Database performance is the number-one cause of poor performance for scalable web applications, and the problem is magnified in cloud environments where I/O and bandwidth are generally slower and less predictable than in dedicated data centers. Database sharding is a highly effective method of removing the database scalability barrier by operating on top of proven RDBMS products such as MySQL and Postgres as well as the new NoSQL database platforms. In this session, you'll learn what it really takes to implement sharding, the role it plays in the effective end-to-end lifecycle management of your entire database environment, why it is crucial for ensuring reliability, and how to choose the best technology for a specific application. We'll also share a case study on a high-volume social networking application that demonstrates the effectiveness of database sharding for scaling cloud-based applications.

1. Scaling your Database in the Cloud Spring2011 Presented by: Cory Isaacson, CEO CodeFutures Corporation http://www.dbshards.com View the video of this presentation here!

2. Introduction Who I am Cory Isaacson, CEO of CodeFutures Providers of dbShards Author of Software Pipelines Partnerships: Rightscale The leading Cloud Management Platform Leaders in database scalability, performance, and high-availability for the cloud… …based on real-world experience with dozens of cloud-based applications …social networking, gaming, data collection, mobile, analytics Objective is to provide useful experience you can apply to scaling (and managing) your database tier… …especially for high volume applications

3. Challenges of cloud computing Cloud provides highly attractive service environment Flexible, scales with need (up or down) No need for dedicated IT staff, fixed facility costs Pay-as-you-go model Cloud services occasionally fail Partial network outages Server failures… …by their nature cloud servers are “transient” Disk volume issues Cloud-based resources are constrained CPU I/O Rates… …the “Cloud I/O Barrier”

4. Typical Application Architecture

5. Scaling in the Cloud Scaling Load Balancers is easy Stateless routing to app server Can add redundant Load Balancers if needed DNS round-robin or intelligent routing for larger sites If one goes down… …failover to another Scaling Application Servers is easy Stateless Sessions can easily transition to another server Add or remove servers as need dictates If one goes down… …failover to another

6. Scaling in the Cloud Scaling the Database tier is hard “Stateful” by definition (and necessity…) Large, integrated data sets… 10s of GBs to TBs (or more…) Difficult to move, reload I/O dependent… …adversely affected by cloud service failures …and slow cloud I/O If one goes down… …ouch! Databases form the “last mile” of true application scalability Initially simple optimization produces the best result Implement a follow-on scalability strategy for long-term performance goals… …plus a high-availability strategy is a must

7. All CPUs wait at the same speed… The Cloud I/O Barrier

8. More Database scalability challenges Databases have many other challenges that limit scalability ACID compliance… …especially Consistency …user contention Operational challenges Failover Planned, unplanned Maintenance Index rebuild Restore Space reclamation Lifecycle Reliable Backup/Restore Monitoring Application Updates Management

9. Database slowdown is not linear…

10. Challenges apply to all types of databases Traditional RDBMS (MySQL, Postgres, Oracle…) I/O bound Multi-user, lock contention High-availability Lifecycle management NoSQLDatabases In-memory, Caching, Document…) Reliability, High-availability Limits of a single server …and a single thread Data dumps to disk Lifecycle Management No matter what the technology, big databases are hard to manage… …elastic scaling is a real challenge …degradation from growth in size and volume is a certainty Application-specific database requirements add to the challenge …database design is key… …balance performance vs. convenience vs. data size

11. The Laws of Databases Law #1: Small Databases are fast… Law #2: Big Databases are slow… Law #3: Keep databases small

12. What is the answer? Database sharding is the only effective method for achieving scale, elasticity, reliability and easy management… …regardless of your database technology

13. What is Database Sharding? “Horizontal partitioning is a database design principle whereby rows of a database table are held separately... Each partition forms part of a shard, which may in turn be located on a separate database server or physical location.” Wikipedia

14. The key to Database Sharding…

15. Database Sharding Architecture

16. Database Sharding Architecture

17. Database Sharding… the results

18. Why does Database Sharding work? Maximize CPU/Memory per database instance… …as compared to database size Reduce the size of index trees Speeds writes dramatically No contention between servers Locking, disk, memory, CPU Allows for intelligent parallel processing… …Go Fish queries across shards Keep CPUs busy and productive

