Hazelcast Jet - January 08, 2018

© 2018 Hazelcast Inc. Confidential & Proprietary
2
DISTRIBUTED COMPUTING. SIMPLIFIED.
Hazelcast Jet is a distributed
computing platform for fast processing
of big data sets. Jet is based on a
parallel core engine allowing data-
intensive applications to operate at
near real-time speeds.

Hazelcast Company Snapshot
Hazelcast founded
commercial open source
business model
Offices in Palo
Alto, London,
Istanbul
2008 2012 2013 2014 2015
JCache
released
Greg Luck
co-lead JCache
(JSR107)
Hazelcast introduces
JCache standard
for enterprise caching
40MM
Phone Homes per month
2016 2017
135 Subscribing
customers
Cash flow
positive
Gartner Market Guide for IMDG &
Leader in Forrester IMDG Wave
Report 2015
Office in BrnoHazelcast 1: First open
source IMDG
Hazelcast
Jet, fast Big
Data product
launched
Gartner Market
Guide for IMDG

Why Hazelcast Jet?
Performance – Low-latency at scale
Highly optimized reading from and writing to IMDG
Embeddable and light-weight

Hazelcast Jet Key Differentiators
PRODUCTIVE
ENTERPRISE GRADE STORAGEFLEXIBLE INTEGRATION
HIGH PERFORMANCE
• One solution provides both batch and stream
processing.
• Based on Hazelcast legendary ease-of-use.
• Variety of connectors enable easy integration into data
processing pipeline.
• Scaling, failure handling and recovery are all automated
for ease of operational management.
• Exactly-Once or At-Least Once guarantees.
• Jet team provides quick response to new feature
requests.
• Embeds Hazelcast IMDG as an operational storage for
enterprise-grade reliability and resilience.
• Most widely deployed IMDG with tens of thousands of
deployments in production worldwide.
• High-Performance integration eliminates network transit
latency penalties from moving data back and forth.
• Eliminates the integration issues between processing
and storage.
• Operates as shared application infrastructure or is
embedded directly in applications as a JAR- making
big data processing an application concern.
• Ideal for microservices with light-weight footprint --each
Jet processing job is launched within its own cluster
maximizing service isolation.
• Light-weight makes manipulation with Jet easy for both
developers and DevOps.
• Cloud and container ready. Pivotal Cloud Foundry,
OpenShift, Docker, Azure, AWS, Kubernetes, …
• Distributed computing platform enables fast processing
of big data.
• Parallel and low-latency core engine allows data-
intensive applications to operate at real-time speeds.
• Cooperative multi-threading architecture allows
operation of thousands of jobs simultaneously.
• Scales both vertically and horizontally.

Hazelcast Jet Key Competitive Differentiators
6
• High Performance | Industry Leading Performance
• Works great with Hazelcast IMDG | Source, Sink, Enrichment
• Very simple to program | Leverages existing standards
• Very simple to deploy | Embed 10MB jar or Client Server
• Works in every Cloud | Same as Hazelcast IMDG
• For Developers by Developers | Code it

Hazelcast Jet Product Overview
7

Hazelcast Jet Architecture
8

Hazelcast Jet Use Cases
Real-time Stream
processing
Distributed
In-Memory Computation
• Big Data in near real-time
• Plugged into data processing pipeline
• Distributed, in-memory computation
• Aggregating, joining multiple sources,
filtering, transforming, enriching
• Elastic scalability
• Super fast
• High availability
• Fault tolerant
• Simple, modern API
• Distributed, in-memory computation
• Transforming and querying
distributed data structures
• Computation close to the data
• Distributed clustering
• High availability
• Fault tolerance
Data-Processing
Microservices
• Data-processing microservices
• Isolation of services with many,
small clusters
• Service registry
• Network discovery
• Inter-process messaging
• Fully embeddable
• Spring Cloud, Boot Data Services

Jet Streaming Performance
*
* Spark had all performance options turn on including Tungsten

Performance - Throughput
*
* Spark had all performance options turn on including Tungsten

