Chapter 2. Clustered caches

Even with synchronous replication, distributed caches are not linearizable. For transactional caches, they do not support serialization/snapshot isolation.

For example, a thread is carrying out a single put request:

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

But another thread might see the values in a different order:

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

The reason is that read can return the value from any owner, depending on how fast the primary owner replies. The write is not atomic across all the owners. In fact, the primary commits the update only after it receives a confirmation from the backup. While the primary is waiting for the confirmation message from the backup, reads from the backup will see the new value, but reads from the primary will see the old one.

Distributed caches split entries into a fixed number of segments and assign each segment to a list of owner nodes. Replicated caches do the same, with the exception that every node is an owner.

The first node in the list of owners is the primary owner. The other nodes in the list are backup owners. When the cache topology changes, because a node joins or leaves the cluster, the segment ownership table is broadcast to every node. This allows nodes to locate keys without making multicast requests or maintaining metadata for each key.

The numSegments property configures the number of segments available. However, the number of segments cannot change unless the cluster is restarted.

Likewise the key-to-segment mapping cannot change. Keys must always map to the same segments regardless of cluster topology changes. It is important that the key-to-segment mapping evenly distributes the number of segments allocated to each node while minimizing the number of segments that must move when the cluster topology changes.

Hashing configuration

You can configure ConsistentHashFactory implementations, including custom ones, with embedded caches only.

<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();

Capacity factors allocate the number of segments based on resources available to each node in the cluster.

The capacity factor for a node applies to segments for which that node is both the primary owner and backup owner. In other words, the capacity factor specifies is the total capacity that a node has in comparison to other nodes in the cluster.

The default value is 1 which means that all nodes in the cluster have an equal capacity and Data Grid allocates the same number of segments to all nodes in the cluster.

However, if nodes have different amounts of memory available to them, you can configure the capacity factor so that the Data Grid hashing algorithm assigns each node a number of segments weighted by its capacity.

The value for the capacity factor configuration must be a positive number and can be a fraction such as 1.5. You can also configure a capacity factor of 0 but is recommended only for nodes that join the cluster temporarily and should use the zero capacity configuration instead.

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

{
  "infinispan" : {
    "cache-container" : {
      "zero-capacity-node" : "true"
    }
  }
}

infinispan:
  cacheContainer:
    zeroCapacityNode: "true"

new GlobalConfigurationBuilder().zeroCapacityNode(true);

Data Grid nodes create local replicas when they retrieve entries from another node in the cluster. L1 caches avoid repeatedly looking up entries on primary owner nodes and adds performance.

The following diagram illustrates how L1 caches work:

Figure 2.4. L1 cache

In the "L1 cache" diagram:

L1 caching performance

Enabling L1 improves performance for read operations but requires primary owner nodes to broadcast invalidation messages when entries are modified. This ensures that Data Grid removes any out of date replicas across the cluster. However this also decreases performance of write operations and increases memory usage, reducing overall capacity of caches.

L1 cache configuration

<distributed-cache l1-lifespan="5000"
                   l1-cleanup-interval="60000">
</distributed-cache>

{
  "distributed-cache": {
    "l1-lifespan": "5000",
    "l1-cleanup-interval": "60000"
  }
}

distributedCache:
  l1Lifespan: "5000"
  l1-cleanup-interval: "60000"

ConfigurationBuilder builder = new ConfigurationBuilder();
builder.clustering().cacheMode(CacheMode.DIST_SYNC)
         .l1()
         .lifespan(5000, TimeUnit.MILLISECONDS)
         .cleanupTaskFrequency(60000, TimeUnit.MILLISECONDS);

Server hinting increases availability of data in distributed caches by replicating entries across as many servers, racks, and data centers as possible.

When Data Grid distributes the copies of your data, it follows the order of precedence: site, rack, machine, and node. All of the configuration attributes are optional. For example, when you specify only the rack IDs, then Data Grid distributes the copies across different racks and nodes.

Server hinting can impact cluster rebalancing operations by moving more segments than necessary if the number of segments for the cache is too low.

