CONFIDENTIAL
1
Praveen Kumar
Emerging Software Platforms,
Global Software Engineering
Mar 2014
Equinix Big Data Platform & Cassandra
Confidential – © 2013 Equinix Inc. www.equinix.com 2
Big Data at Equinix
~2 million
Alarms
~200k
interconnections
~250k
Electrical circuits
Sensors across 95+ IBXs
~40k
Infrastructure objects
Confidential – © 2013 Equinix Inc. www.equinix.com 3
Big Data at Equinix
Sensors across 95+ IBXs
Lead to / produce
Support for multiple protocols
Push as well pull methods
Time series data
Cross sectional dataNot so clean data
High velocity
Clean data Lots and lots of noise
Some useful intel
Confidential – © 2013 Equinix Inc. www.equinix.com 4
Big Data at Equinix
What do we use(or plan to use) this data for?
Customer Presentment Billing
Operations New Product & Services
Confidential – © 2013 Equinix Inc. www.equinix.com 5
Big Data at Equinix
Use-case analysis : 80-20 rule
~80% of use-cases analyzed act upon “Hot Data”
~80% of data for most of use-cases analyzed is time-series.
All “quick win” use-cases need data mediation, aggregation and roll-up for
presentment.
Real-time to near real-time processing of events
Collection, processing and storage technologies suitable for
time-series data.
Collection, mediation, cross-referencing and co-relation of
data from different sources; roll-up and aggregate.
Confidential – © 2013 Equinix Inc. www.equinix.com 6
Big Data at Equinix
Our Approach : Equinix Big Data Platform
§  Common platform to be shared by all initial Big
Data use cases – multi tenancy
§  Built on inexpensive hardware using free or
inexpensive software
§  Seamless & massive scalability using scale-out
§  High reliability - partial failover, graceful
degradation, self-healing, self-balancing
§  Data ingestion and processing capabilities for
high volumes at high velocity
§  Support for structured and semi-structured data
§  Provides real-time processing abilities
§  Provides parallel processing capabilities
§  Support for low latency queries, wide range
scan queries and search
§  Provides abstraction via connectors,
frameworks and libraries
§  Support for low latency queries, wide range
scan queries and search
§  Support for predictive analytics using machine
learning
Immediate requirements
Long term goals
Big Data Platform - Logical Architecture (technology agnostic)
Confidential – © 2013 Equinix Inc. www.equinix.com 7
Big Data at Equinix
Requirements & Technologies considered for Big Data Platform
Confidential – © 2013 Equinix Inc. www.equinix.com 8
Big Data at Equinix
Grand Finale
Hadoop Ecosystem vs. DataStax Enterprise
SearchSearch
SearchSearch
AnalyticsAnalytics
StorageStorageAnalyticsAnalytics
StorageStorage
StorageStorage
Hadoop	
  Distributed	
  File	
  System
(Storage/Analytics)
NameNode Secondary	
  Name	
  Node
Data	
  Nodes	
  (Storage)
HBase	
  (Storage/Analytics)
Hbase	
  Master
Hbase	
  Region	
  Servers
Hbase	
  Master
Search
Management	
  
Services
Cloudera	
  Manager
Solr	
  Nodes
Zookeeper
Pros
•  Scalability
•  Cloud readiness
•  Resource availability
•  Industry momentum
•  Product eco-system
maturity
•  Technical support
Cons
•  Infrastructure footprint
•  Operational Complexity
•  Learning curve
•  Availability
•  Total cost of ownership
Pros
•  Infrastructure footprint
•  Operational ease
•  Scalability
•  Availability
•  Cloud readiness
•  Learning curve
•  Resource availability
•  Technical support
•  Total cost of ownership
Cons
•  Industry momentum
•  Product eco-system
maturity
Confidential – © 2013 Equinix Inc. www.equinix.com 9
Criteria	
   Cassandra	
   HBase	
  
CAP Theorem Focus Availability, Partition-Tolerance Consistency, Availability
Data Partitioning
Supports ordered & random partitioning, random
partitioning is recommended.