Performance - ForkJoin
• Jet is faster than JDK’s j.u.s implementation. The issue is in the JDK the
character stream reports "unknown size" and thus it cannot be parallelized
• If you first load the data into a List in RAM then run the JDK’s j.u.s it comes out a
little faster

Jet Application Deployment Options
IMDG IMDG
• No separate process to manage
• Great for microservices
• Great for OEM
• Simplest for Ops – nothing extra
Embedded
Application
Java API
Application
Java API
Application
Java API
Java Client
Application
Client-Server
Jet
Jet Jet
Jet
Jet Jet
Java Client
Application
Java Client
Application
Java Client
Application
• Separate Jet Cluster
• Scale Jet independent of applications
• Isolate Jet from application server lifecycle
• Managed by Ops

Jet with Hazelcast Deployment Choices
Jet Compute
Cluster
Hazelcast
IMDG Cluster
Sink Enrichment
Message
Broker
(Kafka)
Data
Enrichment
HDFS
Jet Cluster
Sink
Source /
Enrichment
Good when:
• Where source and sink are primarily Hazelcast
• Jet and Hazelcast have equivalent sizing needs
Good when:
• Where source and sink are primarily Hazelcast
• Where you want isolation of the Jet cluster

Stream Processing
• Support for events arriving out of order via Watermarks
• Sliding, Tumbling and Session window support

Fault Tolerance – Distributed State Snapshots
Exactly-Once, At-Least Once or No Guarantee to
optimize between performance and correctness
Distributed State Snapshots to back-up running
computations
Resilience with backups distributed and replicated
across the cluster to prevent losing data when member
fails
Simplicity as the snapshots are stored in embedded in-
memory structures. No further infrastructure is
necessary

Job Management & Fault Tolerance
• Job state and lifecycle saved to IMDG IMaps and benefit
from their performance, resilience, scale and persistence
• Automatic re-execution of part of the job in the event of a
failed worker
• Tolerant of loss of nodes, missing work will be recovered
from last snapshot and re-executed
• Cluster can be scaled without interrupting jobs – new jobs
benefit from the increased capacity
• State and snapshots can be persisted to resume after
cluster restart

Processing Guarantees
Guarantee Snapshots Performance
None No Fastest
At-Least Once Yes Slower
Exactly-Once Yes Slower

Data Input and Output
• Hazelcast Icache (JCache), (batch and streaming of
changes)
• Hazelcast IMap (batch and streaming of changes)
• Hazelcast IList (batch)
• HDFS (batch)
• Kafka (streaming)
• Socket (text encoding) (streaming)
• File (batch)
• FileWatcher (streaming – as new files appear)
• Custom, as sources and sinks are blocking Processors

Stream and Batch Processing
Maps, Caches
and Lists
Databases
IoT
Custom
Connector
Enterprise
Applications
Hazelcast IMDG
HDFS
Stream
Batch or
Stream
Batch or
Stream
Batch or
Stream
Batch
Ingest
Alerts
Enterprise
Applications
Interactive
Analytics
Hazelcast IMDG
Maps, Caches,
Lists
Output
Databases
Enrichment
Kafka
Stream
File /
File Watcher Batch
Socket
Stream
Kafka
HDFS
FileBatch
Database
Database
Events Stream
(via Striim CDC)

Business Use Cases

Field Equipment Monitoring and Analytics
Hazelcast Jet Enables Real-time Monitoring of Sensors and Processing of Streaming Data from
Field Equipment
Industries such as Oil & Gas use sensor data to automate decision making. Events reported from oil rig sensors are evaluated in
the field so that rig settings are adjusted in real time. Hazelcast Jet is ideal for applications that require near real-time insight.
Deployment simplicity makes it a optimal choice for both remote field and data center applications.
Why most processing solutions fall short… Why is Hazelcast Jet ideal?
Traditional data processing frameworks don’t allow direct embedding into
an application or independent system, forcing users to install them as a
secondary standalone server complicating deployment of the solution.
Jet can be used as an embedded library --avoiding need for separate server
tier. Jet doesn’t have any dependencies, so no further infrastructure
necessary to run a cluster. Jet allows users to build self-contained
applications.
Field data processing spans from sensors to on-rig processing, and from
there to processing within the data center --involving multiple execution
environments and hardware architectures. Traditional tools require
different runtimes (single-node, multi-node, different architectures),
resulting in higher system TCO.
Jet is a lightweight tool with a small memory footprint. There is just one
distribution to be used among all deployments. Running on a JVM, Jet can be
used on a variety of hardware platforms.
Traditional complex event processing solutions are built for single
machine/server deployments and cannot be scaled out easily.
Jet can be scaled out easily, as it’s built on top of a massively parallel,
distributed computing core (Hazelcast IMDG). It’s designed to scale out.
Most processing solutions experience growing latencies as the workload
throughput increases.
Jet is designed for low latency under the most grueling workloads with latency
remaining relatively flat even up to 5,000,000 aggregations per second.

