Chuck Neerdaels VP of Storage and Edge Services Y! Scalable Storage and Delivery Service
Cloud Storage & Edge Services Overview <ul><li>Bulk Storage  (Video, Mail attachments, Images, …) : </li></ul><ul><ul><li>...
CS&E cont. <ul><li>Messaging and Replication:  </li></ul><ul><ul><li>Many variations </li></ul></ul><ul><li>Edge: </li></u...
Typical Use Cases User Data and Preferences Image/Video Storage & Delivery RSS Feeds Caching, Load Balancing
A Quick Demo…
more cute puppy photos
Want a real demo?  Sports.yahoo.com Travel.yahoo.com Finance.yahoo.com Mail.yahoo.com News.yahoo.com *.yahoo.com
Infrastructure and Platform Service Reality <ul><li>Demos are only notable when they fail </li></ul><ul><li>On-line 7x24x3...
But what if they didn’t… <ul><li>Globally: </li></ul><ul><ul><li>Sites offline </li></ul></ul><ul><ul><li>Load balancing f...
What Makes an On-line Cloud Service at Y!… Global
Internal Expectations <ul><li>Global Solution, Including Emerging Markets </li></ul><ul><li>Fully Managed </li></ul><ul><l...
Common Challenges <ul><li>Speed of light </li></ul><ul><li>Huge traffic spike potential </li></ul><ul><li>Cost  </li></ul>...
A Little More Detail: MObStor & Sherpa
MObStor: Blowing Through the Cloudiness Global Servers Caching Protocol Termination Authentication Content Routing Local S...
MObStor: Trying to Increase 9’s DNS GSLB Hardware VIP Scalable Session  Manager Internal DNS with Load Feedback Hardware V...
Sherpa/PNUTs
Sherpa/PNUTs DNS GSLB Hardware VIP Scalable Router & Session Manager Consistency vs. Availability Knobs Local Replication ...
Physics and Economics for a Global Cloud…
What’s your target SLA? 0  Distance between producer/consumer *Illustrative, and not intended to portray reality Latency S...
What’s your budget? Latency Replication Costs *Illustrative, and not intended to portray reality $$$ $
Lessons Learned <ul><li>Intuition is often wrong – let  data  drive your data </li></ul><ul><li>Provide hooks, experimenta...
Why would any of you care about a Y! proprietary cloud?
Committed to Open Source <ul><li>Component approach – current pipeline: </li></ul><ul><li>ZooKeeper (already available) </...
<ul><li>YAHOO! IS OPEN SOURCING </li></ul><ul><li>Cloud building block </li></ul><ul><li>Processes up to 35K req/second pe...
Questions? Y! Scalable Storage and Delivery Service
Upcoming SlideShare
Loading in …5
×

Yahoo! Scalable Storage and Delivery Services

4,063 views
3,952 views

Published on

From Chuck Neerdaels' presentation "Yahoo! Scalable Storage and Delivery Services" at the 2009 Cloud Computing Expo in Santa Clara, CA, USA. Here's the talk description on the Expo's site: http://cloudcomputingexpo.com/event/session/541

