# AWS Anaheim Keynote 2018

# Keynote Speaker: Stephan Orban

• General Manager, AWS

# Speaker: Stephen Felisan, CIO, Edmunds

• 20m car shoppers ech month
• Pivoted in 1 year
• Culture for speed
• Data strategy
• Embrace serverless

# Data Strategy

• S3 as data layer for everything
• Sensitive data secured in S3

# Serverless

• Banquet image processing system on Lambda
  • Pregenerate every image in every aspect ratio
  • Msg sent via HTTP with image info in body
  • Lambda does resizing
  • Over 1000 lambdas running in parallel generated over 1bn images in 3 days
  • Saved 1bn dollars
  • Now developing deep learning image classification system with more metadata
• 13.9 avg page load time to 2.3 s
• 2 years ahead of competition
• Software agility means business agility

---

# Compute Services (Stephan Orban)

Back to Stephan Orbal

• From server to serverless
• Focus on the code rather than infrastructure, focusing on business logic.

# Virtual Machines

• Generally monolithic components
• Advantage: Fast upgrades on hardware
• New T3s have 30% speed increase over T2s.
• Higher end machines have large custom memory footprints
• F1s have field programmable gates
• Able to mix component elements, so M5 memory and F1 programmable gates

# Containers

• Container registry
• Easy deployment of containers thanks to Elastic Container Service (ECS)
• Multiple availability zones
• Integrated into the rest of the AWS environment
• Ability orchistrate tens of thousands of containers
• Elastic Kubernetes Service (EKA) allows ease of kubernetes hosting
• Fargate could be called "Serverless for containers". You are only billed for actual processing time

# Serverless

• For when all you care about it the business logic, don't care about infrastructure
• Scales to forever
• Only pay for what you use
• Really nice free tier 😉
• Wired into tons of AWS services

# Lambda Enterprise customer examples

Two companies who leverage Lambda to improve their systems and scale up

• Finra
• Autodesk

---

# Speaker: Ethan Kaplan, CPO, Fender Digital

• Yes, the 70 year old guitar company
• From the beginning and they used iterative engineering to create products
• Fender Digital has 40 employees in IT, 2000 overall
• Stats
  • 45% of players are new players in the ast year
  • 90% abandon guitar in a year
  • 20% of people who continue will buy 10 guitars over their lifetime
  • Mission: Change the stats
• Fender Play mobile was designed to change the stats. Free structured apps designed to help people learn.
• Needed to be agile and scale up rapidly
• They produce 30 lessons per day, how do they get that to an app
• Use Video Processing and Lambda to process incoming videos.
• Serverless had a huge cost benefit for them
• Rely API GAteway / Serverless to read Apple/Google callbacks to put data into DynamoDB. Eventually data goes to Redshift for analytics. Lessons learned:
  • DynamoDB has a learning curve - knowing how to add data to make it easy to query is important.
  • Instrumentation could be better
  • Lambda cold start issues
  • Build/deploy times for a Go App started taking a huge amount of time. Compilation was a beast.
• 40 services in use
• 21x traffic this year but 20% cheaper by moving everything to serverless
• Factories will be using IoT to control humidity and automation

---

# Data & Analytics (Stephan Orban)

• It can become expensive to run databases
  • Costs increased for cloud datastores
• Running AI/ML servers is also a cost

# Migration services

• 87,000+ databases into the AWS ecosystem

# Aurora

• MySQL or PostgreSQL
• RDS in memory
• Very fast

# Aurora Serverless

• MySQL or PostgreSQL
• RDS in memory
• Very fast
• On-demand, auto-scaling database for applications with unpredictable or cyclical workload

# Analytics

• About 80% of what we consider analytics is not analytics. Examples:
  • Data wrangling
  • Data transformations
  • Moving data to the right place (data is in silos)
• Mission: Flip it so 20% of analytics is not analytics
  • Get data
  • Automate data wrangling
  • lifecycle management
  • Provide access to data

# Using S3 as the analytics data storage

• Unmatched reliability, accessbility, durability
• Really good security
• Object level controls
• Versioning
• Lifecycle policies including archiving for long term storage
• Most ways to bring data in and out
• Tons of partner integrations
• Nigh-infinite formats because S3 is a file system!

# S3 Select and Glacier Select

• Run SQL expressions against s3 files
• Only pay for the compute resources while executing your jobs

# AWS Glue

• For loading data into S3, Redshift, and other data lakes
• Generates ETL scripts, loads transformations, and more
• Dump your data and off you go!
• Only pay for the compute resources while executing your jobs

# Machine Learning (Stephan Orban)

• Amazon retail uses machine learning to evaluate whethor or not an order is probably fraudulent.
• 70% of things watched on Amazon Prime and Netflix are driven by recommendations made by machine learning
• Machine learning had long been the domain of experts at Amazon
• Took too long to implement ML for internal Amazon efforts, so they made it easier for staff.
• 250% growth of ML on AWS this year
• 80% of ML queries are estimated to run on AWS

# AWS SageMaker

• Lets you run all the platforms (Torch, etc)
• Wrapper for Jupyter Notebook

# Other tools

• LEX: NTLK
• Polly: Text to speech
• Rekognition/Rekognition Video
• Transcribe
• Translate