Scaling the Web: Databases & NoSQL

This is an introduction to relational and non-relational databases and how their performance affects to scaling a web application.

This is a recording of a guest Lecture I gave at the University of Texas school of Information.

In this talk I address the technologies and tools Gowalla (gowalla.com) uses including memcache, redis and cassandra.

Find more on my blog:
http://schneems.com

Transcript

1. Scaling the Web: Databases & NoSQL Richard Schneeman @schneems works for @Gowalla Wed Nov 10 2011
2. whoami • @Schneems • BSME with Honors from Georgia Tech • 5 + years experience Ruby & Rails • Work for @Gowalla • Rails 3.1 contributor : ) • 3 + years technical teaching
3. Traffic
4. Compounding Traffic ex. Wikipedia
5. Compounding Traffic ex. Wikipedia
6. Gowalla
7. Gowalla • 50 best websites NYTimes 2010 • Founded 2009 @ SXSW • 1 million+ Users • Undisclosed Visitors • Loves/highlights/comments/stories/guides • Facebook/Foursquare/Twitter integration • iphone/android/web apps • public API
9. Gowalla Backend • Ruby on Rails • Uses the Ruby Language • Rails is the Framework
10. The Web is Data • Username => String • Birthday => Int/ Int/ Int • Blog Post => Text • Image => Binary-file/blob Data needs to be stored to be useful
11. Database
12. Gowalla Database • PostgreSQL • Relational (RDBMS) • Open Source • Competitor to MySQL • ACID compliant • Running on a Dedicated Managed Server
13. Need for Speed • Throughput: • The number of operations per minute that can be performed • Pure Speed: • How long an individual operation takes.
14. Potential Problems • Hardware • Slow Network • Slow hard-drive • Insufficient CPU • Insufficient Ram • Software • too many Reads • too many Writes
15. Scaling Up versus Out • Scale Up: • More CPU, Bigger HD, More Ram etc. • Scale Out: • More machines • More machines • More machines • ...
16. Scale Up • Bigger faster machine • More Ram • More CPU • Bigger ethernet bus • ... • Moores Law • Diminishing returns
17. Scale Out • Forget Moores law... • Add more nodes • Master/ Slave Database • Sharding
18. Master DB Slave DB Slave DB Slave DB Slave DB Write Copy Read Master/Slave
19. Master & Slave +/- • Pro • Increased read speed • Takes read load off of master • Allows us to Join across all tables • Con • Doesn’t buy increased write throughput • Single Point of Failure in Master Node
20. Users in USA Read Sharding Write Users in Europe Users in Asia Users in Africa
21. Sharding +/- • Pro • Increased Write & Read throughput • No Single Point of failure • Individual features can fail • Con • Cannot Join queries between shards
22. What is a Database? • Relational Database Managment System (RDBMS) • Stores Data Using Schema • A.C.I.D. compliant • Atomic • Consistent • Isolated • Durable
23. RDBMS • Relational • Matches data on common characteristics in data • Enables “Join” & “Union” queries • Makes data modular
24. Relational +/- • Pros • Data is modular • Highly flexible data layout • Cons • Getting desired data can be tricky • Over modularization leads to many join queries • Trade off performance for search-ability
25. Schema Storage • Blueprint for data storage • Break data into tables/columns/rows • Give data types to your data • Integer • String • Text • Boolean • ...
26. Schema +/- • Pros • Regularize our data • Helps keep data consistent • Converts to programming “types” easily • Cons • Must seperatly manage schema • Adding columns & indexes to existing large tables can be painful & slow
27. ACID • Properties that guarante a database transaction are processed reliably • Atomic • Consistent • Isolated • Durable
28. ACID • Atomic • Any database Transaction is all or nothing. • If one part of the transaction fails it all fails “An Incomplete Transaction Cannot Exist”
29. ACID • Consistent • Any transaction will take the database from one consistent state to another “Only Consistent data is allowed to be written”
30. ACID • Isolated • No transaction should be able to interfere with another transaction “the same field cannot be updated by two sources at the exact same time” a = 0 a += 1 a += 2 } a = ??
