AWS re:Invent 2016: Analyzing Streaming Data in Real-time with Amazon Kinesis Analytics (BDM304)
[AWS LA Media & Entertainment Event 2015]: Raising the Bar on Video Streaming Quality with AWS
-
Upload
amazon-web-services -
Category
Technology
-
view
672 -
download
1
Transcript of [AWS LA Media & Entertainment Event 2015]: Raising the Bar on Video Streaming Quality with AWS
Raising the Bar on Video Streaming
Quality by Utilizing AWS
Amazon Video Case Study
Alex Dunlap, GM CloudFront
© 2015, Amazon Web Services, Inc. or its Affiliates. All rights reserved©
Min
ute
s V
iew
ed
pe
r C
us
tom
er
High Reliability =
No re-buffering
Low Reliability =
Frequent re-buffering
Why is QoS so important to the business?
Min
ute
s V
iew
ed
pe
r C
us
tom
er
High Reliability =
No re-buffering
Low Reliability =
Frequent re-buffering
Why is QoS so important to the business?
What to Expect from the Session
Learn what video quality means and how it’s measured
Understand how Amazon Video ensures customers get
the best possible video quality
See how AWS helps Amazon Video achieve their goals
Building workflows in the cloud means focusing
on functions that matter to customers.
OTT delivery functions
Ingest TranscodePublish /
StoreDeliver Analyze
Mezzanine Files
Contribution Feeds
Multi Protocol
Single Codec
Integration
Media Fragments
Manifests
Players
Derivatives
Recordings
Cache
Usage Analytics
Security Feedback
Prepare Assets for Multiscreen Deliver Assets and Measure
Combine multiple sources of data to get a full
picture
Origin CDN Players / Devices
• 4xx errors
• 5xx errors
• Requests per
minute
• Cache hit/miss
• Network
congestion (to
origin and to
player)
• Rebuffer events
• Fatal error events
• Average bitrate
• Time to first frame
AV: Overview
Amazon Video: watch
anytime, anywhere
Amazon:
a brand that
customers
trust
Subscription
service plus
buy or rent
Exclusive
& Original
content
Premium video: the core of
what we’re offering
AV: Overview
Tradeoffs
become
important
Reliability
Playback
start time
Video
quality
Priority of Tradeoffs
We invent new technologies to
push the boundary where these
tradeoffs must be made.
1
23
Premium video: the core of
what we’re offering
AV: Overview
Tradeoffs
become
important
Measuring
& improving
playback
Premium video: the core of
what we’re offering
AV: Overview
Tradeoffs
become
important
Measuring
& improving
playback
Customer
trust
Premium video: the core of
what we’re offering
AV: Overview
Tradeoffs
become
important
Measuring
& improving
playback
Customer
trust
We noticed that you experienced
poor video playback while
watching the following rental on
Amazon Video: Casablanca.
We’re sorry for the inconvenience
and have issued you a refund for
the following amount: $2.99. We
hope to see you again soon.
Challenges
Key challenge:Deliver premium video reliably over an inherently unreliable
network to hundreds of devices at various specs
Challenges
2009: The early days
Lack of
customer
metrics with
quantifiable
business
impact
Basic
handling of
degradation
and failure
scenarios
Many
devices,
each
worked
differently
Challenges
2009: The early days
Poor insight
into network
congestion
Low CDN cache
hit ratios
No feedback
loop
Measuring the
customer experience
Different devices,
different
implementations,
no simple plugin
Solutions
Solutions
Min
ute
s V
iew
ed
pe
r C
us
tom
er
High Reliability =
No re-buffering
Low Reliability =
Frequent re-buffering
Solutions
Improving the video player
Concurrent
downloads
to maximize
throughput
Adaptive
playback
CDN
switching
DNS failure
Client knows
to fall back to
other CDNs
Solutions
CDN Balancer
Automatic
prioritization
of traffic to
CDNs based
on playback
performance
Understand
performance
by streaming
protocol,
device, end-
user network,
geography,
and CDN
Observe recent
performance to
ensure a problem
experienced by
one customer is a
problem avoided
by others
Then > Now
Reduced the percent of
streams with playback
defects by 87% while traffic
has grown by 270x
Results:
2011 2012 2013 2014 2015
Am
azo
n V
ideo
Cu
sto
mers
Web Services (Purchasing, Catalog,
Vending, etc.)
