AWS Reinvent Conference 2014

I had a free conference pass to the AWS Reinvent Conference, but I just got back from Las Vegas, so I decided to watch the streaming broadcast.

Wednesday Keynotes

– Medical use case
– 1 PB per month storage


– 100,000 writes per second, 5x reads
– 6 copies across 3 DCs, backups to S3
– immediate restart (no log replay)
– pre-warmed cache


– CodeDeploy
– internal developers like it, miss it

Hybrid Mode


– customer-focused, rather than competitor or product-focused
– saved $350 million
– customers were surprised initially
– how many vendors give back money?


– moved old DC, to AWS first
– 2013 Partner of the Year
– worked with AWS on new security features
– moving everything else now

Johnson & Johnson

– lots of M&A, but divisions operate locally
– need analytics
– users asked for better, faster, cheaper, offering more of the same thing wasn’t enough
– AWS makes their engineers happy

AWS Security Keynote
Stephen Schmidt, GM & CISO, AWS

– security is not new to Amazon
– more visible than most on-premises
– one bar – same for everybody
– objects have records, logs and timestamps in Cloudtrail
– sign up for SMS on changes
– HSM launched in Europe first because of demand
– Cause of Error (COE) investigation for each problem
– security engineers are embedded in service teams, CISO reports to CEO. Security is job 1
– pen testing black and white
– collect ddos cannons and mitigate, often beforehand

Infrastructure as Code
Alex Corley and David Winter, AWS
Tom Wanielista, Simple

– Tom needed PCI. So had to start over with VPC, scan each node and binary, etc.

The Quest for the Last 9: Building Highly Available Services from the Ground Up
Khawaja Shams, Technical Advisor, AWS

– Cross-Region Replication


– ad server
– 1 million QPS on DynamoDB, 1 PB logs (compressed) per month
– rather scale engineering team for products, not infra
– 1% downtime is $1 million on $100 million in revenue, plus your customers

Managing the Pace of Innovation: Behind the Scenes at AWS
Charlie Bell, SVP of Utility Computing Services, AWS
Jim Scharf, Technical Advisor, AWS

– “a team should be small enough to be fed on 2 pizzas”
– more than 40 price reductions in 8 years

James Hamilton, VP & Distinguished Engineer, AWS

– great talk, deep and thought-provoking at the same time, as always for JH

– 1 million active clients, storage doubled in 12 months, 5x size of combined competitors
– cost of networking at 8% and growing is anti-Moore’s Law – 2 reasons: indefensible vendor profits, servers getting fatter
– just the support cost was $10+ millions, availability better with own gear because better understanding
– tested on 3 megawatt DC with 8,000 servers. rented of course.
– team metrics must reflect customer complaints, or simply wrong
– 11 regions, over 28 data center buildings
– East-West replication at 74 ms can only be async with data loss on emergency cutover and a 1-week cleanup, so nobody wants to fail over
– reason for AZs is to limit blast-radius while having sync replication under 2 ms
– SR-IOV reduces avg. latency 2x, outliers down 10x
– built 2 power plants because utilities move too slow. Slightly cheaper, but very handy to have power engineers on staff now.
– Aurora is mysql with steroidal storage engine

Thursday Keynotes

Werner Vogels, AWS

– single button deploy to another region

– lego set for building a music service
– music industry could not be made more complicated if you tried
– with what I know now, I wouldn’t touch the music industry with a barge pole. I should have written a taxi app.

The Weather Company

– most distributed cable channel in USA (, weather underground)

Pristine Heath Care

– NICU broadcast with Google Glass
– AWS and Docker
– joined AWS startup program

EC2 Container Services

– new
– demo with Redis

Ben Golub, CEO

– Docker just turned 18 months old
– AWS is respectful of Docker isms

Extreme Availability for Mission-Critical Applications
Raul Frias and Eduardo Horai, Solution Architects, AWS
Andre Fatala, CDO, Magazine Luiza

– challenges of doing HA IT in South America
– misc.

Building Mission Critical Database Applications: A Conversation with AWS Customers about Best Practices
Swami Sivasubramanian, General Manager, AWS NoSQL

Financial Times

– loves the Internet, audience is biggest ever, and more global audience
– UCS “pay as you go, without the go” 🙂
– 50x faster and 20x less. on Redshift. “some of the analysts thought it was broken”
– “we have machines older than you. we cut them up into bookends and give them to the IT staff.”
– easy to demo Redshift. a week to do setup, a few weeks to try it out. easier than months-long RFP process.
– MapReduce is a lot of conversion effort and retraining for staff compared to SQL
– getting HA vertically is exponentially more expensive per 9. We prefer to re-think availability around AWS.
– kafka
– microservices
– no pets: “the day that a server with 464 day uptime goes down would be a terrible day.”
– rebooting for shellshock wasn’t a concern since they reboot all the time
– SAN will just propagate block-level corruption to all apps, so not HA enough
– “data debt is the worst type of technical debt, because you have to go through the whole company to collect it”
– wish list: would like more inference and semantic database querying, few available are ready to scale now.


– lost 3 drives in disk array, had to wait 5 days for replacements. gave up on colo after that and moved to AWS
– 1 PB per month
– 100 ms deadline per request from request to price quote (pulled out of a hat, now industry standard)
– “we spend less on Dynamodb than on snacks” (operating redis in-house cost 2 FTE plus 32 machines)
– they are paying half of a published Cassandra ring doing 1 million ops
– need consistent latency below 5 ms, scalable, cheap
– initial capacity planning is a guess
– for every 10 ad companies that die per day, 100 start up
– moving to Route 53 => 4xELBs => stateless EC2 with health monitoring daemon that dies after 5 min, ships logs, terminates. ASG.
– good backup and restore makes experimenting cheaper.
– wish list: would like in-memory Redshift to be 100x faster and PB-scale.


– “we don’t have that many engineers working on Dynamo, but I can’t tell you how many”

State of the Union: Amazon Compute Services
Peter De Santis, Vice President, Amazon Compute Services, AWS
Matt Garman, Vice President, Amazon EC2, AWS

– Nov. 2011 1064 nodes Top 500 #42, only one on ethernet
– enhanced networking instances have high pps
– HGST does hard disk drive design on c3, a million permutations in 8 hours, 70,000 cores. $5,593.94
vs. in-house cluster of $15 million
– C4 instance has dedicated network paths for EBS, network, etc. now at no extra cost

Develop an Enterprise-wide Cloud Adoption Strategy
Blake Chism, Senior Consultant, AWS Professional Services
Miha Kralj, Principal Consultant, AWS Professional Services

– Professional Services helps document enterprise architecture
– for EA “it’s easier to slice the elephant into smaller pieces”
– interesting modeling slides
– download the whitepaper, first of a series of 8
– interesting blue bike outfit.


– big system – 21,000 miles of rail
– “business process, people, technology” are 3 components
– mainframe, software that needs to be re-architected, ready to move

Notre Dame

– started in EC2 classic, likes auto-scaling, moving to VPC
– committed to moving 80% of IT services to cloud in next 3 years
– Google apps user
– hosting external auth in AWS for traveling users
– AWS Professional Services monitors big games

Brendan Gregg, Netflix: Performance Tuning EC2 Instances Where the Internet lives Amazon details how it does networking in its data centers
A Day in the Life of a Billion Packets (CPN401) | AWS re:Invent 2013
The Stunning Scale of AWS and What it Means for the Future of the Cloud

This entry was posted in API Programming, Business, Cloud, Conferences, Linux, MySQL, Open Source, Oracle, Storage, Tech. Bookmark the permalink.

Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.