Ordered Partitioning. Load balancing
achieved through resharding.
Distributed System P2P architecture (Amazon Dynamo)
Master / Slave via HDFS, Zookeeper for
coordination
Administration & Maintenance Medium High
Single Write Master No (R+W+1 to get Strong Consistency) Yes
Multi-tenancy Yes Yes
Secondary indexes
Supports secondary indexes on CF where column
name is known.
Does not natively support secondary indexes.
Consistency Tunable Consistency Strict consistency (Not ACID)
Hot Spot Problem
No, distributes load across nodes using random
partition strategy.
Yes, one node may handle most of the traffic
due to ordered partition.
Multi-Data Center Support
and Disaster Recovery
Asynchronous replication via WAN Asynchronous replication via WAN
Single point of failure Ring topology, there is no single point of failure.
Although there exists a concept of a master
server, HBase itself does not depend on it
heavily. HBase cluster can keep serving data
even if the master goes down. Hadoop
namenode is a single point of failure.
Commercial vendors Datastax, Acunu Clodera, Hortonworks
Cassandra Vs. HBase
Big Data at Equinix
Confidential – © 2013 Equinix Inc. www.equinix.com 10
Why DSE Cassandra
Big Data at Equinix
Support for Analytics
Integrated search using Solr
Security features
Cluster management capabilities
Commercial support
DataStax would probably list lots of more reasons, these are the reasons
relevant to us.
Confidential – © 2013 Equinix Inc. www.equinix.com 11
Big Data at Equinix
Grand Finale
Hadoop Ecosystem vs. DataStax Enterprise
SearchSearch
SearchSearch
AnalyticsAnalytics
StorageStorageAnalyticsAnalytics
StorageStorage
StorageStorage
Hadoop	
  Distributed	
  File	
  System
(Storage/Analytics)
NameNode Secondary	
  Name	
  Node
Data	
  Nodes	
  (Storage)
HBase	
  (Storage/Analytics)
Hbase	
  Master
Hbase	
  Region	
  Servers
Hbase	
  Master
Search
Management	
  
Services
Cloudera	
  Manager
Solr	
  Nodes
Zookeeper
Pros
•  Scalability
•  Cloud readiness
•  Resource availability
•  Industry momentum
•  Product eco-system
maturity
•  Technical support
Cons
•  Infrastructure footprint
•  Operational Complexity
•  Learning curve
•  Availability
•  Total cost of ownership
Pros
•  Infrastructure footprint
•  Operational ease
•  Scalability
•  Availability
•  Cloud readiness
•  Learning curve
•  Resource availability
•  Technical support
•  Total cost of ownership
Cons
•  Industry momentum
•  Product eco-system
maturity
ü Sold
Confidential – © 2013 Equinix Inc. www.equinix.com 12
Big Data at Equinix
How far are we on our Big Data journey?
ü  Pilot use-case from PoC to Production
ü  Moved network statistics use case from RRD
based solution to DSE Cassandra
ü  Build in progress for
§  power monitoring use cases
§  data center monitoring
§  network monitoring
In-plans
Ø  Recommendation engine on interconnection
platform
Ø  Use case analysis and technology selection for
connected data sets
Ø  Building data science capabilities for use cases
requiring predictive modeling
A few data points
Physical bare metal boxes for DSE
nodes
Densely packed data nodes with 4TB
storage on each node, 96GB RAM
About ~250 million records a day
Also used for log analysis for internal
IT systems monitoring use-cases
Confidential – © 2013 Equinix Inc. www.equinix.com 13
Big Data at Equinix
Experience so far
Lack of standards based connectors / drivers
DataStax has developed a Java Driver, but doesn’t support JDBC
No data visualization tools to access from Cassandra for low-latency access
No data access tools (Toad equivalent) available yet; DevCenter is not there yet
We
used Astyanax and are evaluating DataStax java driver
built libraries to abstract Astyanax for application engineering teams
built rest services for data access by applications
Good reliability
Not many instances of nodes being down
Handled loads even when nodes were down
Confidential – © 2013 Equinix Inc. www.equinix.com 14
Big Data at Equinix
Where do we go from here??
