Facebook Hadoop Data & Applications

Hadoop and Hive at Facebook Data and Applications

Your Company Logo Here

Wednesday, June 10, 2009  Santa Clara Marriott 

Dhruba Borthakur, Ding Zhou

Who generates this data?

Lots of data is generated on Facebook »  200 million active users »  20 million users update their statuses at least

once each day »  More than 850 million photos uploaded to the site

each month »  More than 8 million videos uploaded each month »  More than 1 billion pieces of content (web links,

news stories, blog posts, notes, photos, etc.) shared each week

http://www.slideshare.net/guest5b1607/text-analytics-summit-2009-roddy-lindsay-social-media-happiness-petabytes-and-lols

»  Hadoop/Hive Warehouse ›  4800 cores, 2 PetaBytes

total size

»  Other Hadoop Clusters •  HDFS-Scribe cluster: 320

cores, 160 TB total size •  Hadoop Archival Cluster :

80 cores, 200TB total size •  Test cluster : 800 cores,

150 TB total size

Where do we store parts of this data?

Data Collection using Scribe

Web Servers  Scribe MidTier 

Network Storage and Servers 

Hadoop Hive Warehouse Oracle RAC  MySQL 

Data Collection using Scribe and HDFS

Web Servers 

Scribe MidTier 

RealBme Hadoop Cluster 

Hadoop Hive Warehouse Oracle RAC 

MySQL  Hadoop Scribe Integration

Data Archive: Move old data to cheap storage

Cheap NAS 

Hadoop Archival Cluster 20TB per node 

Hadoop Archive Node 

NFS 

Hive Query 

distcp 

Hadoop Warehouse 

HADOOP‐5048 

Hive User Interfaces

Hive shell access

Hive Web UI

Data Analysis at Facebook

»  Business Intelligence ›  Growth and monetization strategies ›  Product insights & decisions ›  Philosophy: build meta tools and provide easy access to data

»  Artificial Intelligence ›  Recommendation & ranking products ›  Advertising optimization ›  Text analytics ›  Philosophy: model inference; data preparation; model building;

»  Top-level site metrics

BI: Build centralized reporting tools

Bird-view of user growth by countries

Comparing certain metrics between user groups

»  Example: “Find the number of status updates mentioning ‘swine flu’ per day last month”

»  SELECT a.date, count(1) »  FROM status_updates a »  WHERE a.status LIKE “%swine flu%” »  AND a.date >= ‘2009-05-01’ AND a.date <= ‘2009-05-31’ »  GROUP BY a.date

BI: Make AdHoc reporting easy

Build site metric dashboard in a day

»  Data collection: ›  Define metrics and log format (Hive schema) ›  Add logging to the site (Scribe logging) ›  Create a Hive table partitioned by date ›  Set up metric ETL cron job (Hive -> mysql/oracle)

»  Data visualization (using mysql) »  Data access (adhoc query using Hive)

Build Machine Learning Products on Hadoop/Hive

•  Recommendation & ranking •  Advertising optimization •  Text analytics

What applications the user may like

»  Recommend apps based on social and demographic popularity

»  User-app log is huge »  Joining user-app log with

user demographics is difficult

»  Hive for data aggregation

Who the user wants to connect

»  Take existing edges and user feedbacks as labels

»  Build regression models based on user profile and local graph features

»  Too many friends of friends »  Model trained by sampling

»  Hive for model inference »  Hive for feature selection

»  Market research & ad tool

»  Extract popular words from user content

»  Slice by age, gender, region »  Sentiment analysis »  Keyword association

»  Hadoop used for text analytics

What users are talking about (Lexicon)

laid-off

Words associated with vodka

What ads the user might click on

»  Predict user-ad click-through

»  Ads click data is sparse so sampling can miss info

»  Many ML algorithms are iterative thus not easy for hadoop

»  Hadoop for model training

Build ensemble ML models on Hadoop

»  Each mapper trains a number of models

»  Each model output as a intermediate feature

»  Model selection at reducer »  A regression model is built

on selected features

ds1 ds2 ds3 ds4

ensembles

Train models locally Cross-Test models locally

Models assembled by ensemble methods Model inference in a second Hadoop job

In summary

»  So Zuckerberg’s urgent questions are answered; »  So celebrities know where their fans are from; »  So we know one can like vodka and lemonade at the same time; »  It’s fun playing with the data;

Dhruba Borthakur, Ding Zhou

»  Hadoop and Hive at Facebook »  Support product strategy and decision; »  Recommendation & ranking products; »  Advertising optimization; »  Text analytics tools;

dhruba@, dzhou@