netflix cloud architecture...continued
TRANSCRIPT
Incumbents in 1997 ¨ Blockbuster, Hollywood
Videos, Movie Gallery ¤ Rented DVDs out from
prime locations ¤ Kept inventory of about
2000 titles with 3 to 4 copies of each
¤ Mostly expensive DVDs with new movies and some long time hits
¨ New movies typically release to big screen and then release via DVD
¨ DVDs rent while its new ¨ DVD availability window
is relatively small ¨ Charged late fee to
maximize turns
3
Must impose steep rental fees and steeper late fees
The Red Envelope (1999-2008) ¨ Removed late fee as they
saw opportunity in long tail of DVDs
¨ Stocked 120,000 movies mailed out from 50 hubs
¨ Matched renter interest with a low cost movie from the long tail
¨ Announced $1M prize: Help find a better recommendation algorithm
¨ 4 years and a team of 10+ PhDs solved that
¨ In 8 years, Blockbuster, Hollywood video et. Al. are almost out of business
4
Recommend movies in the long tail, personalized to the customer, lowering costs
Streaming Movies – 2008 and beyond ¨ Netflix lets users manage their DVD
queue for what they would rent next ¨ Netflix had their own problems
¤ DVD theft ¤ DVD piracy ¤ FBI involved in the US ¤ Could not grow internationally with DVD
renting model
¨ Streaming model looked very appealing ¨ Cost to stream out DVD via CDN was a
few pennies but postage is 10-100 times more
¨ Streaming lets them grow internationally
¨ 20M US subscribers ¨ Started streaming to US subscribers
from their datacenter ¨ Datacenter faced cooling issues and
downtime ¨ They issued customer refunds which
directly impacted their bottom line ¨ Netflix at that time was making
about 1.6 billion in revenue and had 400 employees
5
Netflix Decisions 6
¨ Who will manage the streaming backend? ¨ What devices will users watch streaming content on? ¨ On the client end
¤ What should be the client experience?
¤ Should Netflix create their own device?
¤ Should Netflix partner with device makers?
¤ If so, which device makers should Netflix pursue?
¤ What about DRM to prevent piracy and honor IP?
¨ On the cloud end ¤ Who will manage the streaming backend?
¤ Who will maintain and upkeep the data center?
¤ How to deal with data center growth?
Decided against creating own device 9
¨ HDMI ports on household TVs were already occupied by a variety of devices
¨ Introducing one more Netflix device wasn’t the best
Partnerships with Device makers 10
¨ Grew subscribers 2008-2010 by partnering with device makers ¤ Game Consoles
n Xbox 360, Wii, PS3
¤ Mobile Devices n Apple iPad, iPhone, iPod Touch, Windows Phone 7 n Android was not in this list until late 2011 (DRM concerns)
Netflix Cloud Story 12
¨ Circa 2008, Netflix had a single data center ¤ Single point of failure (SPOF) ¤ Approaching the limits on
n Cooling n Power n Physical space n Network traffic capacity
Netflix Cloud Story 13
¨ Strategic choices included: ¤ Build more data centers
n Manage more data centers
¤ Outsource capacity planning and scale out n Focus on core competency
Netflix Cloud Story 14
¨ Winning Choice: ¤ Outsource capacity planning and scale out
n Leverage a leading IAAS provider n Amazon Web Services
¨ In two years, from start of outsourcing, Netflix served over 90% traffic out of AWS
¨ This excluded video streaming traffic which came from CDNs
How Netflix Uses AWS 15
¨ AWS provided various IAAS offering at that time but Netflix apps needed more
• Data SimpleDB,RDS,S3, EBS
• Load Handling EC2, EMR, ASG, ELB
• Monitoring CloudWatch
AWS ? Netflix apps
Netflix developers filled the gap 16
¨ Netflix developers ported Netflix platform to the AWS
AWS Netflix Platform
Netflix App
Netflix Platform • Platform.JAR • Middle-tier Load Balancing • Discovery • Encryption Services • Caching • Configuration
How Netflix Used AWS 17
ELB
EC2 EC2 EC2 EC2 EC2
EC2 instances for Q EC2 instances for merch
Simple DB S3 SQS
Edge services (such as www)
Netflix mid-tier services (such as Q or merchandizing)
AWS IAAS
?
How Netflix Used AWS 18
ELB
EC2 EC2 EC2 EC2 EC2
Middle Tier Load Balancer and Discovery
EC2 instances for Q EC2 instances for merch
Simple DB S3 SQS
Edge services (such as www)
Netflix mid-tier services (such as Q or merchandizing)
AWS IAAS
Netflix platform
Netflix stack on EC2 19
Netflix Application
Platform and other infrastructure jar files
Guest OS/Virtual Machine/Host OS/Machine
Discovery client
Mid tier load balancer
Crypto Client
S3 client Simple DB client SQS client
Configuration client
Persistent Data: Simple DB 20
¨ Moved large transactional dataset to Simple DB ¤ Rental history info: Everything that any one has watched
on Netflix (including streaming and DVD rentals) ¨ Netflix uses Simple DB to manage
¤ Thousands of domains ¤ Over 1Tera byte of online transaction data ¤ Billions of rows of data
Persistent Storage: S3 21
¨ Simple key-value store organized as objects in buckets ¤ Each AWS account can create multiple buckets ¤ Each bucket can hold unlimited number of objects
¨ Netflix uses S3 to store data that does not fit into Simple DB ¤ Logs from streaming devices ¤ Files used in movie encoding ¤ Truncated tail of rental history
Data Replication: Item Replicator 22
¨ IR: Item Replicator: Netflix homegrown solution ¨ Keeps data in sync between Data Center and Cloud
¤ Replication is mostly unidirectional n Data Center to the Cloud
Netflix Migration to Cloud 23
¨ Netflix case is just one example. ¨ There are several cases of businesses in a variety of
industry that are migrating to the cloud