Graph databases
Batch processing (Hadoop, Spark , MapReduce ??)
Interactive queries
Online data processing
Data analytics
Data science and machine learning
Data visualization tools and applications
Developer toolkits
We are hiring
Big Data Architect
Big Data Engineers
Data Scientists
send resume at
pkumar@equinix.com
CONFIDENTIAL
15
Thank you!
•  pkmr.work@gmail.com
•  pkumar@equinix.com
•  www.equinix.com
EQUINIX?
Confidential – © 2013 Equinix Inc. www.equinix.com 17
WHO IS EQUINIX?
Confidential – © 2013 Equinix Inc. www.equinix.com 18
GLOBAL
DATA CENTERS
95+ Data Centers
9M+ Square Feet
99.999% Uptime Record
INTERCONNECTION
950+ Networks
110,000+ Cross Connects
BUSINESS
ECOSYSTEMS
Equinix Marketplace™
4,000+ Businesses
Revenue Opportunities
MOVING TOWARDS THE FUTURE | PLATFORM
Equinix: A Platform for Growth
Solid. Powerful. Growing.
$1.8B
IN ANNUALIZED
REVENUE
MEMBER OF THE NASDAQ 100
$7B
INVESTMENTS
IN EXPANSION
15 COUNTRIES
5 CONTINENTS
31 MARKETS
Confidential – © 2013 Equinix Inc. www.equinix.com 21
HOW WE’RE DIFFERENT | GLOBAL FOOTPRINT
Where You Are. Where You Need To Be.
90%
PASS THROUGH EQUINIX DATA CENTERS
OVER
OF INTERNET ROUTES
950+NETWORK PROVIDERS
450+
CLOUD & SaaS
PROVIDERS
CONFIDENTIAL
24
Thank you!
•  pkmr.work@gmail.com
•  pkumar@equinix.com
•  www.equinix.com

Cassandra Day SV 2014: Apache Cassandra at Equinix for High Performance, Scalability and Short Response Time

  • 1.
    CONFIDENTIAL 1 Praveen Kumar Emerging SoftwarePlatforms, Global Software Engineering Mar 2014 Equinix Big Data Platform & Cassandra
  • 2.
    Confidential – ©2013 Equinix Inc. www.equinix.com 2 Big Data at Equinix ~2 million Alarms ~200k interconnections ~250k Electrical circuits Sensors across 95+ IBXs ~40k Infrastructure objects
  • 3.
    Confidential – ©2013 Equinix Inc. www.equinix.com 3 Big Data at Equinix Sensors across 95+ IBXs Lead to / produce Support for multiple protocols Push as well pull methods Time series data Cross sectional dataNot so clean data High velocity Clean data Lots and lots of noise Some useful intel
  • 4.
    Confidential – ©2013 Equinix Inc. www.equinix.com 4 Big Data at Equinix What do we use(or plan to use) this data for? Customer Presentment Billing Operations New Product & Services
  • 5.
    Confidential – ©2013 Equinix Inc. www.equinix.com 5 Big Data at Equinix Use-case analysis : 80-20 rule ~80% of use-cases analyzed act upon “Hot Data” ~80% of data for most of use-cases analyzed is time-series. All “quick win” use-cases need data mediation, aggregation and roll-up for presentment. Real-time to near real-time processing of events Collection, processing and storage technologies suitable for time-series data. Collection, mediation, cross-referencing and co-relation of data from different sources; roll-up and aggregate.
  • 6.
