the local cache, without removing it from the other nodes. To do this, it must call Cache.putForExternalRead(key, value) instead of Cache.put(key, value).

Invalidation mode can be used with a shared cache store. A write operation will both update the shared store, and it would remove the stale values from the other nodes' memory. The benefit of this is twofold: network traffic is minimized as invalidation messages are very small compared to replicating the entire value, and also other caches in the cluster look up modified data in a lazy manner, only when needed.

An invalidation cache can also be configured with a special cache loader, ClusterLoader. When ClusterLoader is enabled, read operations that do not find the key on the local node will request it from all the other nodes first, and store it in memory locally. In certain situation it will store stale values, so only use it if you have a high tolerance for stale values.

Invalidation mode can be synchronous or asynchronous. When synchronous, a write blocks until all nodes in the cluster have evicted the stale value. When asynchronous, the originator broadcasts invalidation messages but doesn’t wait for responses. That means other nodes still see the stale value for a while after the write completed on the originator.

Transactions can be used to batch the invalidation messages. Transactions acquire the key lock on the primary owner. To find more about how primary owners are assigned, please read the Key Ownership section.

Entries written to a replicated cache on any node will be replicated to all other nodes in the cluster, and can be retrieved locally from any node. Replicated mode provides a quick and easy way to share state across a cluster, however replication practically only performs well in small clusters (under 10 nodes), due to the number of messages needed for a write scaling linearly with the cluster size. Data Grid can be configured to use UDP multicast, which mitigates this problem to some degree.

Each key has a primary owner, which serializes data container updates in order to provide consistency. To find more about how primary owners are assigned, please read the Key Ownership section.

Replicated mode can be synchronous or asynchronous.

If transactions are enabled, write operations are not replicated through the primary owner.

Distribution tries to keep a fixed number of copies of any entry in the cache, configured as numOwners. This allows the cache to scale linearly, storing more data as nodes are added to the cluster.

As nodes join and leave the cluster, there will be times when a key has more or less than numOwners copies. In particular, if numOwners nodes leave in quick succession, some entries will be lost, so we say that a distributed cache tolerates numOwners - 1 node failures.

The number of copies represents a trade-off between performance and durability of data. The more copies you maintain, the lower performance will be, but also the lower the risk of losing data due to server or network failures. Regardless of how many copies are maintained, distribution still scales linearly, and this is key to Data Grid’s scalability.

The owners of a key are split into one primary owner, which coordinates writes to the key, and zero or more backup owners. To find more about how primary and backup owners are assigned, please read the Key Ownership section.

A read operation will request the value from the primary owner, but if it doesn’t respond in a reasonable amount of time, we request the value from the backup owners as well. (The infinispan.stagger.delay system property, in milliseconds, controls the delay between requests.) A read operation may require 0 messages if the key is present in the local cache, or up to 2 * numOwners messages if all the owners are slow.

A write operation will also result in at most 2 * numOwners messages: one message from the originator to the primary owner, numOwners - 1 messages from the primary to the backups, and the corresponding ACK messages.

Just as replicated mode, distributed mode can also be synchronous or asynchronous. And as in replicated mode, asynchronous replication is not recommended because it can lose updates. In addition to losing updates, asynchronous distributed caches can also see a stale value when a thread writes to a key and then immediately reads the same key.

Transactional distributed caches use the same kinds of messages as transactional replicated caches, except lock/prepare/commit/unlock messages are sent only to the affected nodes (all the nodes that own at least one key affected by the transaction) instead of being broadcast to all the nodes in the cluster. As an optimization, if the transaction writes to a single key and the originator is the primary owner of the key, lock messages are not replicated.

cache.get(k) -> v1
cache.put(k, v2)
cache.get(k) -> v2

cache.get(k) -> v2
cache.get(k) -> v1

<cache-container zero-capacity-node="true" />

new GlobalConfigurationBuilder().zeroCapacityNode(true);

<distributed-cache name="distributedCache" owners="2" segments="100" capacity-factor="2" />

Configuration c = new ConfigurationBuilder()
   .clustering()
      .cacheMode(CacheMode.DIST_SYNC)
      .hash()
         .numOwners(2)
         .numSegments(100)
         .capacityFactor(2)
   .build();

<transport initial-cluster-size="4" initial-cluster-timeout="30000" />

GlobalConfiguration global = new GlobalConfigurationBuilder()
   .transport()
   .initialClusterSize(4)
   .initialClusterTimeout(30000, TimeUnit.MILLISECONDS)
   .build();

