Gene Yaacobi, Infrastructure Manager at Shutterstock, presents how his team uses active probing of their applications from around the world to help with choosing new data center locations.
Unraveling Multimodality with Large Language Models.pdf
Using Data to Determine Where to Build a New Data Center
1. Using Data to Determine
Where to Build a New Data
Center
Eugene Yaacobi, Shutterstock
2. Who Am I?
• Nerd
• Manager, Infrastructure @
Shutterstock
• I also like cars
3. A Little About
Shutterstock
• We are a stock photo, footage,
and music company
• 12 years old - founded in 2003
• Made in NY
• Offices in Amsterdam, Berlin,
Chicago, Dallas, Denver,
London, Los Angeles,
Montreal, Paris, San
Francisco, and Silicon Valley
• Over 600 employees
• 4 images sold every second
• Focus on volume, unbiased
search experience.
• Translated into 20 languages
• Accepting 9 currencies
• Customers in 150+ countries
4. A Little About
Shutterstock (cont.)
• Two-sided marketplace
• Over 1.3 million active
customers from 150 countries
• 80,000 contributors in 100
countries
• 60 mm images; over 50,000
images added every day (look
at the counter on the site and
round down to the nearest
million)
• 3mm+ video clips; over 90%
HD, 160,000 4K clips, and
3,000 videos added every day
• 500 mm paid downloads to
date, selling 4 images per
second
• $350 mm paid to contributors
since 2003
5. Infrastructure at
Shutterstock
• The service that every other
service is built on:
• 3 Datacenters
• 4,000 servers running
CentOS
• Servers are mostly
virtualized
• ~500 network devices
• Brocade load balancers
• Juniper routers and
switches, running JunOS
6. The Infrastructure Group at Shutterstock
• First responders to site issues
• 1OC for immediate alerts
• Made up of five teams:
• SRE - training
• Tools - internal tooling
• Storage - Purchasable asset storage
• Assets - Hardware acquisition
• Traffic - Network engineering
9. Every second counts! 40% of people abandon a website that takes more than 3
seconds to load.
10. If an e-commerce site is making
$100,000 per day, a 1 second page
delay could potentially cost you $2.5
million in lost sales every year
This makes executives sad
pandas
20. Like what?
• Ad-Hoc Troubleshooting
• Monitor and Alert on Overall Site Performance
• Does the site function properly and efficiently in all
locations?
• If not:
• Which locations are performing under expectations?
• Why are these locations seeing degraded
performance?
Started my life as a sysadmin when I was in Junior High.
I found a Red Hat 7.2 disc inside of a Red Hat For Dummies book, and popped the disc in Eventually I figured out my way around the command line
I thought it would be a great idea to make some money with my uber linux skills and started a hosting company.
Since being old enough to hold real jobs, I’ve mostly worked at various managed hosting companies and lead teams at them.
Joined Shutterstock in 2013 as part of the WebOps group
while on that group I built out an Openstack cluster,
ultimately moved into the Infrastructure Group
where I took over our datacenter and network operations before transitioning to my position today as Manager of Infrastructure.
We are a stock photo, footage, and music company:
12 years old - founded in 2003
Made in NY
Offices in Amsterdam, Berlin, Chicago, Dallas, Denver, London, Los Angeles, Montreal, Paris, San Francisco, and Silicon Valley
Over 600 employees
4 images sold every second
Focus on volume, unbiased search experience.
Translated into 20 languages
Accepting 9 currencies
Customers in 150+ countries
Everything you build or sell lives on top of the basic Infrastructure we design, scale & maintain.
Our goal is to build & maintain a first class Infrastructure to enable
+ Sales and Accounts teams to keep our customers delighted
+ Product and Engineering teams to rapidly make new products
Finance teams to support all business units
What is the environment like
Servers: CentOS
Network: JunOS
4k Running Instances - Mostly Virtual
3 Datacenters
Interesting facts
Stacked end to end, the amount of hard drives we use today is as tall as the Eiffel Tower
Power Statistic from 2020 Talk
We are the first responders
Infrastructure Supports over 30 product teams
1OC
SEV-0/1
Rest of team works on longer term projects/assists as needed
A day in the life of a network Engineer
Office Build-outs
WiFi Improvement Projects
Standardization / Automation
Standards Process (Using GitHub for +1,0,-1)
Troubleshooting (should somehow tie in how we use ThousandEyes)
Segway into Infra building out a new site
https://blog.kissmetrics.com/loading-time/
https://blog.kissmetrics.com/loading-time/
Many ways to figure this out
Google Analytics (see where your traffic is coming from)
I picked 10 places that I’d like to visit for this example and setup a
Let the data collect
I built a report in ThousandEyes with some things you may want to look at
Average page load times
Average response times
Sometimes a performance issue is related to the application and the network isn’t the bottleneck
In my test example, I will use South Korea
The important thing is to have the Connect, SSL, and Receive as low as possible
By being closer to the end-user we can shorten these times and have better performance
Adhoc Troubleshooting
Better visualization into how pervasive a reported issue is:
Is it global?
Is it only occurring in one location?
If so, is it widespread in that location, or just a few ISPs having trouble?