A newer version of Hazelcast Platform is available.

View latest

Data Affinity

6.0-SNAPSHOT 5.5 5.4 5.3 5.2 5.1 5.0

Data affinity ensures that related entries exist on the same member. If related data is on the same member, operations can be executed without the cost of extra network calls and extra wire data. This feature is provided by using the same partition keys for related data.

PartitionAware

Co-location of related data and computation

Hazelcast has a standard way of finding out which member owns/manages each key object. The following operations are routed to the same member, since all them are operating based on the same key "key1".

HazelcastInstance hazelcastInstance = Hazelcast.newHazelcastInstance();
Map mapA = hazelcastInstance.getMap( "mapA" );
Map mapB = hazelcastInstance.getMap( "mapB" );
Map mapC = hazelcastInstance.getMap( "mapC" );

// since map names are different, operation will be manipulating
// different entries, but the operation will take place on the
// same member since the keys ("key1") are the same
mapA.put( "key1", value );
mapB.get( "key1" );
mapC.remove( "key1" );

// lock operation will still execute on the same member
// of the cluster since the key ("key1") is same
hazelcastInstance.getLock( "key1" ).lock();

// distributed execution will execute the 'runnable' on the
// same member since "key1" is passed as the key.
hazelcastInstance.getExecutorService().executeOnKeyOwner( runnable, "key1" );

When the keys are the same, entries are stored on the same member. But we sometimes want to have related entries stored on the same member, such as a customer and his/her order entries. We would have a customers map with customerId as the key and an orders map with orderId as the key. Since customerId and orderId are different keys, a customer and his/her orders may fall into different members in your cluster. So how can we have them stored on the same member? We create an affinity between customer and orders. If we make them part of the same partition then these entries will be co-located. We achieve this by making orderKey s PartitionAware.

final class OrderKey implements PartitionAware, Serializable {

    private final long orderId;
    private final long customerId;

    OrderKey(long orderId, long customerId) {
        this.orderId = orderId;
        this.customerId = customerId;
    }

    @Override
    public Object getPartitionKey() {
        return customerId;
    }

    @Override
    public String toString() {
        return "OrderKey{"
                + "orderId=" + orderId
                + ", customerId=" + customerId
                + '}';

Notice that OrderKey implements PartitionAware and that getPartitionKey() returns the customerId. These make sure that the Customer entry and its Orders are stored on the same member.

HazelcastInstance hazelcastInstance = Hazelcast.newHazelcastInstance();
Map mapCustomers = hazelcastInstance.getMap( "customers" );
Map mapOrders = hazelcastInstance.getMap( "orders" );

// create the customer entry with customer id = 1
mapCustomers.put( 1, customer );

// now create the orders for this customer
mapOrders.put( new OrderKey( 21, 1 ), order );
mapOrders.put( new OrderKey( 22, 1 ), order );
mapOrders.put( new OrderKey( 23, 1 ), order );

Assume that you have a customers map where customerId is the key and the customer object is the value. You want to remove one of the customer orders and return the number of remaining orders. Here is how you would normally do it.

public static int removeOrder( long customerId, long orderId ) throws Exception {
    IMap<Long, Customer> mapCustomers = instance.getMap( "customers" );
    IMap mapOrders = hazelcastInstance.getMap( "orders" );

    mapCustomers.lock( customerId );
    mapOrders.remove( new OrderKey(orderId, customerId) );
    Set orders = orderMap.keySet(Predicates.equal( "customerId", customerId ));
    mapCustomers.unlock( customerId );

    return orders.size();
}

There are couple of things you should consider.

  • There are four distributed operations there: lock, remove, keySet, unlock. Can you reduce the number of distributed operations?

  • The customer object may not be that big, but can you not have to pass that object through the wire? Think about a scenario where you set order count to the customer object for fast access, so you should do a get and a put, and as a result, the customer object is passed through the wire twice.

Instead, why not move the computation over to the member (JVM) where your customer data resides. Here is how you can do this with distributed executor service.

  1. Send a PartitionAware Callable task.

  2. Callable does the deletion of the order right there and returns with the remaining order count.

  3. Upon completion of the Callable task, return the result (remaining order count). You do not have to wait until the task is completed; since distributed executions are asynchronous, you can do other things in the meantime.

