1. Introduction to cloud computing Jiaheng Lu Department of Computer Science Renmin University of China www.jiahenglu.net

2. Advanced MapReduce Application Reference: Jimmy Lin http://www.umiacs.umd.edu/~jimmylin/clo ud-2008-Fall/schedule.html

3. Managing Dependencies Remember: Mappers run in isolation You have no idea in what order the mappers run You have no idea on what node the mappers run You have no idea when each mapper finishes Tools for synchronization: Ability to hold state in reducer across multiple key- value pairs Sorting function for keys Partitioner Cleverly-constructed data structures

4. Motivating Example Term co-occurrence matrix for a text collection M = N x N matrix (N = vocabulary size) M ij : number of times i and j co-occur in some context (for concreteness, let’s say context = sentence) Why? Distributional profiles as a way of measuring semantic distance Semantic distance useful for many language processing tasks e.g., Mohammad and Hirst (EMNLP, 2006)

5. MapReduce: Large Counting Problems Term co-occurrence matrix for a text collection = specific instance of a large counting problem A large event space (number of terms) A large number of events (the collection itself) Goal: keep track of interesting statistics about the events Basic approach Mappers generate partial counts Reducers aggregate partial counts

6. First Try: “Pairs” Each mapper takes a sentence: Generate all co-occurring term pairs For all pairs, emit (a, b) → count Reducers sums up counts associated with these pairs Use combiners!

7. “Pairs” Analysis Advantages Easy to implement, easy to understand Disadvantages Lots of pairs to sort and shuffle around (upper bound?)

8. Another Try: “Stripes” Idea: group together pairs into an associative array Each mapper takes a sentence: Generate all co-occurring term pairs (a, b) → 1 (a, c) → 2 (a, d) → 5 (a, e) → 3 (a, f) → 2 a → { b: 1, c: 2, d: 5, e: 3, f: 2 } a → { b: 1, d: 5, e: 3 } a → { b: 1, c: 2, d: 2, f: 2 } a → { b: 2, c: 2, d: 7, e: 3, f: 2 }

9. Another Try: “Stripes” Reducers perform element-wise sum of associative arrays a → { b: 1, d: 5, e: 3 } a → { b: 1, c: 2, d: 2, f: 2 } a → { b: 2, c: 2, d: 7, e: 3, f: 2 } +

10. “Stripes” Analysis Advantages Far less sorting and shuffling of key-value pairs Can make better use of combiners Disadvantages More difficult to implement Underlying object is more heavyweight Fundamental limitation in terms of size of event space

11. Cluster size: 38 cores Data Source: Associated Press Worldstream (APW) of the English Gigaword Corpus (v3), which contains 2.27 million documents (1.8 GB compressed, 5.7 GB uncompressed)

12. Conditional Probabilities How do we compute conditional probabilities from counts? Why do we want to do this? How do we do this with MapReduce?

13. P(B|A): “Pairs” For this to work: Must emit extra (a, *) for every b n in mapper Must make sure all a’s get sent to same reducer (use Partitioner) Must make sure (a, *) comes first (define sort order) (a, b 1 ) → 3 (a, b 2 ) → 12 (a, b 3 ) → 7 (a, b 4 ) → 1 … (a, *) → 32 (a, b 1 ) → 3 / 32 (a, b 2 ) → 12 / 32 (a, b 3 ) → 7 / 32 (a, b 4 ) → 1 / 32 … Reducer holds this value in memory

14. P(B|A): “Stripes” Easy! One pass to compute (a, *) Another pass to directly compute P(B|A) a → {b 1 :3, b 2 :12, b 3 :7, b 4 :1, … }

15. Synchronization in Hadoop Approach 1: turn synchronization into an ordering problem Sort keys into correct order of computation Partition key space so that each reducer gets the appropriate set of partial results Hold state in reducer across multiple key-value pairs to perform computation Approach 2: construct data structures that “bring the pieces together” Each reducer receives all the data it needs to complete the computation

16. Issues and Tradeoffs Number of key-value pairs Object creation overhead Time for sorting and shuffling pairs across the network Size of each key-value pair De/serialization overhead Combiners make a big difference! RAM vs. disk and network Arrange data to maximize opportunities to aggregate partial results

17. Data Types in Hadoop WritableDefines a de/serialization protocol. Every data type in Hadoop is a Writable. WritableComprableDefines a sort order. All keys must be of this type (but not values). IntWritable LongWritable Text … Concrete classes for different data types.