Published in: Technology

Yahoo! Scalable Storage and Delivery Services

  1. 1. Chuck Neerdaels VP of Storage and Edge Services Y! Scalable Storage and Delivery Service
  2. 2. Cloud Storage & Edge Services Overview <ul><li>Bulk Storage (Video, Mail attachments, Images, …) : </li></ul><ul><ul><li>MObStor </li></ul></ul><ul><ul><li>DORA (Clustered Object Store) </li></ul></ul><ul><li>Structured Storage (key-value and user data): </li></ul><ul><ul><li>Sherpa/PNUTs </li></ul></ul><ul><ul><li>User Database (UDB) </li></ul></ul><ul><ul><li>UDS/UPS </li></ul></ul>
  3. 3. CS&E cont. <ul><li>Messaging and Replication: </li></ul><ul><ul><li>Many variations </li></ul></ul><ul><li>Edge: </li></ul><ul><ul><li>ySquid </li></ul></ul><ul><ul><li>Traffic Server </li></ul></ul><ul><ul><li>Core Proxy Infrastructure </li></ul></ul><ul><ul><li>Multi-vendor CDN </li></ul></ul><ul><ul><li>Global DNS & local load balancing technology </li></ul></ul>
  4. 4. Typical Use Cases User Data and Preferences Image/Video Storage & Delivery RSS Feeds Caching, Load Balancing
  5. 5. A Quick Demo…
  6. 6.
  7. 7. more cute puppy photos
  8. 8.
  9. 9. Want a real demo? Sports.yahoo.com Travel.yahoo.com Finance.yahoo.com Mail.yahoo.com News.yahoo.com *.yahoo.com
  10. 10. Infrastructure and Platform Service Reality <ul><li>Demos are only notable when they fail </li></ul><ul><li>On-line 7x24x365 </li></ul><ul><ul><ul><li>Through upgrades, attacks, outages, human errors, … </li></ul></ul></ul><ul><ul><ul><li>Legitimate use cases can be as challenging as abuse </li></ul></ul></ul><ul><li>It’s all assumed to, “just work” </li></ul><ul><li>But, what if they didn’t… </li></ul>
  11. 11. But what if they didn’t… <ul><li>Globally: </li></ul><ul><ul><li>Sites offline </li></ul></ul><ul><ul><li>Load balancing fails </li></ul></ul><ul><ul><li>Images, CSS, and JavaScript not served </li></ul></ul><ul><ul><li>No logins and personalization </li></ul></ul><ul><ul><li>Normalized RSS feeds go stale </li></ul></ul><ul><ul><li>Site thumbnails fail </li></ul></ul><ul><ul><li>Server-to-server communication ceases </li></ul></ul><ul><li>Ads </li></ul><ul><ul><li>Ads not served </li></ul></ul><ul><ul><li>Ad pipelines halt </li></ul></ul><ul><li>Flickr: </li></ul><ul><ul><li>No more video or images </li></ul></ul><ul><li>Mail: </li></ul><ul><ul><li>Webmail slow of offline, attachments missing </li></ul></ul><ul><li>Mobile & Social: </li></ul><ul><ul><li>Features go offline </li></ul></ul><ul><li>Sports </li></ul><ul><ul><li>Overwhelmed every Sunday </li></ul></ul><ul><ul><li>Fantasy Sports draft data gets stale </li></ul></ul>Yahoo! sites would become stale, boring, slow, non-profitable, or non-existent
  12. 12. What Makes an On-line Cloud Service at Y!… Global
  13. 13. Internal Expectations <ul><li>Global Solution, Including Emerging Markets </li></ul><ul><li>Fully Managed </li></ul><ul><li>Multi-tenant (Finance) </li></ul><ul><li>Single tenant (Customers) </li></ul><ul><li>Elastic </li></ul><ul><li>Predictable SLA’s </li></ul><ul><li>Simple to provision and manage </li></ul><ul><li>Providing the illusion of infinite bandwidth, capacity, (and single tenancy)… </li></ul>
  14. 14. Common Challenges <ul><li>Speed of light </li></ul><ul><li>Huge traffic spike potential </li></ul><ul><li>Cost </li></ul><ul><li>Partitioned network failures are assumed </li></ul><ul><li>Consumer/User intuition </li></ul>
  15. 15. A Little More Detail: MObStor & Sherpa
  16. 16. MObStor: Blowing Through the Cloudiness Global Servers Caching Protocol Termination Authentication Content Routing Local Servers Auto Expiration De-duplication Object Placement Re-replication
  17. 17. MObStor: Trying to Increase 9’s DNS GSLB Hardware VIP Scalable Session Manager Internal DNS with Load Feedback Hardware VIP Local Replication Global Replication Separate Metadata Replication
  18. 18. Sherpa/PNUTs
  19. 19. Sherpa/PNUTs DNS GSLB Hardware VIP Scalable Router & Session Manager Consistency vs. Availability Knobs Local Replication Global Replication Ordered Messaging
  20. 20. Physics and Economics for a Global Cloud…
  21. 21. What’s your target SLA? 0 Distance between producer/consumer *Illustrative, and not intended to portray reality Latency Speed of light + network degradation
  22. 22. What’s your budget? Latency Replication Costs *Illustrative, and not intended to portray reality $$$ $
  23. 23. Lessons Learned <ul><li>Intuition is often wrong – let data drive your data </li></ul><ul><li>Provide hooks, experimental feedback, and mobility </li></ul><ul><li>N-way global replication is really expensive </li></ul><ul><li>Customers don’t understand 95/5 billing </li></ul><ul><li>Customers don’t understand CAP theorem </li></ul><ul><li>Verify all provisioning </li></ul><ul><li>Strive for quality, but plan for failure </li></ul>
  24. 24. Why would any of you care about a Y! proprietary cloud?
  25. 25. Committed to Open Source <ul><li>Component approach – current pipeline: </li></ul><ul><li>ZooKeeper (already available) </li></ul><ul><li>Traffic Server (just released!) </li></ul><ul><li>More to come </li></ul>
  26. 26. <ul><li>YAHOO! IS OPEN SOURCING </li></ul><ul><li>Cloud building block </li></ul><ul><li>Processes up to 35K req/second per box </li></ul><ul><li>Extensible Framework </li></ul><ul><li>Edge Caching, Load Balancing, Proxying </li></ul><ul><li>Leveraged in Edge, Storage, Serving services </li></ul><ul><li>Supports more than half of all Yahoo! traffic </li></ul><ul><li>Available via the Apache Software Foundation </li></ul>
  27. 27. Questions? Y! Scalable Storage and Delivery Service

×