Field Equipment Monitoring and Analytics Case
Study
CUSTOMER SUCCESS
Leading oil & gas system
integrator, specializing in
acquisition, persistence,
secure transportation and
dissemination of high-
frequency sensor data
10’s of 1000’s of events
per second are gathered
throughout the process
CHALLENGE
• Use sensor from oil wells data to automate decision-making:
- Keep latency low and consistent at scale, while system turns insights into decisions executed in a closed loop
- Simplify system deployment and maintenance in remote locations with limited bandwidth
- Use standards that extend system viability and maintenance lowering TCO
• Requires high application availability/performance for early issue detection to avoid production loss and
optimize well productivity
SOLUTION
• Hazelcast Jet is the processing backbone of application monitoring well sensors. API connects multiple
sensor data streams with varying formats and frequency. After joining and filtering, sliding and session
windows compute data insights to decisions
• Embedded Hazelcast IMDG is operational data store for easy scaling, whether on bare metal or in AWS
• Real-time analysis and correlation of event data from various devices and sensors throughout the process
allow immediate action. Jet adjusts rig settings in real time.
WHY HAZELCAST JET
• Performance and scalability
• In-memory data store with parallel processing enables scalable real-time analytics
• Open source, standards-based avoids vendor lock-in

Payment Processing
Hazelcast Jet Modernizes Payment Processing, Accelerating Bank Performance and Improving
Customer Experience
Hazelcast Jet is an ideal technology for payment processing systems as it offers the lowest end-to-end processing latencies on a
consistent basis. This is the result of Jet combining data processing and messaging in a single solution with elastically scalable
storage for caching data from 3rd party systems.
Why most processing solutions fall short… Why is Hazelcast Jet ideal?
Most processing solutions experience growing latencies as the workload
throughput increases.
Jet is designed for low latency under the most grueling workloads with latency
remaining relatively flat even up to 5,000,000 aggregations per second.
Most data processing solutions rely on 3rd party tools for messaging and
storage increasing deployment and maintenance complexity.
Jet embeds Hazelcast IMDG, which provides messaging and storage
capabilities in a single solution simplifying deployment and maintenance while
lowering TCO.
Traditional solutions for service orchestration are built for single
machine/server deployments and cannot be scaled out easily.
Jet scales easily, as it’s built on top of a massively parallel, distributed
computing core (Hazelcast IMDG). It’s designed to scale out.
The elasticity of a Jet cluster allows adding and removing computing and
storage resources in accordance with changing processing demand.

Payment Processing Case Study
CUSTOMER SUCCESS
A global information
technology solutions
company
Processing 10’s of 1,000’s
of payments per second
today.
Built-in scalability to
support future business.
CHALLENGE
• Before settling a transaction, payment processing systems check the merchant details by forwarding them to
the card’s issuing bank or association for verification, and carry out anti-fraud measures
• Each step in this pipeline requires the lowest possible latency to deliver a positive customer experience
• With 24/7 global operations and hard SLAs, resiliency and automatic recovery are a must-have
SOLUTION
• Within the payment processing application, Jet acts as the pipeline for each payment process step
• The payment management application orchestrates XML payment instructions and forwards them to the
respective card’s issuing bank or association for verification, then carries out anti-fraud measures before
settling transactions
• Multiple Jet processing jobs are pipeline components. Hazelcast IMDG distributed IMaps are used for
transaction ingestion and messaging
WHY HAZELCAST JET
• High-performance connectors between Jet and IMDG enable low-latency operations; consistent low latency
of the Hazelcast platform keeps the CGI payment management application within strictest SLA requirements
• Automatic recovery of the Jet cluster achieves high-availability even during failures
• Open source, standards-based avoids vendor lock-in