Server hinting configuration

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

{
  "infinispan" : {
    "cache-container" : {
      "transport" : {
        "cluster" : "MyCluster",
        "machine" : "LinuxServer01",
        "rack" : "Rack01",
        "site" : "US-WestCoast"
      }
    }
  }
}

cacheContainer:
  transport:
    cluster: "MyCluster"
    machine: "LinuxServer01"
    rack: "Rack01"
    site: "US-WestCoast"

GlobalConfigurationBuilder global = GlobalConfigurationBuilder.defaultClusteredBuilder()
  .transport()
  .clusterName("MyCluster")
  .machineId("LinuxServer01")
  .rackId("Rack01")
  .siteId("US-WestCoast");

In a distributed cache, a key is allocated to a list of nodes with an opaque algorithm. There is no easy way to reverse the computation and generate a key that maps to a particular node. However, Data Grid can generate a sequence of (pseudo-)random keys, see what their primary owner is, and hand them out to the application when it needs a key mapping to a particular node.

Following code snippet depicts how a reference to this service can be obtained and used.

// 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");

The service is started at step 2: after this point it uses the supplied Executor to generate and queue keys. At step 3, we obtain a key from the service, and at step 4 we use it.

Lifecycle

KeyAffinityService extends Lifecycle, which allows stopping and (re)starting it:

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

The service is instantiated through KeyAffinityServiceFactory. All the factory methods have an Executor parameter, that is used for asynchronous key generation (so that it won’t happen in the caller’s thread). It is the user’s responsibility to handle the shutdown of this Executor.

The KeyAffinityService, once started, needs to be explicitly stopped. This stops the background key generation and releases other held resources.

The only situation in which KeyAffinityService stops by itself is when the cache manager with which it was registered is shutdown.

Topology changes

When the cache topology changes, the ownership of the keys generated by the KeyAffinityService might change. The key affinity service keep tracks of these topology changes and doesn’t return keys that would currently map to a different node, but it won’t do anything about keys generated earlier.

As such, applications should treat KeyAffinityService purely as an optimization, and they should not rely on the location of a generated key for correctness.

In particular, applications should not rely on keys generated by KeyAffinityService for the same address to always be located together. Collocation of keys is only provided by the Grouping API.

Complementary to the Key affinity service, the Grouping API allows you to co-locate a group of entries on the same nodes, but without being able to select the actual nodes.

By default, the segment of a key is computed using the key’s hashCode(). If you use the Grouping API, Data Grid will compute the segment of the group and use that as the segment of the key.

When the Grouping API is in use, it is important that every node can still compute the owners of every key without contacting other nodes. For this reason, the group cannot be specified manually. The group can either be intrinsic to the entry (generated by the key class) or extrinsic (generated by an external function).

To use the Grouping API, you must enable groups.

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

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

If you have control of the key class (you can alter the class definition, it’s not part of an unmodifiable library), then we recommend using an intrinsic group. The intrinsic group is specified by adding the @Group annotation to a method, for example:

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;
   }
   }
}

If you don’t have control over the key class, or the determination of the group is an orthogonal concern to the key class, we recommend using an extrinsic group. An extrinsic group is specified by implementing the Grouper interface.

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

    Class<T> getKeyType();
}

If multiple Grouper classes are configured for the same key type, all of them will be called, receiving the value computed by the previous one. If the key class also has a @Group annotation, the first Grouper will receive the group computed by the annotated method. This allows you even greater control over the group when using an intrinsic group.

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;
   }
}

Grouper implementations must be registered explicitly in the cache configuration. If you are configuring Data Grid programmatically:

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

Or, if you are using XML:

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

Advanced API

AdvancedCache has two group-specific methods:

Both methods iterate over the entire data container and store (if present), so they can be slow when a cache contains lots of small groups.

Because the Cache interface extends java.util.Map, write methods like put(key, value) and remove(key) return the previous value by default.

In some cases, the return value may not be correct:

Transactional caches return the correct previous value in cases 3 and 4. However, transactional caches also have a gotcha: in distributed mode, the read-committed isolation level is implemented as repeatable-read. That means this example of "double-checked locking" won’t work:

Cache cache = ...
TransactionManager tm = ...

tm.begin();
try {
   Integer v1 = cache.get(k);
   // Increment the value
   Integer v2 = cache.put(k, v1 + 1);
   if (Objects.equals(v1, v2) {
      // success
   } else {
      // retry
   }
} finally {
  tm.commit();
}

The correct way to implement this is to use cache.getAdvancedCache().withFlags(Flag.FORCE_WRITE_LOCK).get(k).

In caches with optimistic locking, writes can also return stale previous values. Write skew checks can avoid stale previous values.