19. Breaking the Cloud I/O Barrier

20. Database types Traditional RDBMS Proven SQL language… …although ORM can change that Durable, transactional, dependable… …perceived as inflexible NoSQL Typically in-memory… …the real speed advantage Various flavors… Key/Value Store Document database Hybrid Store complex documents using a key In-memory and disk based

21. Black box vs. Relational Sharding Both utilize sharding on a data key… …typically modulus on a value or consistent hash Black box sharding is automatic… …attempts to evenly shard across all available servers …no developer visibility or control …can work acceptably for simple, non-indexed NoSQL data stores …easily supports single-row/object results Relational sharding is defined by the developer… …selective sharding of large data …explicit developer control and visibility …tunable as the database grows and matures …more efficient for result sets and searchable queries …preserves data relationships

22. How Database Sharding works for NoSQL

23. Relational Sharding Shard-Tree Root Table Shard-Tree Child Tables Global Tables

24. How Relational Sharding works

25. How Relational Sharding works Shard key recognition in SQL SELECT * FROM customerWHERE customer_id = 1234 INSERT INTO customer(customer_id, first_name, last_name, addr_line1,…)VALUES(2345, ‘John’, ‘Jones’, ‘123 B Street’,…) UPDATE customerSET addr_line1 = ‘456 C Avenue’WHERE customer_id = 4567

26. What about Cross-Shard result sets?

27. Cross-shard result set example Go Fish (no shard key) SELECT country_id, count(*) FROM customerGROUP BY country_id

28. More on Cross-Shard result sets… Black Box approach requires “scatter gather” for multi-row/object result set… …common with NoSQL engines …forces use of denormalized lists …must be maintained by developer/application code …Map Reduce processing helps with this (non-realtime) Relational sharding provides access to meaningful result sets of related data… …aggregation, sort easier to perform …logical search operations more natural …related rows are stored together

29. Make your Application Shard-Safe For sharding to be efficient, every sharded table should include a shard key. SELECT WHERE clause statements should include a shard key where possible so that the query goes directly to a single shard rather than being executed against multiple shards in parallel. INSERT, UPDATE and DELETE WHERE CLAUSE statements must include a shard key so that the data is written to the correct shard. Avoid cross-shard inner joins. For example, joining one customer to another is not feasible as both customers may be in different shards. This can be accomplished, but requires complex logic and can adversely affect performance.

30. What about High-Availability? Can you afford to take your databases offline: …for scheduled maintenance? …for unplanned failure? …can you accept some lost transactions? By definition Database Sharding adds failure points to the data tier A proven High-Availability strategy is a must… …system outages…especially in the cloud …planned maintenance…a necessity for all database engines

32. No rapid failover for continuous operation

33. Maintenance difficult in production environments

35. …3 active-active instances for in-memory databases

36. Support for…

37. …fail-down

38. …failover

39. …maintenance

40. …automated backup/restore

42. Database Sharding…elastic shards Expand the number of shards… …divide a single shard into N new shards …or rebalance existing shards across N new shards Contract the number of shards… …consolidate N shards into a single shard Due to large data sizes, this takes time… …regardless of data architecture Requires High-Availability to ensure no downtime… …perform scaling on live replica in background …fail back to active instances

43. Database Sharding…the future Ability to leverage proven database engines… …SQL …NoSQL …Caching Hybrid database infrastructures using the best of all technologies In-memory and disk-based solutions co-existing Allow developers to select the best database engine for a given set of application requirements… …seamless context-switching within the application …use the API of choice Improved management… …monitoring …configuration “on-the-fly” …dynamic elastic shards based on demand

44. Database Sharding summary Database Sharding is the only effective tool for scaling your database tier in the cloud… …or anywhere Use Database Sharding for any type of engine… …SQL …NoSQL …Cache Use the best engine for a given application requirement… …avoid the “one-size-fits-all” trap …each engine has its strengths to capitalize on Ensure your High-Availability is proven and bulletproof… …especially critical on the cloud …must support failure and maintenance

45. Questions/Answers Cory Isaacson CodeFutures Corporation cory.isaacson@codefutures.com http://www.dbshards.com