APIs

Pipeline API
• General purpose, declarative API: both powerful and simple to use
• Recommended as the best place to start using Jet
• Supports fork, join, cogroup, map, filter, flatmap, reduce, groupby
• Works with all sinks and sources
• Is a DSL which is put through a planner and converted to DAG plan
for execution
• Batch and Streaming (window support in 0.6)
See https://github.com/hazelcast/hazelcast-jet-code-samples

Pipeline API Code Sample
JetInstance jet = Jet.newJetInstance();
Pattern delimiter = Pattern.compile("W+");
Pipeline p = Pipeline.create();
p.drawFrom(Sources.<Long, String>readMap(BOOK_LINES))
.flatMap(e -> traverseArray(delimiter.split(e.getValue().toLowerCase()))) .filter(word ->
!word.isEmpty())
.groupBy(wholeItem(), counting())
.drainTo(Sinks.writeMap(COUNTS));
Job job = jet.newJob(p);
job.join();

Distributed java.util.stream API
• Jet adds distributed support for the java.util.stream API for
Hazelcast Map, List and Cache
• Supports all j.u.s. operations such as:
− map(), flatMap(), filter(), reduce(), collect(), sorted(), distinct()
• Lambda serialization is solved by creating Serializable versions of
the interfaces
• j.u.s streams are converted to Processor API (DAG) for execution
• Strictly a batch processing API
• Easiest place to start, but we recommend the Pipeline API to
exploit all features of Jet
See https://github.com/hazelcast/hazelcast-jet-code-samples

j.u.s API
JetInstance jet = Jet.newJetInstance();
Jet.newJetInstance();
IStreamMap<Long, String> lines = jet.getMap("lines");
Map<String, Long> counts = lines
.stream()
.flatMap(m -> Stream.of(PATTERN.split(m.getValue().toLowerCase())))
.filter(w -> !w.isEmpty())
.collect(DistributedCollectors.toIMap(”counts", w -> w,
w -> 1L, (left, right) -> left + right));

DAG API: Powerful, Low Level API
DAG describes how vertices are connected to each other:
DAG dag = new DAG();
// nil -> (docId, docName)
Vertex source = dag.newVertex("source", readMap(DOCID_NAME));
// (docId, docName) -> lines
Vertex docLines = dag.newVertex("doc-lines",
nonCooperative(flatMap((Entry<?, String> e) ->
traverseStream(docLines(e.getValue()))))
);
// line -> words
Vertex tokenize = dag.newVertex("tokenize",
flatMap((String line) -> traverseArray(delimiter.split(line.toLowerCase()))
.filter(word -> !word.isEmpty()))
);
// word -> (word, count)
Vertex accumulate = dag.newVertex("accumulate", accumulateByKey(wholeItem(),
AggregateOperations.counting()));
// (word, count) -> (word, count)
Vertex combine = dag.newVertex("combine", combineByKey(AggregateOperations.counting()));
// (word, count) -> nil
Vertex sink = dag.newVertex("sink", writeMap("counts"));
return dag.edge(between(source.localParallelism(1), docLines))
.edge(between(docLines.localParallelism(1), tokenize))
.edge(between(tokenize, accumulate).partitioned(wholeItem(), HASH_CODE))
.edge(between(accumulate, combine).distributed().partitioned(entryKey()))
.edge(between(combine, sink));

Building Custom Processors
• Unified API for sinks, sources and intermediate steps
• Not required to be thread safe
• Each Processor has an Inbox and Outbox per inbound and outbound
edge
• Two main methods to implement:
boolean tryProcess(int ordinal, Object item)
– Process incoming item and emit new items by populating the outbox
boolean complete()
– Called after all upstream processors are also completed. Typically used for
sources and batch operations such as group by and distinct
• Non-cooperative processors may block indefinitely
• Cooperative processors must respect Outbox when emitting and yield if
Outbox is already full

