Think Distributed: The Hazelcast Way
- 2. About Me
Senior Solutions Architect
Worked with Terracotta
In-memory Distributed Systems since 2009
Java Programmer since 1998
rahul@hazelcast.com
Rahul Gupta
follow me @wildnez
- 4. Why Hazelcast?
• Scale-out Computing enables cluster capacity to be increased
or decreased on-demand
• Resilience with automatic recovery from member failures
without losing data while minimizing performance impact on
running applications
• Programming Model provides a way for developers to easily
program a cluster application as if it is a single process
• Fast Application Performance enables very large data sets to be
held in main memory for real-time performance
- 10. Distributed Map
import java.util.concurrent.ConcurrentMap;
import com.hazelcast.core.Hazelcast;
import com.hazelcast.core.HazelcastInstance;
public static void main(String[] args) {
HazelcastInstance h = Hazelcast.newHazelcastInstance();
ConcurrentMap<Integer, String> map = h.getMap("myMap");
map.put(1, "Paris");
map.put(2, "London");
map.put(3, "San Francisco");
String oldValue = map.remove(2);
}
- 12. Persistence API
public class MapStorage
implements MapStore<String, User>, MapLoader<String, User> {
// Some methods missing ...
@Override public User load(String key) { return loadValueDB(key); }
@Override public Set<String> loadAllKeys() { return loadKeysDB(); }
@Override public void delete(String key) { deleteDB(key); }
@Override public void store(String key, User value) {
storeToDatabase(key, value);
}
}
<map name="users">
<map-store enabled="true">
<class-name>com.hazelcast.example.MapStorage</class-name>
<write-delay-seconds>0</write-delay-seconds>
</map-store>
</map>
- 13. JCache API
// Retrieve the CachingProvider which is automatically baced by
// the chosen Hazelcast server or client provider
CachingProvider cachingProvider = Caching.getCachingProvider();
// Create a CacheManager
CacheManager cacheManager = cachingProvider.getCacheManager();
// Cache<String, String> cache = cacheManager
// .getCache( name, String.class, String.class );
// Create a simple but typesafe configuration for the cache
CompleteConfiguration<String, String> config =
new MutableConfiguration<String, String>()
.setTypes( String.class, String.class );
- 14. JCache API
// Create and get the cache
Cache<String, String> cache = cacheManager
.createCache( "example", config );
// Alternatively to request an already existing cache
Cache<String, String> cache = cacheManager
.getCache( name, String.class, String.class );
// Put a value into the cache
cache.put( "world", "Hello World" );
// Retrieve the value again from the cache
String value = cache.get( "world" );
System.out.println( value );
- 15. Eviction
• Unless deleted, entries remain in the map.
• Use eviction policies to prevent OOM situations.
• Eviction Triggers run LFU or LRU.
time-to-live
max-idle-seconds
max-size (PER_NODE(number of entries),
PER_PARTITION(number of entries),
USED_HEAP_SIZE(mb),
USED_HEAP_PERCENTAGE(mb))
• Setting eviction-percentage removes that % of entries when eviction is
triggered.
- 16. Features Description
MultiMap Store multiple values against one Key
Replicated Map Cluster wide replication, all entries on all nodes. Good for read heavy use
cases
Near Cache Map Entries on Client/Application local memory
RingBuffer Stores data in a ring-like structure, like a circular array with given capacity
More Distributed Structures
- 17. Java Collection API: Map, List, Set, Queue
Jcache
High Density Memory Store
Hibernate 2nd Level Cache
Web Session Replication: Tomcat, Jetty
Predicate API: Indexes, SQL Query
Persistence: Map/Queue Store & Loader. Write Behind/Through
Spring Compliance
Transactions: Local & XA
WAN & DR Replication
IM Data Store (Caching) Features
- 18. Java: Will it make the cut?
Garbage Collection limits heap usage. G1 and
Balanced aim for <100ms at 10GB.