Content Delivery
Telemetry Storage
Video Transcoding
using AWS
Content Origin
Storage
Amazon Video
Japan
Prime Video
Austria
Amazon Originals
Amazon S3
Amazon RedshiftAmazon EC2
Amazon SWFAmazon EC2
Prime Video
Japan
Prime Video
UK & Germany
CloudFront
Then > Now
AV’s Long Tail Demand Profile Presents a Challenge to
CDNs
More Popular Titles Less Popular Titles
% o
f S
essio
ns w
ith
ou
t err
ors
an
d
reb
uff
ers
Other CDNs
Why doesn’t the traditional CDN work for long tail media?
Limited storage
at the edge
Least recently
used eviction
policy
Media files are big
and getting bigger
Objects requested frequently
High cache hit ratios
Good playback experience
Objects requested infrequently
High cache miss ratios
Poor playback experience (latency, rebuffers)
Solution: Build Infrastructure Optimized for
Throughput and Storage
Interconnect
Fabric
Tier 1:
Transit Layer
Tier 2:
Caching Layer
Border
Transit / Peers
Interconnect
Fabric
Border
Transit / Peers
Large-
Object Store
Large-
Object Store
Large-
Object Store
Backbone
Benefits
Benefits
Direct peering
with major ISPs in
multiple internet
exchange
facilities
Improved
ThroughputBetter Caching Lower Costs
Petabytes of content
storage at the edge
Decouples storage
from network
capacity; scale each
separately as
demand warrants
Our relationship with AV also influenced other media
streaming specific improvements to CloudFrontIntelligent Pre-fetching
5 Mbps
2.5 Mbps
1 Mbps
512 Kbps
256 Kbps
Dynamic Manifest Support
CloudFront
Edge
Pre-fetch video fragments into cache at
the requested and adjacent bitrates to
reduce cache misses.
AV URL Vending
Service
Dynamic Manifest
Service
Media FragmentsMedia
Fragments
Built support for requesting a dynamically
generated manifest to optimize bitrate
availability based on device to improve
quality of playback.
.m3u8
Results: CloudFront winning traffic where we have
deployed new sites and raising the bar on quality
CloudFront Traffic GrowthCloudFront QoS Improvements
we
ek 1
we
ek 3
we
ek 5
we
ek 7
we
ek 9
we
ek 1
1
we
ek 1
3
we
ek 1
5
we
ek 1
7
we
ek 1
9
we
ek 2
1
we
ek 2
3
we
ek 2
5
we
ek 2
7
we
ek 2
9
we
ek 3
1
we
ek 3
3
we
ek 3
5
we
ek 3
7
Peak T
hro
ug
hp
ut
(Gb
ps
)
Peak Throughput of Amazon Video served via CloudFront
we
ek 1
we
ek 3
we
ek 5
we
ek 7
we
ek 9
we
ek 1
1
we
ek 1
3
we
ek 1
5
we
ek 1
7
we
ek 1
9
we
ek 2
1
we
ek 2
3
we
ek 2
5
we
ek 2
7
we
ek 2
9
we
ek 3
1
we
ek 3
3
we
ek 3
5
we
ek 3
7
% o
f S
essio
ns w
ith
ou
t err
ors
an
d
reb
uff
ers
Quality of Service of Amazon Video served via CloudFront
Results: Dramatically improve performance on long tail
portion of content
More Popular Titles Less Popular Titles
% o
f S
essio
ns w
ith
ou
t err
ors
an
d
reb
uff
ers
CloudFront Other CDNs