Here is an example code.

HazelcastInstance hazelcastInstance = Hazelcast.newHazelcastInstance();

    public int removeOrder(long customerId, long orderId) throws Exception {
        IExecutorService executorService = hazelcastInstance.getExecutorService("ExecutorService");

        OrderDeletionTask task = new OrderDeletionTask(customerId, orderId);
        Future<Integer> future = executorService.submit(task);
        int remainingOrders = future.get();

        return remainingOrders;
    }

    public static class OrderDeletionTask
            implements Callable<Integer>, PartitionAware, Serializable, HazelcastInstanceAware {

        private long orderId;
        private long customerId;
        private HazelcastInstance hazelcastInstance;

        public OrderDeletionTask() {
        }

        public OrderDeletionTask(long customerId, long orderId) {
            this.customerId = customerId;
            this.orderId = orderId;
        }

        @Override
        public Integer call() {
            IMap<Long, Customer> customerMap = hazelcastInstance.getMap("customers");
            IMap<OrderKey, Order> orderMap = hazelcastInstance.getMap("orders");

            customerMap.lock(customerId);

            Predicate predicate = Predicates.equal("customerId", customerId);
            Set<OrderKey> orderKeys = orderMap.localKeySet(predicate);
            int orderCount = orderKeys.size();
            for (OrderKey key : orderKeys) {
                if (key.orderId == orderId) {
                    orderCount--;
                    orderMap.delete(key);
                }
            }

            customerMap.unlock(customerId);

            return orderCount;
        }

        @Override
        public Object getPartitionKey() {
            return customerId;
        }

        @Override
        public void setHazelcastInstance(HazelcastInstance hazelcastInstance) {
            this.hazelcastInstance = hazelcastInstance;
        }
    }

The following are the benefits of doing the same operation with distributed ExecutorService based on the key:

  • only one distributed execution (executorService.submit(task)), instead of four

  • less data is sent over the wire

  • less lock duration, i.e., higher concurrency, for the Customer entry since lock/update/unlock cycle is done locally (local to the customer data)

Partitioning Strategies

Another way of storing the related data on the same location is using/implementing the class PartitioningStrategy. Normally (if no partitioning strategy is defined), Hazelcast finds the partition of a key first by converting the object to binary and then by hashing this binary. If a partitioning strategy is defined, Hazelcast injects the key to the strategy and the strategy returns an object out of which the partition is calculated by hashing it.

Hazelcast offers the following out-of-the-box partitioning strategies:

  • DefaultPartitioningStrategy: Default strategy. It checks whether the key implements PartitionAware. If it implements, the object is converted to binary and then hashed, to find the partition of the key.

  • StringPartitioningStrategy: Works only for string keys. It uses the string after @ character as the partition ID. For example, if you have two keys ordergroup1@region1 and customergroup1@region1, both ordergroup1 and customergroup1 fall into the partition where region1 is located.

  • StringAndPartitionAwarePartitioningStrategy: Works as the combination of the above two strategies. If the key implements PartitionAware, it works like the DefaultPartitioningStrategy. If it is a string key, it works like the StringPartitioningStrategy.

You can configure the partitioning strategies:

  • for each map, or

  • globally (applied to all of the data structures in your cluster).

The per map and global partitioning strategies are supported on the member side. Hazelcast Java clients only support the global strategy.

Per Map Partitioning Strategy Configuration

The following shows example configurations. For the declarative configurations (XML, YAML), you use the partition-strategy element. For the programmatic approach, you use the setPartitioningStrategyClass() method.

  • XML

  • YAML

  • Java Member API

<hazelcast>
    ...
    <map name="myMap">
        <partition-strategy>
             com.hazelcast.partition.strategy.StringAndPartitionAwarePartitioningStrategy
             ///OR
             ///YourCustomPartitioningStrategyClass (1)
        </partition-strategy>
    </map>
    ...
</hazelcast>
hazelcast:
  map:
    myMap:
      partition-strategy: com.hazelcast.partition.strategy.StringAndPartitionAwarePartitioningStrategy
      # OR
      # partition-strategy: YourCustomPartitioningStrategyClass (1)
Config config = new Config();
MapConfig mapConfig = config.getMapConfig("myMap");
PartitioningStrategyConfig psConfig = mapConfig.getPartitioningStrategyConfig();
psConfig.setPartitioningStrategyClass( "StringAndPartitionAwarePartitioningStrategy" );

// OR
psConfig.setPartitioningStrategy(YourCustomPartitioningStrategy); (1)
...
1 You can define your own partition strategy by implementing the class PartitioningStrategy. To enable your implementation, add the full class name to your Hazelcast configuration using either the declarative or programmatic approach, as exemplified above.
All the cluster members must have the same partitioning strategy configurations.