18. Complex Data Types in Hadoop How do you implement complex data types? The easiest way: Encoded it as Text, e.g., (a, b) = “a:b” Use regular expressions to parse and extract data The hard way: Define a custom implementation of WritableComprable Must implement: readFields, write, compareTo Computationally efficient, but slow for rapid prototyping

19. Yahoo ！ PNUTS and Hadoop

20. babycenter epicurious Search Results of the Future yelp.com answers.com LinkedIn webmd Gawker New York Times

21. What’s in the Horizontal Cloud? Common Approaches to QA, Production Engineering, Performance Engineering, Datacenter Management, and Optimization Common Approaches to QA, Production Engineering, Performance Engineering, Datacenter Management, and Optimization ID & Account Management Monitoring & QoS Shared Infrastructure Metering, Billing, Accounting Horizontal Cloud Services Edge Content Services e.g., YCS, YCPI Provisioning & Virtualization e.g., EC2 Batch Storage & Processing e.g., Hadoop & Pig Operational Storage e.g., S3, MObStor, Sherpa Other Services Messaging, Workflow, virtual DBs & Webserving Security Simple Web Service API’s

22. Yahoo! Cloud Stack Provisioning (Self-serve) Horizontal Cloud Services …YCSYCPI Brooklyn EDGE Monitoring/Metering/Security Horizontal Cloud Services …Hadoop BATCH Horizontal Cloud Services …SherpaMOBStor STORAGE Horizontal Cloud Services VM/OS… APP Horizontal Cloud Services VM/OSyApache WEB Data Highway Serving Grid PHPApp Engine

23. Yahoo! CCDI Thrust Areas Fast Provisioning and Machine Virtualization: On demand, deliver a set of hosts imaged with desired software and configured against standard services Multiple hosts may be multiplexed onto the same physical machine. Batch Storage and Processing: Scalable data storage optimized for batch processing, together with computational capabilities Operational Storage: Persistent storage that supports low- latency updates and flexible retrieval Edge Content Services: Support for dealing with network topology, communication protocols, caching, and BCP Rest of today’s talk

24. Web Data Management Large data analysis (Hadoop) Structured record storage (PNUTS/Sherpa) Blob storage (SAN/NAS) Scan oriented workloads Focus on sequential disk I/O $ per cpu cycle CRUD Point lookups and short scans Index organized table and random I/Os $ per latency Object retrieval and streaming Scalable file storage $ per GB

25. The World Has Changed Web serving applications need: Scalability! Preferably elastic Flexible schemas Geographic distribution High availability Reliable storage Web serving applications can do without: Complicated queries Strong transactions

26. PNUTS / SHERPA To Help You Scale Your Mountains of Data

27. Yahoo! Serving Storage Problem Small records – 100KB or less Structured records – lots of fields, evolving Extreme data scale - Tens of TB Extreme request scale - Tens of thousands of requests/sec Low latency globally - 20+ datacenters worldwide High Availability - outages cost $millions Variable usage patterns - as applications and users change 27

28. The PNUTS/Sherpa Solution The next generation global-scale record store Record-orientation: Routing, data storage optimized for low-latency record access Scale out: Add machines to scale throughput (while keeping latency low) Asynchrony: Pub-sub replication to far-flung datacenters to mask propagation delay Consistency model: Reduce complexity of asynchrony for the application programmer Cloud deployment model: Hosted, managed service to reduce app time-to-market and enable on demand scale and elasticity 28

29. E 75656 C A 42342 E B 42521 W C 66354 W D 12352 E F 15677 E What is PNUTS/Sherpa? E 75656 C A 42342 E B 42521 W C 66354 W D 12352 E F 15677 E CREATE TABLE Parts ( ID VARCHAR, StockNumber INT, Status VARCHAR … ) CREATE TABLE Parts ( ID VARCHAR, StockNumber INT, Status VARCHAR … ) Parallel database Geographic replication Structured, flexible schema Hosted, managed infrastructure A 42342 E B 42521 W C 66354 W D 12352 E E 75656 C F 15677 E 29

30. What Will It Become? E 75656 C A 42342 E B 42521 W C 66354 W D 12352 E F 15677 E E 75656 C A 42342 E B 42521 W C 66354 W D 12352 E F 15677 E E 75656 C A 42342 E B 42521 W C 66354 W D 12352 E F 15677 E CREATE TABLE Parts ( ID VARCHAR, StockNumber INT, Status VARCHAR … ) CREATE TABLE Parts ( ID VARCHAR, StockNumber INT, Status VARCHAR … ) Parallel database Geographic replication Indexes and views Structured, flexible schema Hosted, managed infrastructure

31. What Will It Become? E 75656 C A 42342 E B 42521 W C 66354 W D 12352 E F 15677 E E 75656 C A 42342 E B 42521 W C 66354 W D 12352 E F 15677 E E 75656 C A 42342 E B 42521 W C 66354 W D 12352 E F 15677 E Indexes and views