<transport
    cluster="MyCluster"
    machine="LinuxServer01"
    rack="Rack01"
    site="US-WestCoast" />

// 1. Obtain a reference to a cache
Cache cache = ...
Address address = cache.getCacheManager().getAddress();

// 2. Create the affinity service
KeyAffinityService keyAffinityService = KeyAffinityServiceFactory.newLocalKeyAffinityService(
      cache,
      new RndKeyGenerator(),
      Executors.newSingleThreadExecutor(),
      100);

// 3. Obtain a key for which the local node is the primary owner
Object localKey = keyAffinityService.getKeyForAddress(address);

// 4. Insert the key in the cache
cache.put(localKey, "yourValue");

public interface Lifecycle {
   void start();
   void stop();
}

Configuration c = new ConfigurationBuilder()
   .clustering().hash().groups().enabled()
   .build();

<distributed-cache>
   <groups enabled="true"/>
</distributed-cache>

class User {
   ...
   String office;
   ...

   public int hashCode() {
      // Defines the hash for the key, normally used to determine location
      ...
   }

   // Override the location by specifying a group
   // All keys in the same group end up with the same owners
   @Group
   public String getOffice() {
      return office;
   }
   }
}

public interface Grouper<T> {
    String computeGroup(T key, String group);

    Class<T> getKeyType();
}

public class KXGrouper implements Grouper<String> {

   // The pattern requires a String key, of length 2, where the first character is
   // "k" and the second character is a digit. We take that digit, and perform
   // modular arithmetic on it to assign it to group "0" or group "1".
   private static Pattern kPattern = Pattern.compile("(^k)(<a>\\d</a>)$");

   public String computeGroup(String key, String group) {
      Matcher matcher = kPattern.matcher(key);
      if (matcher.matches()) {
         String g = Integer.parseInt(matcher.group(2)) % 2 + "";
         return g;
      } else {
         return null;
      }
   }

   public Class<String> getKeyType() {
      return String.class;
   }
}

Configuration c = new ConfigurationBuilder()
   .clustering().hash().groups().enabled().addGrouper(new KXGrouper())
   .build();

<distributed-cache>
   <groups enabled="true">
      <grouper class="com.acme.KXGrouper" />
   </groups>
</distributed-cache>

Scattered mode is very similar to Distribution Mode as it allows linear scaling of the cluster. It allows single node failure by maintaining two copies of the data (as Distribution Mode with numOwners=2). Unlike Distributed, the location of data is not fixed; while we use the same Consistent Hash algorithm to locate the primary owner, the backup copy is stored on the node that wrote the data last time. When the write originates on the primary owner, backup copy is stored on any other node (the exact location of this copy is not important).

This has the advantage of single Remote Procedure Call (RPC) for any write (Distribution Mode requires one or two RPCs), but reads have to always target the primary owner. That results in faster writes but possibly slower reads, and therefore this mode is more suitable for write-intensive applications.

Storing multiple backup copies also results in slightly higher memory consumption. In order to remove out-of-date backup copies, invalidation messages are broadcast in the cluster, which generates some overhead. This makes scattered mode less performant in very big clusters (this behaviour might be optimized in the future).

When a node crashes, the primary copy may be lost. Therefore, the cluster has to reconcile the backups and find out the last written backup copy. This process results in more network traffic during state transfer.

Since the writer of data is also a backup, even if we specify machine/rack/site ids on the transport level the cluster cannot be resilient to more than one failure on the same machine/rack/site.

Currently it is not possible to use scattered mode in transactional cache. Asynchronous replication is not supported either; use asynchronous Cache API instead. Functional commands are not implemented neither but these are expected to be added soon.

The cache is configured in a similar way as the other cache modes, here is an example of declarative configuration:

<scattered-cache name="scatteredCache" />

And this is how you can configure it programmatically:

Configuration c = new ConfigurationBuilder()
   .clustering().cacheMode(CacheMode.SCATTERED_SYNC)
   .build();

Scattered mode is not exposed in the server configuration as the server is usually accessed through the Hot Rod protocol. The protocol automatically selects primary owner for the writes and therefore the write (in distributed mode with two owner) requires single RPC inside the cluster, too. Therefore, scattered cache would not bring the performance benefit.