RightScale Conference Santa Clara 2011: Scaling an application in the cloud often hits the most common bottleneck – the database tier. Not only is database performance the number one cause of poor application performance, but also the issue is magnified in cloud environments where I/O and bandwidth is generally slower and less predictable than in dedicated data centers. Database sharding is a highly effective method of removing the database scalability barrier, operating on top of proven RDBMS products such as MySQL and Postgres – as well as the new NoSQL database platforms. One critical aspect often given too little consideration is monitoring and continuous operation of your databases, including the full lifecycle, to ensure that they stay up.

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

2. Introduction
 Who I am
 Cory Isaacson, CEO of CodeFutures
 Providers of dbShards
 Author of Software Pipelines
 Leaders in scalability, performance, high-availability
and database solutions…
 …based on real-world experience with dozens of cloud-
based and data center hosted 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 database
scalability
 Databases grow…
 …and get slower
 Why?
 User content, locking
 Connection limitations
 Index Depth
 Table Sizes
 I/O bottlenecks
 Challenges apply regardless of hosting
environment

4. Challenges in the cloud
environment
 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”

5. Typical Application Architecture

6. Scaling your application
 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
 Add or remove servers as need dictates
 If one goes down…
 …failover to another

7. Scaling your application
 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!

8. Scaling your application
 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
 The best time to plan your database scalability
strategy is now
 don‟t wait until it‟s a “3-alarm fire”

9. More Database scalability
challenges
 Databases have many other challenges that limit scalability
 ACID compliance…
 …especially Consistency
 …user contention
 Preserving Data Relationships…
 Referential Integrity
 Operational challenges
 Failover
 Planned, unplanned
 Maintenance
 Index rebuild
 Restore
 Space reclamation
 Lifecycle
 Reliable Backup/Restore
 Monitoring
 Application Updates
 Management

10. FamilyBuilder
 Innovator in Facebook applications
 Among first 500 apps worldwide
 50MM Users, 100,000 new users/day

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

12. Database slowdown is not
linear…
Database Load Curve
10000
9000
8000
7000
Load Time
6000
5000
4000 Time
3000 Expon. (Time)
2000
1000
0
0 10 20 30 40
Data File (GB)
GB Load Time (Min)
.9 1
1.3 2.5
3.5 11.7
39.0 10 days…

13. Challenges apply to all types of
databases
 Traditional RDBMS (MySQL, Postgres, Oracle…)
 I/O bound
 Multi-user, lock contention
 High-availability
 Lifecycle management
 NoSQL, “NewSQL”, Column Databases, Caching…
 Reliability
 Limits of a single server
 …and a single thread
 Data dumps to disk
 High-availability
 Lifecycle Management

14. Challenges apply to all types of
databases
 No matter what the technology, big databases
are hard to manage
 elasticscaling is a real challenge
 degradation from growth in size and volume is a
certainty
 application-specific database requirements add to
the challenge
 Sound database design is key…
 balance performance vs. convenience vs. data
size

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

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

17. 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

18. What is Database Sharding?
 Start with a big monolithic database
 break it into smaller databases
 across many servers
 using a key value

19. The key to Database
Sharding…

20. Database Sharding Architecture

21. Database Sharding Architecture

22. Database Sharding… the
results

23. 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
 reads are faster too
 aggregate, list queries are generally much faster
 No contention between servers
 locking, disk, memory, CPU
 Allows for intelligent parallel processing
 Go Fish queries across shards
 Keep CPUs busy and productive

24. Breaking the I/O Barrier

25. Database types
 Traditional RDBMS
 Proven
 SQL language…
 …although ORM can change that
 Durable, transactional, dependable…
 …but inflexible
 NoSQL/NewSQL
 In-memory, memory-cache to disk, single-threaded…
 …the real speed advantage
 Various flavors…
 Key/Value Store
 Document database
 Hybrid

26. 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

27. How Database Sharding works for
NoSQL

28. How Relational Sharding works
Primary Shard Table
Shard Child Tables
Global Tables

29. How Relational Sharding works

30. How Relational Sharding works
 Shard key recognition in SQL
 SELECT * FROM customer
WHERE customer_id = 1234
 INSERT INTO customer
(customer_id, first_name, last_name, addr_line1,
…)
VALUES
(2345, „John‟, „Jones‟, „123 B Street‟,…)
 UPDATE customer
SET addr_line1 = „456 C Avenue‟
WHERE customer_id = 4567

31. What about Multi-Shard result
sets?

32. Multi-shard result set example
 Go Fish (no shard key)
 SELECT
country_id, count(*) FROM customer
GROUP BY country_id

33. More on Multi-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

34. The Database Sharding
Process

35. 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
 Disaster-recovery…
 …multi-region
 …multi-cloud

36. Database Sharding…elastic
shards

37. Database Sharding summary
 Database Sharding is the most effective tool for scaling your
database tier in the cloud…
 …or anywhere
 Use Database Sharding for any type of engine…
 …SQL
 …NoSQL
 …Column Db
 …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

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