Global Partitioning Strategy Configuration

You can also set a global strategy which is applied to all the data structures in your cluster. This can be done by setting the hazelcast.partitioning.strategy.class property.

The following shows example configurations.

  • XML

  • YAML

  • Java Member/Client API

<hazelcast>
    ...
    <properties>
        <property name="hazelcast.partitioning.strategy.class">
            com.hazelcast.partition.strategy.StringAndPartitionAwarePartitioningStrategy
        </property>
    </properties>
    ...
</hazelcast>
hazelcast:
  properties:
    hazelcast.partitioning.strategy.class: com.hazelcast.partition.strategy.StringAndPartitionAwarePartitioningStrategy
Config config = new Config();
config.setProperty( "hazelcast.partitioning.strategy.class", "com.hazelcast.partition.strategy.StringAndPartitionAwarePartitioningStrategy" );

You can specify the aforementioned out-of-the-box strategies or your custom partitioning strategy.

You can also use other system property configuring options as explained in the Configuring with System Properties section.

Attribute Based Partitioning Strategy

Attribute based partitioning (AttributePartitioningStrategy) is a strategy such that the map entries having the same partition keys live in the same cluster member. It can be configured per map only, and cannot be used globally.

To configure this strategy for a map, you can provide the partitioning-attributes configuration property under the map element. If this property is provided, it will automatically supersede partitioning-strategy (if any one of them is configured as explained in the previous sections) and the strategy will be set to AttributePartitioningStrategy. Note that, when provided, partitioning-attributes should contain at least one attribute name.

Dynamic configuration is supported, which means that you can add or change the configuration of this strategy without restarting the cluster.

It is important to note that AttributePartitioningStrategy is ONLY supported on the member-side.

Overall, AttributePartitioningStrategy is a powerful tool for partitioning data across a cluster in Hazelcast. By following the guidelines above, you can easily configure it for your specific use case.

Example Usage

Here is an example flow that shows how attribute based partitioning strategy works.

This example assumes that you have a map named myMap of type IMap<Person, Long> where Person has { public Long id; public String name; public Long orgId; }.

  1. Start a cluster by providing a map configuration that includes the attributes id and name (or you can add this configuration while a cluster is running).

    • XML

    • YAML

    <hazelcast>
    <map name="myMap">
        <partition-attributes>
            <attribute>id</attribute>
            <attribute>name</attribute>
        </partition-attributes>
        ...
    </map>
</hazelcast>
    hazelcast:
  map:
    myMap:
      partition-attributes:
        - name: "id"
        - name: "name"
      ...

  2. When data is added to the map using the put method, the strategy configured for this map is used to determine the partition ID of the new entry. Specifically, the AttributePartitioningStrategy extracts the id and name attributes from the given object, and creates an Object[] array out of them. This partition key is then used by the member logic to calculate the partition ID of the new entry.

    To give an example, assume that you have a key/value pair Person(1, "John", 3), 1, where id=1, name=John, and orgId=3.

    Add this map entry using myMap.put(new Person(1, "John", 3), 1);. The key of this entry (new Person(1, "John", 3)) will be passed to the configured strategy.

  3. The AttributePartitioningStrategy that is configured for this map takes the key and only extracts the id and name attributes from it (1 and John), and creates the {1, "John"} partition key.

  4. The member logic then calculates the partition ID out of the {1, "John"} partition key, and puts the entry to the cluster member which has the calculated partition ID.

  5. A new entry having the same partition key will be put to the same cluster member.

When you want to query for an entry with the key new Person(1, "John", 3), the exact same turn of events described above occurs, except the operation will be get instead of put.

lifebuoy

Help and support