Scaling Hibernate with Terracotta
Alex Miller (@puredanger)

Architecture and Scale

Applications in Three Tiers
Client
Server
Database

How do we scale?
Client Client Client
Server Server Server
Database

But what about failover?

Option 1: Servers share state
Client Client Client
Server Server Server
Database

How?
• RMI
• JMS
• Custom (JGroups, etc)

Option 2: Database (“stateless”)
Client Client Client
Server Server Server
Database

Stateless == “State in the database”
• Database load
• Serialization to and from the database
• Network bandwidth
• Object / relational mapping due to impedance mismatch

Option 3: Terracotta!
Client Client Client
Server Server Server
Terracotta Terracotta Terracotta
Terracotta
Terracotta
Database Server
Server
Instance
Instance

Why is this better?
• Hub and spoke vs peer to peer
• Avoid object / relational translation
• Avoid serialization
• Reduce database overload
• Programming model you already know
• Focus on your app, not scalability and high availability

What state should go in Terracotta?
Appropriateness Memory Terracotta Database
Data Lifetime

How does Terracotta work?

Core Concepts
• Roots
• Instrumented Classes
• Locks
• Integration Modules

Clustered Heap
App App app = new App();
app.run();

Clustered Heap
App EmployeeManager employeeMgr =
new EmployeeManager();
Employee
Manager
Object lock
Object lock = new Object();
Map<String,Employee> =
new HashMap<String,Employee>();
HashMap
employees

Clustered Heap
App
Employee
" U9 9 4 1 1 "
M anag er
Ob ject lock
Employee.AC T
IV E = = 1
Employee " Gandalf"
HashM ap
employees
Employee employee =
new Employee(
“U99411”,
“Gandalf”,
Employee.ACTIVE);

Clustered Heap
App employeeMgr.addEmployee(employee);
Employee
"U99411"
Manager
Object lock
Employee.ACT
IVE == 1
"U99411" Employee "Gandalf"
HashMap
employees
synchronized(lock) {
employees.put(e.getId(), e);
}

Clustered Heap
App
Employee
"U99411"
Manager
Object lock
Employee.ACT
IVE == 1
"U99411" Employee "Gandalf"
HashMap
employees
"U23526" Employee "U23526"
Employee.ACT
IVE = 1
"U99411"

Configuration
• Roots
• app.App.employeeMgr - a root of the clustered heap
• Instrumented classes
• app.* - instrument app classes at load time
• Locks
• employeeMgr.addEmployee() -> cluster synchronized blocks

Integration Modules
• Pre-built configuration
• Integration with specific 3rd-party libraries
• Hibernate
• Spring
• Ehcache
• Quartz
• Lucene
• many more...

Scaling Hibernate
• Hibernate Second Level Cache
• Clustered detached entities
• Close session, detach entities
• Cluster those entities in Terracotta
• Merge with new session later

Hibernate Caching
Application Thread Application Thread
Session Session 1st Level Cache
Cache Cache Cache
Concurrency Concurrency Concurrency
Hibernate
Strategy Strategy Strategy
CacheProvider 2nd Level Cache
Cache Cache Cache
Region Region Region
Database

Entity and collection caches
• Entity and collection cache regions
• Mark a Hibernate entity or a collection in an entity as @Cacheable
• Specify a cache concurrency strategy
• ReadOnly, ReadWrite, NonstrictReadWrite, Transactional
• Turn on second level caching in the Hibernate config

Query Cache
• Query cache regions
• Mark HQL, Criteria, Query as cacheable
• Store result set id values
• Timestamp cache region - last update time for each entity type
• Useful for caching natural key lookups (non-primary key)
• ...but lots of hidden issues

Terracotta Hibernate Second Level Cache
• Easy integration and configuration
• Supports entity, collection, and query cache regions
• Supports read-only, read-write, and nonstrict-read-write cache
concurrency strategies
• Hibernate-specific tooling
• High performance with cache coherency

Enabling Second Level Cache
• Mark your entities with a cache concurrency strategy
• In hibernate.cfg.xml: <cache usage="read-write"/>
• With annotations: @Cache(usage=CacheConcurrencyStrategy.READ_WRITE)
• hibernate.cfg.xml
• <property name="cache.use_second_level_cache">true</property>
• <property name="cache.provider_class">
org.terracotta.hibernate.TerracottaHibernateCacheProvider</property>

Enabling Second Level Cache
• Define the tc-hibernate-cache.xml in your classpath
<?xml version=”1.0” encoding=”UTF-8”?>
<terracotta-hibernate-cache-configuration>
<default-configuration>
<time-to-idle-seconds>7200</time-to-idle-seconds>
<time-to-live-seconds>7200</time-to-live-seconds>
</default-configuration>
<cache>
<region-name>org.terracotta.authinator.domain.Account</region-name>
<!-- as many region-names here as you want -->
<configuration>
<time-to-idle-seconds>600</time-to-idle-seconds>
<time-to-live-seconds>600</time-to-live-seconds>
</configuration>
</cache>
</terracotta-hibernate-cache-configuration>
• Add the Terracotta Hibernate cache provider jar to your classpath
• -cp terracotta-hibernate-cache-1.0.0.jar
• Add the Terracotta Hibernate cache agent jar to your command line
• -javaagent:terracotta-hibernate-agent-1.0.0.jar

Tooling Demo

Performance - Read-Only Comparison
Throughput Latency
200K 100
Transactions per second
80
150K
Avg Latency (ms)
60
100K
40
50K
20
0K 0
Database IMDG EhcacheTerracotta Database IMDG EhcacheTerracotta

How else is Terracotta used?
• Distributed cache
• Clustered HTTP sessions
• Batch processing
• Grid
• Messaging and events

Who Uses It?
• e-Commerce
• Online gaming
• Financial services
• Travel & leisure
• Social networking

Thanks!
• Terracotta Open Source JVM clustering:
• http://www.terracotta.org
• Apress: “The Definitive Guide to Terracotta”
• by Ari Zilka, Alex Miller, Geert Bevin, Jonas
Boner, Orion Letizi, Taylor Gautier
• Alex Miller
• @puredanger
• http://tech.puredanger.com