Hazelcast Jet Professional Support

Meet Your Business SLAs with
Hazelcast Jet Professional Support
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
FEATURE COMPARISON HAZELCAST JET PROFESSIONAL SUPPORT
24x7 Support Window
Service-Level Agreement 4 hours
Quarterly Review of Feature Requests P
Quarterly Review of the Hazelcast Jet Roadmap P
Email, IM, Phone Support Contacts 2
Hazelcast Jet Production Assurance for Initial Deployment/Charge
Control with Hazelcast Simulator P
Maintenance Patches P
Access to Hot Fix Patches P
IP Compliance Assurance with Black Duck ® Protex™ P
• Minimized downtime with 24/7 response to
support tickets, continuous delivery of
patches, and Hot Fix Patches
• Lower cost of operation of DevOps for your
Hazelcast Jet apps with simulation/
configuration testing
• Certified IP Compliance guarantees avoiding
legal exposure from open source

Hazelcast Jet Production Assurance for Initial
Deployment/Change Control with Hazelcast Simulator
Production simulation of your Hazelcast Jet
environment including hardware, virtualization,
operating system, configuration, feature usage,
topology, sizing and load
Includes assistance with ongoing simulation and
testing of your Hazelcast Jet platform and
deployed applications using Hazelcast Simulator
• Simulate the expected throughput/latency with
your specific requirements during pre-
production
• Test if Hazelcast Jet behaves as expected
when implementing new functionality, load
characteristics or expansion
• Standardize your application deployment
process with a battery of rigorous tests that we
will run for you before initial deployment and on
each upgrade

Hot Fix Patches
Personalized release of Hazelcast Jet software
dedicated to addressing your specific reported issue
Made to the exact version of Hazelcast Jet that you
are running with only the specific bug fix applied
Hot Fix Patches can thus be applied to your
production system with no or minimal testing
We reproduce the problem, and apply the patch in
our Simulator production simulation tool, to ensure we
go red before we go green

IP Compliance Assurance with
Black Duck® Protex™
• Full insight into Hazelcast
Jet IP compliance with
Black Duck® Protex™
reporting to help you
understand license
obligations, conflicts and
risks
• Black Duck Protex is the
leading solution for
managing open source
compliance
• Regular code scans before
minor and major releases
• We provide you a copy of
the report on demand

Hazelcast Jet Product Roadmap

Hazelcast Jet 0.6 Features (03/2018)
40
Features Description
Distributed Computation
Distributed data processing framework. It’s based on directed acyclic graphs (DAGs) and in-memory
computation
Robust stream processing
Fault tolerance based on distributed snapshots. When there is a failure (network failure or split, node
failure), Jet uses the latest state snapshot and automatically restarts
Ex-Once and At-least Once Jet can be configured to at-least-once or exactly-once semantics to favor performance or consistency
Windowing with Event-time
Improved streaming support including windowing support with event-time semantics. Out of the box
support for tumbling, sliding and session window aggregations
High-Level Pipeline API Convenient API to implement data processing pipeline
Pivotal Cloud Foundry Tile Hazelcast Jet is available in Pivotal Cloud Foundry
Spring Integration Integration with Spring Framework and Spring Boot
High-Performance Hazelcast
Integrations
Integrates with both embedded and non-embedded Hazelcast instances to be used for reading and
writing data for distributed computation. Change event reader allows streaming from Hazelcast IMDG
File and Socket Connector Ability to read and write unbounded data from text sockets and files
Kafka and HDFS Connector Reads and writes data streams to the Kafka message broker and Hadoop Distributed File System

Hazelcast 2018 Jet Roadmap
41
Features Description
Elasticity Dynamic rescaling of the computation
Diagnostics and management Management and monitoring features for Jet
More Connectors JMS | JDBC
1.0 Maturity To reach 1.0 maturity in 2018

Hazelcast Jet - January 08, 2018

More Related Content

What's hot (20)

Similar to Hazelcast Jet - January 08, 2018 (20)

Recently uploaded (20)

Hazelcast Jet - January 08, 2018

Editor's Notes