Unused
Memory
64GB
4GB 4s
Heap
Java Apps Memory Bound
GC Pause
Time
Available
Memory
GC
Off-Heap Storage
No low-level CPU access
Java is challenged as an infrastructure language
despite its newly popular usage for this
Heap
- 19. GC Pause
64 Gb
JVM
4 Gb JVM 1
4 Gb JVM 2
4 Gb JVM 3
4 Gb JVM 4
4 Gb JVM 5
4 Gb JVM 16
GC Pause
0 GB
64 GB
64 Gb in-memory data
time
Standard Impediments of Caching
- 20. GC Pause
64 Gb
JVM
4 Gb JVM 1
4 Gb JVM 2
4 Gb JVM 3
4 Gb JVM 4
4 Gb JVM 5
4 Gb JVM 16
GC Pause
0 Gb
64 Gb
64 Gb in-memory data
time
64 Gb
HD
2 Gb JVM
GC Pause
HD Memory
Caching with HD Memory
- 21. On-Heap Memory Store
(Objects Stored as Objects)
High-Density Memory Store
(Objects Serialized and Stored as
Bytes)
s.m.Unsafe s.m.Unsafe
2-4GB
(Limited by Garbage
Collection)
0-1TB
(Limited by
Machine RAM)
Memory Stores
•Member
•Client
(Near Cache)
RAM in JVM Process
APIs
JCache
(ICache)
Map
(IMap)
HD Memory
- 22. On Heap Vs. High-Density
Memory Management
On Heap Memory HD Memory
0 MB HD 3.3 GB
3.9 GB Heap Storage 0.6 GB
9 (4900 ms) Major GC 0 (0 ms)
31 (4200 ms) Minor GC 356 (349 ms)
Node Used
Heap
Total
Heap
Max.
Heap
Heap Usage
Percentage
Used Heap: 0.2 GB
192.168.1.10:5701 57 MB 229 MB 910 MB 6.28%
Memory Utilization
Home Offheap-test
Node Used Heap: 3.9 GB
192.168.1.10:5701 3933 MB 4658MB 4653MB 84.45%
Memory Utilization
Home
Used
Heap
Total
Heap
Max.
Heap
Heap Usage
Percentage
Example: On Heap Memory Example: HD Memory
- 23. Hazelcast Servers
Hazelcast Server
JVM [Memory]
A B C
Business Logic
Data Data Data
CE = Compute Engine
Result
Business / Processing Logic
Result
TCP / IP
Client Client
Distributed Computing
- 24. IM Distributed Computing Feature
HD
Cache
Dist.
Compute
Dist.
Message
Java Concurrency API
(Lock, Semaphore, AtomicLong, AtomicReference, Executor Service, Blocking Queue)
Entry and Item Listeners
Entry Processor
Aggregators
Map/Reduce
Data Affinity
Continues Query
Map Interceptors
Delta Update
- 25. Executor Service API
public interface com.hazelcast.core.IExecutorService
extends java.util.concurrent.ExecutorService
HazelcastInstance hz = getHazelcastInstance();
//java.util.concurrent.ExecutorService implementation
IExecutorService es = hz.getExecutorService("name");
es.executeOnAllMembers(buildRunnable());
es.executeOnKeyOwner(buildRunnable(), "Peter");
es.execute(buildRunnable());
Map<..> futures = es.submitToAllMembers(buildCallable());
Future<..> future = es.submitToKeyOwner(buildCallable(), "Peter");
es.submitToAllMembers(buildCallable(), buildCallback());
es.submitToKeyOwner(buildCallable(), "Peter", buildCallback());
- 27. Lock API
HazelcastInstance hz = getHazelcastInstance();
// Distributed Reentrant
Lock lock = hz.getLock("myLock");
lock.lock();
try {
// Do something
} finally {
lock.unlock();
}
Distributed Lock
- 30. Map/Reduce API
HazelcastInstance hz = getHazelcastInstance();
Map users = hz.getMap("users");
JobTracker tracker = hz.getJobTracker("default");
KeyValueSource source = KeyValueSource.fromMap(users);
Job job = tracker.newJob(source);
ICompleteFuture future = job.mapper(new MyMapper())
.reducer(new MyReducer())
.submit();
Map result = future.get();
- 31. Aggregations API
HazelcastInstance hz = getHazelcastInstance();
Map users = hz.getMap("users");
int sum = users.aggregate(
Supplier.all((user) -> user.getSalary()),
Aggregations.longSum()
);
- 34. Queue API
interface com.hazelcast.core.IQueue<E>
extends java.util.concurrent.BlockingQueue
HazelcastInstance hz = getHazelcastInstance();
//java.util.concurrent.BlockingQueue implementation
IQueue<Task> queue = hz.getQueue("tasks");
queue.offer(newTask());
queue.offer(newTask(), 500, TimeUnit.MILLISECONDS);
Task task = queue.poll();
Task task = queue.poll(100, TimeUnit.MILLISECONDS);
Task task = queue.take();
- 35. Topic API