    Confidential – ©2013 Equinix Inc. www.equinix.com 6 Big Data at Equinix Our Approach : Equinix Big Data Platform §  Common platform to be shared by all initial Big Data use cases – multi tenancy §  Built on inexpensive hardware using free or inexpensive software §  Seamless & massive scalability using scale-out §  High reliability - partial failover, graceful degradation, self-healing, self-balancing §  Data ingestion and processing capabilities for high volumes at high velocity §  Support for structured and semi-structured data §  Provides real-time processing abilities §  Provides parallel processing capabilities §  Support for low latency queries, wide range scan queries and search §  Provides abstraction via connectors, frameworks and libraries §  Support for low latency queries, wide range scan queries and search §  Support for predictive analytics using machine learning Immediate requirements Long term goals Big Data Platform - Logical Architecture (technology agnostic)
  • 7.
    Confidential – ©2013 Equinix Inc. www.equinix.com 7 Big Data at Equinix Requirements & Technologies considered for Big Data Platform
  • 8.
    Confidential – ©2013 Equinix Inc. www.equinix.com 8 Big Data at Equinix Grand Finale Hadoop Ecosystem vs. DataStax Enterprise SearchSearch SearchSearch AnalyticsAnalytics StorageStorageAnalyticsAnalytics StorageStorage StorageStorage Hadoop  Distributed  File  System (Storage/Analytics) NameNode Secondary  Name  Node Data  Nodes  (Storage) HBase  (Storage/Analytics) Hbase  Master Hbase  Region  Servers Hbase  Master Search Management   Services Cloudera  Manager Solr  Nodes Zookeeper Pros •  Scalability •  Cloud readiness •  Resource availability •  Industry momentum •  Product eco-system maturity •  Technical support Cons •  Infrastructure footprint •  Operational Complexity •  Learning curve •  Availability •  Total cost of ownership Pros •  Infrastructure footprint •  Operational ease •  Scalability •  Availability •  Cloud readiness •  Learning curve •  Resource availability •  Technical support •  Total cost of ownership Cons •  Industry momentum •  Product eco-system maturity
  • 9.
    Confidential – ©2013 Equinix Inc. www.equinix.com 9 Criteria   Cassandra   HBase   CAP Theorem Focus Availability, Partition-Tolerance Consistency, Availability Data Partitioning Supports ordered & random partitioning, random partitioning is recommended. Ordered Partitioning. Load balancing achieved through resharding. Distributed System P2P architecture (Amazon Dynamo) Master / Slave via HDFS, Zookeeper for coordination Administration & Maintenance Medium High Single Write Master No (R+W+1 to get Strong Consistency) Yes Multi-tenancy Yes Yes Secondary indexes Supports secondary indexes on CF where column name is known. Does not natively support secondary indexes. Consistency Tunable Consistency Strict consistency (Not ACID) Hot Spot Problem No, distributes load across nodes using random partition strategy. Yes, one node may handle most of the traffic due to ordered partition. Multi-Data Center Support and Disaster Recovery Asynchronous replication via WAN Asynchronous replication via WAN Single point of failure Ring topology, there is no single point of failure. Although there exists a concept of a master server, HBase itself does not depend on it heavily. HBase cluster can keep serving data even if the master goes down. Hadoop namenode is a single point of failure. Commercial vendors Datastax, Acunu Clodera, Hortonworks Cassandra Vs. HBase Big Data at Equinix
  • 10.
    Confidential – ©2013 Equinix Inc. www.equinix.com 10 Why DSE Cassandra Big Data at Equinix Support for Analytics Integrated search using Solr Security features Cluster management capabilities Commercial support DataStax would probably list lots of more reasons, these are the reasons relevant to us.
  • 11.