32. Scalability Thousands of machines Easy to add capacity Restrict query language to avoid costly queries Geographic replication Asynchronous replication around the globe Low-latency local access High availability and fault tolerance Automatically recover from failures Serve reads and writes despite failures Design Goals 32 Consistency Per-record guarantees Timeline model Option to relax if needed Multiple access paths Hash table, ordered table Primary, secondary access Hosted service Applications plug and play Share operational cost

33. Technology Elements PNUTS Query planning and execution Index maintenance Distributed infrastructure for tabular data Data partitioning Update consistency Replication YDOT FS Ordered tables Applications Tribble Pub/sub messaging YDHT FS Hash tables Zookeeper Consistency service YCA: Authorization PNUTS API Tabular API 33

34. Data Manipulation Per-record operations Get Set Delete Multi-record operations Multiget Scan Getrange Web service (RESTful) API 34

35. Tablets—Hash Table Apple Lemon Grape Orange Lime Strawberry Kiwi Avocado Tomato Banana Grapes are good to eat Limes are green Apple is wisdom Strawberry shortcake Arrgh! Don’t get scurvy! But at what price? How much did you pay for this lemon? Is this a vegetable? New Zealand The perfect fruit NameDescriptionPrice $12 $9 $1 $900 $2 $3 $1 $14 $2 $8 0x0000 0xFFFF 0x911F 0x2AF3 35

36. Tablets—Ordered Table 36 Apple Banana Grape Orange Lime Strawberry Kiwi Avocado Tomato Lemon Grapes are good to eat Limes are green Apple is wisdom Strawberry shortcake Arrgh! Don’t get scurvy! But at what price? The perfect fruit Is this a vegetable? How much did you pay for this lemon? New Zealand $1 $3 $2 $12 $8 $1 $9 $2 $900 $14 NameDescriptionPrice A Z Q H

37. Flexible Schema Posted dateListing idItemPrice 6/1/07424252Couch$570 6/1/07763245Bike$86 6/3/07211242Car$1123 6/5/07421133Lamp$15 Color Red Condition Good Fair

38. Storage units Routers Tablet Controller REST API Clients Local region Remote regions Tribble Detailed Architecture 38

39. Tablet Splitting and Balancing 39 Each storage unit has many tablets (horizontal partitions of the table) Tablets may grow over time Overfull tablets split Storage unit may become a hotspot Shed load by moving tablets to other servers Storage unit Tablet

40. QUERY PROCESSING 40

41. Accessing Data 41 SU 1 Get key k 2 3 Record for key k 4

42. Bulk Read 42 SU Scatter/ gather server SU 1 {k1, k2, … kn} 2 Get k 1 Get k 2 Get k 3

43. Storage unit 1Storage unit 2Storage unit 3 Range Queries in YDOT Clustered, ordered retrieval of records Storage unit 1 Canteloupe Storage unit 3 Lime Storage unit 2 Strawberry Storage unit 1 Router Apple Avocado Banana Blueberry Canteloupe Grape Kiwi Lemon Lime Mango Orange Strawberry Tomato Watermelon Apple Avocado Banana Blueberry Canteloupe Grape Kiwi Lemon Lime Mango Orange Strawberry Tomato Watermelon Grapefruit…Pear? Grapefruit…Lime? Lime…Pear? Storage unit 1 Canteloupe Storage unit 3 Lime Storage unit 2 Strawberry Storage unit 1

44. Updates 1 Write key k 2 7 Sequence # for key k 8 SU 3 Write key k 4 5 SUCCESS 6 Write key k Routers Message brokers 44

45. ASYNCHRONOUS REPLICATION AND CONSISTENCY 45

46. Asynchronous Replication 46

47. Goal: Make it easier for applications to reason about updates and cope with asynchrony What happens to a record with primary key “Alice”? Consistency Model 47 Time Record inserted Update Delete Time v. 1 v. 2 v. 3v. 4 v. 5 v. 7 Generation 1 v. 6 v. 8 Update As the record is updated, copies may get out of sync.

