Protocol Buffers and Hadoop at Twitter

Hadoop and Protocol Buffers at Twitter Kevin Weil -- @kevinweil Analytics Lead, Twitter TM

Outline ‣ Problem Statement ‣ CSV? XML? JSON? Regex? ‣ Protocol Buffers ‣ Codegen, Hadoop and You ‣ Applications ‣ Conclusions and Next Steps

My Background ‣ Studied Mathematics and Physics at Harvard, Physics at Stanford ‣ Tropos Networks (city-wide wireless): mesh routing algorithms, GBs of data ‣ Cooliris (web media): Hadoop and Pig for analytics, TBs of data ‣ Twitter: Hadoop, Pig, HBase, large-scale data analysis and visualization, social graph analysis, machine learning, lots more data

The Challenge ‣ Store some tweets

The Challenge ‣ Store some tweets Store 100 billion tweets

The Challenge ‣ Store 100 billion tweets in a way that is ‣ Robust to changes

The Challenge ‣ Store 100 billion tweets in a way that is ‣ Robust ‣ Efficient in size and speed

The Challenge ‣ Store 100 billion tweets in a way that is ‣ Robust ‣ Efficient ‣ Amenable to large-scale analysis

The Challenge ‣ Store 100 billion tweets in a way that is ‣ Robust ‣ Efficient ‣ Amenable to large-scale analysis ‣ Reusable (especially for other classes of data, like logs, where the size gets really large)

The System ‣ Your (friend’s) hadoop cluster

The Data ‣ kevin@tw-mbp-kweil ~ $ curl http:// ‣ ‣ <?xml version="1.0" encoding="UTF-8"?> <status> api.twitter.com/1/statuses/show/9225259353.xml ‣ <created_at>Wed Feb 17 08:01:13 +0000 2010</created_at> ‣ <id>9225259353</id> ‣ <text>Preparing slides for tomorrow's talk at Y! at the Hadoop User Group: Protobufs and Hadoop at Twitter. See you there? http://bit.ly/9DJcd9</text> ‣ <source><a href="http://www.tweetdeck.com/" rel="nofollow">TweetDeck</a></source> ‣ <truncated>false</truncated> ‣ <in_reply_to_status_id></in_reply_to_status_id> <in_reply_to_user_id></in_reply_to_user_id> Each tweet has 12 fields, 3 of which (user, geo, ‣ ‣ ‣ <favorited>false</favorited> ‣ <in_reply_to_screen_name></in_reply_to_screen_name> ‣ <user> contributors) have subfields ‣ <id>3452911</id> ‣ <name>Kevin Weil</name> ‣ <screen_name>kevinweil</screen_name> ‣ <location>Portola Valley, CA</location> ‣ <description>Analytics Lead at Twitter. Ultra-marathons, cycling, hadoop, lolcats.</description> ‣ <profile_image_url>http://a3.twimg.com/profile_images/220257539/n206489_34325699_8572_normal.jpg</profile_image_url> ‣ <url></url> ‣ <protected>false</protected> ‣ <followers_count>3122</followers_count> ‣ <profile_background_color>B2DFDA</profile_background_color> ‣ <profile_text_color>333333</profile_text_color> ‣ It can change as we add new features ‣ <profile_link_color>93A644</profile_link_color> ‣ <profile_sidebar_fill_color>ffffff</profile_sidebar_fill_color> ‣ <profile_sidebar_border_color>eeeeee</profile_sidebar_border_color> ‣ <friends_count>436</friends_count> ‣ <created_at>Wed Apr 04 19:29:46 +0000 2007</created_at> ‣ <favourites_count>721</favourites_count> ‣ <utc_offset>-28800</utc_offset> ‣ <time_zone>Pacific Time (US & Canada)</time_zone> ‣ <profile_background_image_url>http://s.twimg.com/a/1266345225/images/themes/theme13/bg.gif</profile_background_image_url> ‣ <profile_background_tile>false</profile_background_tile> ‣ <notifications>false</notifications> ‣ <geo_enabled>true</geo_enabled> ‣ <verified>false</verified> ‣ <following>false</following> ‣ <statuses_count>2556</statuses_count> ‣ <lang>en</lang> ‣ <contributors_enabled>false</contributors_enabled> ‣ </user> ‣ <geo/> ‣ <contributors/> ‣ </status> ‣

The Requirements ‣ Splittability ‣ Parsing efficiency ‣ Reusability ‣ Ability to add new fields ‣ Ability to ignore unused fields ‣ Small data size ‣ Hierarchical

Common Formats Parsing Ignore unused Splittable Reusability Add new fields Small data size Hierarchical efficiency fields XML JSON CSV Custom regex (Apache)

Enter Protocol Buffers ‣ “Protocol Buffers are a way of encoding structured data in an efficient yet extensible format. Google uses Protocol Buffers for almost all of its internal RPC protocols and file formats.” ‣ http://code.google.com/p/protobuf ‣ You write IDL describing your data structure ‣ It generates code in your languages of choice to construct, serialize, deserialize, reflect across, etc, your data structure ‣ Like Thrift, but richer and more efficient (except no RPC) ‣ Avro is an exciting up-and-coming alternative