    Confidential – ©2013 Equinix Inc. www.equinix.com 11 Big Data at Equinix Grand Finale Hadoop Ecosystem vs. DataStax Enterprise SearchSearch SearchSearch AnalyticsAnalytics StorageStorageAnalyticsAnalytics StorageStorage StorageStorage Hadoop  Distributed  File  System (Storage/Analytics) NameNode Secondary  Name  Node Data  Nodes  (Storage) HBase  (Storage/Analytics) Hbase  Master Hbase  Region  Servers Hbase  Master Search Management   Services Cloudera  Manager Solr  Nodes Zookeeper Pros •  Scalability •  Cloud readiness •  Resource availability •  Industry momentum •  Product eco-system maturity •  Technical support Cons •  Infrastructure footprint •  Operational Complexity •  Learning curve •  Availability •  Total cost of ownership Pros •  Infrastructure footprint •  Operational ease •  Scalability •  Availability •  Cloud readiness •  Learning curve •  Resource availability •  Technical support •  Total cost of ownership Cons •  Industry momentum •  Product eco-system maturity ü Sold
  • 12.
    Confidential – ©2013 Equinix Inc. www.equinix.com 12 Big Data at Equinix How far are we on our Big Data journey? ü  Pilot use-case from PoC to Production ü  Moved network statistics use case from RRD based solution to DSE Cassandra ü  Build in progress for §  power monitoring use cases §  data center monitoring §  network monitoring In-plans Ø  Recommendation engine on interconnection platform Ø  Use case analysis and technology selection for connected data sets Ø  Building data science capabilities for use cases requiring predictive modeling A few data points Physical bare metal boxes for DSE nodes Densely packed data nodes with 4TB storage on each node, 96GB RAM About ~250 million records a day Also used for log analysis for internal IT systems monitoring use-cases
  • 13.
    Confidential – ©2013 Equinix Inc. www.equinix.com 13 Big Data at Equinix Experience so far Lack of standards based connectors / drivers DataStax has developed a Java Driver, but doesn’t support JDBC No data visualization tools to access from Cassandra for low-latency access No data access tools (Toad equivalent) available yet; DevCenter is not there yet We used Astyanax and are evaluating DataStax java driver built libraries to abstract Astyanax for application engineering teams built rest services for data access by applications Good reliability Not many instances of nodes being down Handled loads even when nodes were down
  • 14.
    Confidential – ©2013 Equinix Inc. www.equinix.com 14 Big Data at Equinix Where do we go from here?? Graph databases Batch processing (Hadoop, Spark , MapReduce ??) Interactive queries Online data processing Data analytics Data science and machine learning Data visualization tools and applications Developer toolkits We are hiring Big Data Architect Big Data Engineers Data Scientists send resume at pkumar@equinix.com
  • 15.
    CONFIDENTIAL 15 Thank you! •  pkmr.work@gmail.com • pkumar@equinix.com •  www.equinix.com
  • 16.
  • 17.
    Confidential – ©2013 Equinix Inc. www.equinix.com 17 WHO IS EQUINIX?
  • 18.
    Confidential – ©2013 Equinix Inc. www.equinix.com 18 GLOBAL DATA CENTERS 95+ Data Centers 9M+ Square Feet 99.999% Uptime Record INTERCONNECTION 950+ Networks 110,000+ Cross Connects BUSINESS ECOSYSTEMS Equinix Marketplace™ 4,000+ Businesses Revenue Opportunities MOVING TOWARDS THE FUTURE | PLATFORM Equinix: A Platform for Growth
  • 19.
    Solid. Powerful. Growing. $1.8B INANNUALIZED REVENUE MEMBER OF THE NASDAQ 100 $7B INVESTMENTS IN EXPANSION
  • 20.
  • 21.
    Confidential – ©2013 Equinix Inc. www.equinix.com 21 HOW WE’RE DIFFERENT | GLOBAL FOOTPRINT Where You Are. Where You Need To Be.
  • 22.
    90% PASS THROUGH EQUINIXDATA CENTERS OVER OF INTERNET ROUTES 950+NETWORK PROVIDERS
  • 23.
  • 24.
    CONFIDENTIAL 24 Thank you! •  pkmr.work@gmail.com • pkumar@equinix.com •  www.equinix.com