public class Example implements MessageListener<String> {
public void sendMessage {
HazelcastInstance hz = getHazelcastInstance();
ITopic<String> topic = hz.getTopic("topic");
topic.addMessageListener(this);
topic.publish("Hello World");
}
@Override
public void onMessage(Message<String> message) {
System.out.println("Got message: " + message.getMessageObject());
}
}
- 38. Distributed Maps
Fixed number of partitions (default 271)
Each key falls into a partition
partitionId = hash(keyData)%PARTITION_COUNT
Partition ownerships are reassigned upon membership change
A B C
- 59. Deployment Options
Great for early stages of rapid
application development and iteration
Necessary for scale up or scale out
deployments – decouples
upgrading of clients and cluster for
long term TCO
Embedded Hazelcast
Hazelcast Node
1
Applications
Java API
Client-Server Mode
Hazelcast
Node 3
Java API
Applications
Java API
Applications
Java API
Applications
Hazelcast
Node 2
Hazelcast
Node 1
Hazelcast Node
2
Applications
Java API
Hazelcast Node
3
Applications
Java API
- 60. Easy API
// Creating a new Hazelcast node
HazelcastInstance hz = Hazelcast.newHazelcastInstance();
// Getting a Map, Queue, Topic, ...
Map map = hz.getMap("my-map");
Queue queue = hz.getQueue("my-queue");
ITopic topic = hz.getTopic("my-topic");
//Creating a Hazelcast Client
HazelcastInstance client = HazelcastClient.newHazelcastClient();
// Shutting down the node
hz.shutdown();
- 62. Hazelcast High Level Roadmap
Hi-Density Caching
In-Memory Data Grid
2014 2015 2016
HD Memory | Advance Messaging
PaaS | Extensions | Integrations | JET
Scalability | Resiliency | Elastic Memory | In-Memory Computing
Advance In-memory Computing Platform
- 64. © 2016 Hazelcast Inc. Confidential & Proprietary 64
Features Description
Modularity In 3.7, Hazelcast is converted to a modular system based around extension points. So clients,
Cloud Discovery providers and integrations to third party systems like Hibernate etc will be
released independently. 3.7 will then ship with the latest stable versions of each.
Redesign of Partition Migration More robust partition migration to round out some edge cases.
Graceful Shutdown
Improvements
More robust shutdown with partition migration on shutdown of a member
Higher Networking Performance A further 30% improvement in performance across the cluster by eliminating notifyAll() calls.
Map.putAll() Performance
Speedup
Implement member batching.
Rule Based Query Optimizer Make queries significantly faster by using static transformations of queries.
Azul Certification Run Hazelcast on Azul Zing for Java 6, 7 or 8 for less variation of latencies due to GC.
Solaris Sparc Support Align HD Memory backed data structure's layouts so that platforms, such as SPARC work.
Verify SPARC using our lab machine.
New Features for JCache Simple creation similar to other Hazelcast Data Structures. E.g.
Command Line Interface New command line interface for common operations performed by Operations.
Non-blocking Vert.x integration New async methods in Map and integration with Vert.x to use them.
New Hazelcast 3.7 Features
- 65. © 2016 Hazelcast Inc. Confidential & Proprietary 65
Features Description
Scala integration for Hazelcast members and Hazelcast client. Implements all Hazelcast
features. Wraps the Java client for client mode and in embedded mode uses the Hazelcast
member directly.
Node.js Native client implementation using the Hazelcast Open Client protocol. Basic feature support.
Python Native client implementation using the Hazelcast Open Client protocol. Supports most Hazelcast
features.
Clojure Clojure integration for Hazelcast members and Hazelcast client. Implements some Hazelcast
features. Wraps the Java client for client mode and in embedded mode uses the Hazelcast
member directly.