31. ACID • Durable • Once a transaction Is committed it will stay that way “Save it once, read it forever”
32. What is a Database? • RDBMS • Relational • Flexible • Has a schema • Most likely ACID compliant • Typically fast under low load or when optimized
33. What is SQL? • Structured Query Language • The language databases speak • Based on relational algebra • Insert • Query • Update • Delete “SELECT Company, Country FROM Customers WHERE Country = 'USA' ”
34. Why people <3 SQL • Relational algebra is powerful • SQL is proven • well understood • well documented
35. Why people </3 SQL • Relational algebra Is hard • Different databases support different SQL syntax • Yet another programming language to learn
36. SQL != Database • SQL is used to talk to a RDBMS (database) • SQL is not a RDBMS
37. What is NoSQL? Not A Relational Database
38. RDBMS
39. Types of NoSQL • Distributed Systems • Document Store • Graph Database • Key-Value Store • Eventually Consistent Systems Mix And Match ↑
40. Key Value Stores • Non Relational • Typically No Schema • Map one Key (a string) to a Value (some object) Example: Redis
41. Key Value Example redis = Redis.new redis.set(“foo”, “bar”) redis.get(“foo”) >> “bar”
42. Key Value Example redis = Redis.new redis.set(“foo”, “bar”) redis.get(“foo”) >> “bar” Key Value Key Value
43. Key Value • Like a databse that can only ever use primary Key (id) YES select * from users where id = ‘3’; NO select * from users where name = ‘schneems’;
44. NoSQL @ Gowalla • Redis (key-value store) • Store “Likes” & Analytics • Memcache (key-value store) • Cache Database results • Cassandra • (eventually consistent, with-schema, key value store) • Store “feeds” or “timelines” • Solr (search index)
45. Memcache • Key-Value Store • Open Source • Distributed • In memory (ram) only • fast, but volatile • Not ACID • Memory object caching system
46. Memcache Example memcache = Memcache.new memcache.set(“foo”, “bar”) memcache.get(“foo”) >> “bar”
47. Memcache • Can store whole objects memcache = Memcache.new user = User.where(:username => “schneems”) memcache.set(“user:3”, user) user_from_cache = memcache.get(“user:3”) user_from_cache == user >> true user_from_cache.username >> “Schneems”
48. Memcache @ Gowalla • Cache Common Queries • Decreases Load on DB (postgres) • Enables higher throughput from DB • Faster response than DB • Users see quicker page load time
49. What to Cache? • Objects that change infrequently • users • spots (places) • etc. • Expensive(ish) sql queries • Friend ids for users • User ids for people visiting spots • etc.
50. Memcache Distributed B C A
51. Memcache Distributed B C A Easily add more nodes D
52. Memcache <3’s DB • We use them Together • If memcache doesn’t have a value • Fetch from the database • Set the key from database • Hard • Cache Invalidation : (
53. Redis • Key Value Store • Open Source • Not Distributed (yet) • Extremely Quick • “Data structure server”
54. Redis Example, again redis = Redis.new redis.set(“foo”, “bar”) redis.get(“foo”) >> “bar”
55. Redis - Has Data Types • Strings • Hashes • Lists • Sets • Sorted Sets
56. Redis Example, sets redis = Redis.new redis.sadd(“foo”, “bar”) redis.members(“foo”) >> [“bar”] redis.sadd(“foo”, “fly”) redis.members(“foo”) >> [“bar”, “fly”]
57. Redis => Likeable • Very Fast response • ~ 50 queries per page view • ~ 1 ms per query • http://github.com/Gowalla/likeable
58. Cassandra • Open Source • Distributed • Key Value Store • Eventually Consistent • Sortof not ACID • Uses A Schema • ColumnFamilies
59. Cassandra Distributed B C A Eventual Consistency D Data In Copied To Extra Nodes ... Eventually
60. Cassandra @ Gowalla{ Activity Feeds
61. Cassandra @ Gowalla • Chronologic • http://github.com/Gowalla/chronologic
62. Should I use NoSQL?
63. Which One?
64. Pick the right tool
65. Tradeoffs • Every Data store has them • Know your data store • Strengths • Weaknesses
66. NoSQL vs. RDBMS • No Magic Bullet • Use Both!!! • Model data in a datastore you understand • Switch to when/if you need to • Understand Your Options
67. Questions? Richard Schneeman @schneems works for @Gowalla