Petabytes and Nanoseconds
- 1. Petabytes and Nanoseconds Distributed Data Storage andthe CAP Theorem FIN talk Robert Greiner Nathan Murray August 21,2014
- 2. CHAPTER The Problems Your phone can add two numbers in the same time it takes light to travel one foot All high frequency trading servers are connected to the NASDAQ network with the same length of cable, so that no party has a speed advantage
- 3. A Common Scenario Web Application RDBMS + =
- 4. The Solution: Scale All the Things!!1
- 5. Why shouldwe scale? Throughput Latency Storage Reliability
- 6. The Solution? Add a load balancer Add more web servers Tune the DB. Indexes,SPs, etc.
- 7. There’sa new bottleneck Generally an RDBMS can becomea bottleneck around 10K transactions per second
- 8. Next Step… Distribute Your Data Each web server can talk to any data storage node Nodes distribute queries and replicate data – lots more complexity!
- 9. Cluster = Additional Complexity
- 10. Enter the CAP Theorem! This guy created the CAP Theorem This guy’s VP Invented the internet
- 11. CAP Theorem: Defined Within a distributed system, you can only make two of the following three guarantees across a write/read pair
- 12. Guarantee 1: Consistency If a value is written, and then fetched, I will alwaysget back the new value Note: not the same as the C in ACID! _
- 13. Guarantee 2: Availability If a value is written, a success message should always be returned. If a subsequent read returns a stale value, or something reasonable, it’s OK. _ Note: not the same as the A in HA!
- 14. Guarantee 3: Partition Tolerance The system will continue to function when network partitions occur –OOP != NP. _ Note: nothing to do with BAC!
- 15. CAP Triangle The CAP Theorem is explained as a triangle C, A or P: Pick two This is true in practice, except…
- 16. When choosing a distributed system… vs.
- 17. … You Can’t Sacrifice Partition Tolerance! NOTDistributed (a.k.a. NOTPartition Tolerant) Available AND Consistent Distributed (a.k.a. Partition Tolerant) Available OR Consistent _ _
- 18. CPvs. AP Synchronous. Waits until partition heals or times out. Asynchronous. Returns a reasonable response always.
- 19. CPvs. AP Synchronous. Waits until partition heals or times out. Asynchronous. Returns a reasonable response always. At a bank, you get a deposit receipt afterthe work is complete At a coffee shop, you get a receipt beforethe work is complete
- 20. CHAPTER Whendo companies care?
- 21. Companies care about internetscale
- 22. Distributed Storage Past 2004 Google’s Map Reduce paper published 2006 Google’s Big Table paper published 2007 Amazon’s Dynamo paper published 2008 Yahoo runs search on Hadoop 2008 Facebook open sources Cassandra 2008 Bitcoin paper published 2009 Yahoo open sources Hadoop 2010 Azure Table Storage released 2012 Google’s Spanner and F1 papers 2013 Amazon releases DynamoDB inside AWS 2014 Google’s Mesa paper published 2015 ????
- 23. Looking forward •Open source implementations of more sophisticated storage systems •Managed services with more advanced capabilities •Google Cloud versions of F1, Spanner, or Mesa? •NoSQL + SQL •Distributed data storage in untrusted environments
- 24. CHAPTER How does this affect me
- 25. Even our most “legacy” clients are already starting to care about internet scale: _
- 27. Scenario Client = Energy Retailer (Independent Sales Force) Sales Agent captures info about potential customer Price generated on-demand based on daily rate curve Quote no longer valid at midnight Each night, rates are updated based on new rate-curve Used to take 4hours Now takes > 24hours (Due to increased demand)
- 28. Current State
- 29. Solution Strategy Assess •Analyze business performance needs •Select non-performing work streams •Filter –(Could/Should) •Prioritize •Performance Baseline / Load Test Strategize •Identify Bottlenecks (CPU/RAM/Network) •Optimization strategy •Technology Selection Implement •POC •Load Test •Optimize •Build
- 30. Optimize Code Scale Up Scale Out Managed Service
- 31. Optimize CodeLevel 1 Least organizational impact No architecture changes required Use existing development processes Risky –Code may be fine Expensive –Dev Resources Time Consuming –Dev + Deploy
- 32. Scale UpLevel 2 Easiest solution Utilize existing infrastructure Little/no architecture changes Low probability of network partitions May not solve the problem long-term Hardware limitations Non-linear improvement (2x RAM != 2x Performance) C/A
- 33. Scale OutLevel 3 Highest throughput Improved system up-time No single point of failure Linear performance increases Use commodity hardware –Hard to scale-up CPU Increased infrastructure / system complexity Increased probability of network partitions Automation complexity A/C
- 34. Managed ServiceLevel 4 Low barrier to entry No additional hardware investment required Treat as extension of existing data center Appliance configuration Globally redundant (cloud) Most organizational change Less control and customization Built-in redundancy and innovation C/A A/C
- 35. Optimize Code(Level 1) •Least organizational impact •No architecture changes required •Use existing development processes •Risky –Code may be fine •Expensive –Dev Resources •Time Consuming –Dev + Deploy Scale Up(Level 2) •Easiest solution •Utilize existing infrastructure •Little/no architecture changes •Reduce probability of network partitions •May not solve the problem long-term •Hardware limitations •Non-linear improvement Scale Out(Level 3) •Highest throughput •Improved system up-time •No single point of failure •Linear performance inc. •Use commodity hardware •Increased infrastructure / system complexity •Increased probability of network partitions •Automation complexity Managed Service(Level 4) •Low barrier to entry •No additional hardware investment required •Treat as extension of existing data center •Appliance configuration •Globally redundant (cloud) •Most organizational change •Less control and customization •High innovation Pick One (Or More!)
- 36. First Attempt
- 37. Good Enough?
- 38. Taking It to the Next Level