Protobuf IDL Example ‣ message Status { ‣ optional string created_at = 1; ‣ optional int64 id = 2; ‣ optional string text = 3; ‣ optional string source = 4; ‣ optional bool truncated = 5; ‣ optional int64 in_reply_to_status_id = 6; ‣ optional int64 in_reply_to_user_id = 7; ‣ optional bool favorited = 8; ‣ optional string in_reply_to_screen_name = 9; ‣ optional message User = 10; ‣ optional message Geo = 11; ‣ optional message Contributors = 12; ‣ message User { ‣ optional int64 id = 1; ‣ optional string name = 2; ‣ ... ‣ } ‣ message Geo { ... } ‣ message Contributors { ... } ‣ }

Protobuf Generated Code ‣ The generated code is: ‣ Efﬁcient (Google quotes 80x vs. |-delimited format)1,2 ‣ Extensible ‣ Backwards compatible ‣ Polymorphic (in Java, C++, Python) ‣ Metadata-rich 1. http://cacm.acm.org/magazines/2010/1/55744-mapreduce-a-ﬂexible-data-processing-tool/fulltext 2. http://code.google.com/p/thrift-protobuf-compare/wiki/Benchmarking

Common Formats Parsing Ignore unused Splittable Reusability Add new fields Small data size Hierarchical efficiency fields XML JSON CSV Custom regex (Apache) Protocol Buffers

But Wait, There’s More ‣ Codegen for data structures is nice... ‣ Next step: codegen for all Hadoop-related code

But Wait, There’s More ‣ Codegen for data structures is nice... ‣ Next step: codegen for all Hadoop-related code ‣ Protocol Buffer InputFormats

But Wait, There’s More ‣ Codegen for data structures is nice... ‣ Next step: codegen for all Hadoop-related code ‣ Protocol Buffer InputFormats ‣ OutputFormats

But Wait, There’s More ‣ Codegen for data structures is nice... ‣ Next step: codegen for all Hadoop-related code ‣ Protocol Buffer InputFormats ‣ OutputFormats ‣ Writables

But Wait, There’s More ‣ Codegen for data structures is nice... ‣ Next step: codegen for all Hadoop-related code ‣ Protocol Buffer InputFormats ‣ OutputFormats ‣ Writables ‣ Pig LoadFuncs and StoreFuncs

But Wait, There’s More ‣ Codegen for data structures is nice... ‣ Next step: codegen for all Hadoop-related code ‣ Protocol Buffer InputFormats ‣ OutputFormats ‣ Writables ‣ Pig LoadFuncs and StoreFuncs ‣ Cascading, Streaming, Dumbo, etc

But Wait, There’s More ‣ Codegen for data structures is nice... ‣ Next step: codegen for all Hadoop-related code ‣ Protocol Buffer InputFormats ‣ OutputFormats ‣ Writables ‣ Pig LoadFuncs and StoreFuncs ‣ Cascading, Streaming, Dumbo, etc ‣ Per Protocol Buffer

Protocol Buffer InputFormats ‣ All objects (hierarchical data, inheritance, etc) ‣ All automatically generated ‣ Efﬁcient, extensible storage and serialization

Pig LoadFuncs ‣ All objects (hierarchical data, inheritance, etc) ‣ All automatically generated ‣ Even the load statement itself is codegen

Where do these work? ‣ Java MapReduce APIs (InputFormats, OutputFormats, Writables) ‣ Deprecated Java MapReduce APIs (same) ‣ Enables Streaming, Dumbo, Cascading ‣ Pig ‣ HBase

Counting Big Data ‣ standard counts, min, max, std dev ‣ How many requests do we serve in a day? ‣ What is the average latency? 95% latency? ‣ Group by response code. What is the hourly distribution? ‣ How many searches happen each day on Twitter? ‣ How many unique queries, how many unique users? ‣ What is their geographic distribution?

Correlating Big Data ‣ probabilities, covariance, influence ‣ How does usage differ for mobile users? ‣ How about for users with 3rd party desktop clients? ‣ Cohort analyses ‣ Site problems: what goes wrong at the same time? ‣ Which features get users hooked? ‣ Which features do successful users use often? ‣ Search corrections, search suggestions ‣ A/B testing

Research on Big Data ‣ prediction, graph analysis, natural language ‣ What can we tell about a user from their tweets? ‣ From the tweets of those they follow? ‣ From the tweets of their followers? ‣ From the ratio of followers/following? ‣ What graph structures lead to successful networks? ‣ User reputation

Research on Big Data ‣ prediction, graph analysis, natural language ‣ Sentiment analysis ‣ What features get a tweet retweeted? ‣ How deep is the corresponding retweet tree? ‣ Long-term duplicate detection ‣ Machine learning ‣ Language detection ‣ ... the list goes on.

Resolution ‣ All we do now is write IDL for the data schema ‣ Get efﬁcient, forward/backwards compatible, splittable data structures automatically generated for us ‣ Get loaders, input formats, output formats, writables, and schemas automatically generated for us ‣ Helps the Twitter analytics team stay agile ‣ Can handle new, complex data without the need for new code, new tests, new bugs ‣ Focus on the analysis, not data formats

Twitter Open Source ‣ Coming soon! (1-2 weeks) http://github.com/kevinweil ‣ All base classes for InputFormats, OutputFormats, Writables, Pig Loaders, etc ‣ For new and deprecated MapReduce API ‣ With and without LZO compression (see http://github.com/ kevinweil/hadoop-lzo) ‣ Protobuf reﬂection helpers ‣ Serialized block storage format for HDFS

Questions? Follow me at twitter.com/kevinweil ‣ If this sounded interesting to you -- that’s because it is. And we’re hiring. TM