48. Example: Social Alice UserStatus AliceBusy West East UserStatus AliceFree UserStatus Alice??? UserStatus Alice??? UserStatus AliceBusy UserStatus Alice___ Busy Free Record Timeline

49. Time v. 1 v. 2 v. 3v. 4 v. 5 v. 7 Generation 1 v. 6 v. 8 Current version Stale version Read Consistency Model 49 In general, reads are served using a local copy

50. Time v. 1 v. 2 v. 3v. 4 v. 5 v. 7 Generation 1 v. 6 v. 8 Read up-to-date Current version Stale version Consistency Model 50 But application can request and get current version

51. Time v. 1 v. 2 v. 3v. 4 v. 5 v. 7 Generation 1 v. 6 v. 8 Read ≥ v.6 Current version Stale version Consistency Model 51 Or variations such as “read forward”—while copies may lag the master record, every copy goes through the same sequence of changes

52. Time v. 1 v. 2 v. 3v. 4 v. 5 v. 7 Generation 1 v. 6 v. 8 Write Current version Stale version Consistency Model 52 Achieved via per-record primary copy protocol (To maximize availability, record masterships automaticlly transferred if site fails) Can be selectively weakened to eventual consistency (local writes that are reconciled using version vectors)

53. Time v. 1 v. 2 v. 3v. 4 v. 5 v. 7 Generation 1 v. 6 v. 8 Write if = v.7 ERROR Current version Stale version Consistency Model 53 Test-and-set writes facilitate per-record transactions

54. Consistency Techniques Per-record mastering Each record is assigned a “master region” May differ between records Updates to the record forwarded to the master region Ensures consistent ordering of updates Tablet-level mastering Each tablet is assigned a “master region” Inserts and deletes of records forwarded to the master region Master region decides tablet splits These details are hidden from the application Except for the latency impact!

55. 55 Mastering A 42342 E B 42521 W C 66354 W D 12352 E E 75656 C F 15677 E A 42342 E B 42521 W C 66354 W D 12352 E E 75656 C F 15677 E A 42342 E B 42521 W C 66354 W D 12352 E E 75656 C F 15677 E Tablet master

56. Bulk Insert/Update/Replace Client Source Data Bulk manager 1.Client feeds records to bulk manager 2.Bulk loader transfers records to SU’s in batches Bypass routers and message brokers Efficient import into storage unit

57. Bulk Load in YDOT YDOT bulk inserts can cause performance hotspots Solution: preallocate tablets

58. Index Maintenance How to have lots of interesting indexes and views, without killing performance? Solution: Asynchrony! Indexes/views updated asynchronously when base table updated

59. SHERPA IN CONTEXT 59

60. Types of Record Stores Query expressiveness Simple Feature rich Object retrieval Retrieval from single table of objects/records SQL S3 PNUTS Oracle

61. Types of Record Stores Consistency model Best effort Strong guarantees Eventual consistency Timeline consistency ACID S3 PNUTS Oracle Program centric consistency Object-centric consistency

62. Types of Record Stores Data model Flexibility, Schema evolution Optimized for Fixed schemas CouchDB PNUTS Oracle Consistency spans objects Object-centric consistency

63. Types of Record Stores Elasticity (ability to add resources on demand) Inelastic Elastic Limited (via data distribution) VLSD (Very Large Scale Distribution /Replication) Oracle PNUTS S3

64. Data Stores Comparison User-partitioned SQL stores Microsoft Azure SDS Amazon SimpleDB Multi-tenant application databases Salesforce.com Oracle on Demand Mutable object stores Amazon S3 Versus PNUTS More expressive queries Users must control partitioning Limited elasticity Highly optimized for complex workloads Limited flexibility to evolving applications Inherit limitations of underlying data management system Object storage versus record management

65. Application Design Space Records Files Get a few things Scan everything Sherpa MObStor Everest Hadoop YMDB MySQL Filer Oracle BigTable 65

66. Alternatives Matrix Elastic Operability Global low latency Availability Structured access Sherpa Y! UDB MySQL Oracle HDFS BigTable Dynamo Updates Cassandra Consistency model SQL/ACID 66

67. Further Reading Efficient Bulk Insertion into a Distributed Ordered Table (SIGMOD 2008) Adam Silberstein, Brian Cooper, Utkarsh Srivastava, Erik Vee, Ramana Yerneni, Raghu Ramakrishnan PNUTS: Yahoo!'s Hosted Data Serving Platform (VLDB 2008) Brian Cooper, Raghu Ramakrishnan, Utkarsh Srivastava, Adam Silberstein, Phil Bohannon, Hans-Arno Jacobsen, Nick Puz, Daniel Weaver, Ramana Yerneni Asynchronous View Maintenance for VLSD Databases, Parag Agrawal, Adam Silberstein, Brian F. Cooper, Utkarsh Srivastava and Raghu Ramakrishnan SIGMOD 2009 (to appear) Cloud Storage Design in a PNUTShell Brian F. Cooper, Raghu Ramakrishnan, and Utkarsh Srivastava Beautiful Data, O’Reilly Media, 2009 (to appear)

68. QUESTIONS? 68