New Hazelcast 3.7 Clients and Languages
- 66. © 2016 Hazelcast Inc. Confidential & Proprietary 66
Features Description
Azure Marketplace Ability to start Hazelcast instances on Docker environments easily. Provides Hazelcast,
Hazelcast Enterprise and Management Center.
Azure Cloud Provider Discover Provider for member discovery using Kubernetes. (Plugin)
AWS Marketplace Deploy Hazelcast, Hazelcast Management Center and Hazelcast Enterprise clusters straight
from the Marketplace.
Consul Cloud Provider Discover Provider for member discovery for Consul (Plugin)
Etcd Cloud Provider Discover Provider for member discovery for Etcd (Plugin)
Zookeeper Cloud Provider Discover Provider for member discovery for Zookeeper (Plugin)
Eureka Cloud Provider Discover Provider for member discovery for Eureka 1 from Netflix. (Plugin)
Docker Enhancements Docker support for cloud provider plugins
New Hazelcast 3.7 Cloud Features
- 69. Features Description
Amazon EC2 EC2 Auto discovery – upgraded with Discovery SPI
Microsoft Azure Available on Azure Marketplace
Pivotal Cloud Foundry Only distributed IMDG to provide on-demand service broker and disk
based persistence
OpenShift Native compliancy
Hazelcast on Cloud – IaaS, PaaS
- 70. Hazelcast on Cloud – SaaS, IaaS, PaaS
Other off-the-shelf cloud based compliancy
• OpenStack
• Google Compute Engine
• Google Platform Services
• jClouds
• Discovery SPI – Everything Everywhere
- 72. What’s Hazelcast Jet?
• General purpose distributed data processing
framework
• Based on Direct Acyclic Graph to model data
flow
• Built on top of Hazelcast
• Comparable to Apache Spark or Apache Flink
7
2
- 76. Service Offerings
Hazelcast (Apache Licensed)
• Professional Subscription – 24x7 support*
Hazelcast Enterprise Support
• Available with Hazelcast Enterprise software subscription - 24x7 support*
Additional Services
• Development Support Subscription – 8x5 support*
• Simulator TCK
• Training
• Expert Consulting
• Development Partner Program
* All subscriptions include Management Center
- 77. © 2016 Hazelcast Inc. Confidential & Proprietary 77
Support Subscriptions
What’s Included
100% SUCCESS RATE ON CUSTOMER ISSUES:
“As usual, the response was timely beyond
expectations, and very good technical content
returned. Exemplary support, hard to find in
any company…”
- Fortune 100 Financial Services Customer
ENTERPRISE HD ENTERPRISE PROFESSIONAL OPEN SOURCE
SUPPORT WINDOW 24/7 24/7 24/7
RESPONSE TIME FOR CRITIAL ISSUES 1 Hour 1 Hour 2 Hours
SUPPORTED SOFTWARE
Hazelcast &
Hazelcast Enterprise
Hazelcast &
Hazelcast Enterprise
Hazelcast
SUPPORT CONTACTS 4 4 2
SUPPORT CHANNELS Email, IM & Phone Email, IM & Phone Email, IM & Phone
PATCH LEVEL FIXES
REMOTE MEETINGS (via GoToMeeting)
CODE REVIEW (with a Senior Solutions Architect) 2 Hours 2 Hours 2 Hours
QUARTERLY REVIEW OF FEATURE REQUES*
QUARTERLY REVIEW OF HAZELCAST ROADMAP*
- 78. © 2016 Hazelcast Inc. Confidential & Proprietary
7
8
Best In Class Support
Support from the Engineers who wrote
the code
SLA Driven – 100% attainment of
support response time
Follow the Sun
Portal, Email and Phone access
Go Red, Go Green. Reproduction of
issues on Simulator. Proof of fix on
Simulator.
Periodic Technical Reviews
Meet your production schedule and
corporate compliance requirements
Ensure the success of your
development team with training and
best practices
- 81. © 2016 Hazelcast Inc. Confidential & Proprietary 81
Release Lifecycle
• Regular Feature release each 4-5 months, e.g. 3.3, 3.4, 3.5
• Maintenance release approximately each month with bug fixes based
on the current feature release, e.g. 3.4.1
• For older versions, patch releases made available to fix issues
• Release End of Life per support contract