Infinispan Hands-On

Typical Applications of Infinispan

• Cache for databases
• Cache for other things
  • Servlet responses
  • Results of expensive computations
  • etc
• NoSQL key/value data store

JEE Application Example

• JBoss Application Server
• Persistence layer with JPA and Hibernate
• REST interface

What is First and Second Level caching in Hibernate?

First-level cache

Second-level cache

Real-Live Example in Three Steps

1. Example Project from Maven Archetype
2. Enable Database Logging
3. Enabling Infinispan L2 Cache

Maven Archetype

    mvn archetype:generate \
    -DarchetypeArtifactId=jboss-javaee6-webapp-archetype \
    -DarchetypeGroupId=org.jboss.spec.archetypes \
    -DarchetypeVersion=7.1.2.Final \
    -DgroupId=de.consol.research \
    -DartifactId=infinispan-jpa-example \
    -Dversion=1.0-SNAPSHOT \
    -Dname="Infinispan JPA Example" \
    -DarchetypeCatalog=http://search.maven.org/remotecontent?filepath=archetype-catalog.xml \
    -DinteractiveMode=false

    cd infinispan-jpa-example
    mvn clean verify jboss-as:run

Example Application

Example Application

Enable Database Logging

    SET TRACE_LEVEL_SYSTEM_OUT 2;

Enable 2nd Level Cache

    <shared-cache-mode>DISABLE_SELECTIVE</shared-cache-mode>
    <properties>
    <property name="hibernate.cache.use_second_level_cache" value="true"/>
    </properties>

Cache Modes

JBoss Management Console

Infinispan's Clustering Modes

• Invalidation: cluster-aware "local" caches that are able to invalidate entries if entries are modified elsewhere in the cluster.
• Replicated cache: Fully Replicated.
• Distributed cache: Consistent Hash.
• Additional (not on JBoss AS managment console): Client-Server mode

Replication Mode

Distribution Mode

• Read often served from remote node
• Writes sent to defined set of nodes
• Inexpensive to implement atomic operations
• Different strategies with joining andeaving nodes.

Client-Server Mode

JCACHE - JSR 107

JSR 107 Core Interfaces

• CachingProvider: Configure CacheManagers
• CacheManager: Acquire Cache
• Cache: ConcurrentMap + eviction + backed by distributed topology
• Entry: Key-value pair
• Expiry

JSR 107 Timetable 2013

Concurrent Map

private Cache<String, Integer> cache;

public void incrementUnsafe(String user) {
    if ( ! cache.containsKey(user) ) {
        cache.put(user, 0);
    }
    int current = cache.get(user);
    cache.put("user", current + 1);
}

Concurrent Map

private Cache<String, Integer> cache;

public synchronized void incrementThreadSafe(String user) {
    if ( ! cache.containsKey(user) ) {
        cache.put(user, 0);
    }
    int current = cache.get(user);
    cache.put("user", current + 1);
}

Concurrent Map

private Cache<String, Integer> cache;

public void incrementAtomic(String user) {
    cache.putIfAbsent(user, 0);
    int current;
    do {
        current = cache.get(user);
    }
    while (!cache.replace(user, current, current + 1));
}

Concurrent Map

Alternatively, caches provide
pessimistic lock / mutate / unlock.

Simple Distributed Hash

Distributed Hash with Back-up on "Neighbor to the Right"

Distributed Hash with Back-up on "Neighbor to the Right"

• Immediately after Node failure: Back-up node may become overloaded, because it gets all keys from failing neighbor.
• Rebalance: Nearly all keys need to be remapped when # nodes decreases / increases.

Distributed Hash with Secondary Back-up Hash

Distributed Hash with Secondary Back-up Hash

• Immediately after Node failure: Back-up node may become overloaded, because it gets all keys from failing neighbor.
• Rebalance: Nearly all keys need to be remapped when # nodes decreases / increases.

Consistent Hashing

Consistent Hashing

• On Node Failure: Load is distributed on #segments / #nodes nodes.
• Rebalance: Only #segments / #nodes segments need to be remapped.

• Default Num Segments for Infinispan is 60
• It does not make sense to configure num segments < num nodes, because then nodes will be empty.

Demo

Demo

Demo

• Num Keys: 600, Num Segments: 6.
• Keys per segment: ca 100
• Num Nodes 6: Everything evenly distributed (100 primary keys per Node)
• Num Nodes 5: One node has about 200 primary keys, the others have 100 primary keys.
• Num Nodes: 7: One node does not have any primary keys, the others have 100 primary keys.