1. How Hadoop Revolutionized Data Warehousing at Yahoo and Facebook
Amr Awadallah
CTO, Cloudera, Inc.
August 5, 2009

2. Problems We Wanted to Solve What is Hadoop? HDFS and MapReduce
Outline
Problems We Wanted to Solve
What is Hadoop?
HDFS and MapReduce
Access Languages for Hadoop
Hadoop vs RDBMSes
Conclusion

3. Our Older Systems Limited Raw Data Access
BI / Reports
Ad hoc Queries &
Data Mining
RDBMS (200GB/day)
ETL Grid
Non-Consumption
Filer heads are a bottleneck
Storage Farm for Unstructured Data (20TB/day)
Mostly Append
Collection
Instrumentation

4. We Needed To Be More Agile (part 1)
Data Errors and Reprocessing
We encountered data errors that required reprocessing, which could happen a long time after the fact. “Tape Data” was cost prohibitive to reprocess, we needed to retain raw-data online for long time periods
Conformation Loss
Conversion of data from raw format to conformed dimensions causes some information loss. We needed access to the original data to recover lost information whenever needed (e.g.: a new browser user agent)
Shrinking ETL Window
The storage filers for raw data started becoming a significant bottleneck as large amounts of data needed to be copied to the ETL grid for processing (e.g. 30 hours to process a day’s worth of data)
Ad Hoc Queries on Raw Data
We wanted to run ad hoc queries against the original raw event data, but the storage filers only store and can’t compute

5. We Needed To Be More Agile (part 2)
Data Model Agility: Schema-on-Read vs Schema-on-Write
We wanted to access data even if it had no schema yet, e.g. frequently a new product or feature will launch but we can’t get their dashboards since their schemas weren’t defined yet
Schema-on-Read is slower in terms of machine time (due to read overhead) but it allows us to evolve in an agile way, then we materialize to relational datamarts when data model stabilizes
Consolidated Repository and Ubiquitous Access
We wanted to eliminate borders and have a single repository where anybody can store, join, and process any of our data bits
Beyond Reporting (Data-As-Product)
Last, but not least, we wanted to process the data in ways that feed directly into the product/business (e.g. Spam Filtering, Ad Targeting, Collaborative Filtering, Multimedia Processing)
Data-As-Product is also referred to as Active DW, Operational BI, Online BI, etc.

6. The Solution: A Store-Compute Grid
Interactive Apps
“Batch” Apps
RDBMS
Ad hoc Queries
& Data Mining
ETL and Aggregations
Storage + Computation
Mostly Append
The solution is to *augment* the current RDBMSes with a “smart” storage/processing system. The original event level data is kept in this smart storage layer and can be mined as needed. The aggregate data is kept in the RDBMSes for interactive reporting and analytics.
Collection
Instrumentation

7. What is Hadoop?
A scalable fault-tolerant grid operating system for data storage and processing
Its scalability comes from the marriage of:
HDFS: Self-Healing High-Bandwidth Clustered Storage
MapReduce: Fault-Tolerant Distributed Processing
Operates on unstructured and structured data
A large and active ecosystem (many developers and additions like HBase, Hive, Pig, …)
Open source under the friendly Apache License
The system is self-healing in the sense that it automatically routes around failure. If a node fails then its workload and data are transparently shifted some where else.
The system is intelligent in the sense that the MapReduce scheduler optimizes for the processing to happen on the same node storing the associated data (or co-located on the same leaf Ethernet switch), it also speculatively executes redundant tasks if certain nodes are detected to be slow.
One of the key benefits of Hadoop is the ability to just upload any unstructured files to it without having to “schematize” them first.
You can dump any type of data into Hadoop then the input record readers will abstract it out as if it was structured (i.e. schema on read vs on write)
Open Source Software allows for innovation by partners and customers. It also enables third-party inspection of source code which provides assurances on security and product quality.
1 HDD = 75 MB/sec, 1000 HDDs = 75 GB/sec, the “head of fileserver” bottleneck is eliminated.

8. Hadoop History
: Doug Cutting and Mike Cafarella started working on Nutch (a web-scale crawler-based search system)
: Google publishes GFS and MapReduce papers
2004: Cutting adds DFS & MapReduce support to Nutch
2006: Yahoo! hires Cutting, Hadoop spins out of Nutch
2007: NY Times converts 4TB of archives over 100 EC2s
2008: Web-scale deployments at Y!, Facebook, Last.fm
April 2008: Fastest sort of a TB, 3.5mins over 910 nodes
May 2009:
Fastest sort of a TB, 62secs over 1460 nodes
Sorted a PB in 16.25hours over 3658 nodes
100s of deployments worldwide (
June 2009: Hadoop Summit 2009 – 750 attendees

9. System Shall Manage and Heal Itself Performance Shall Scale Linearly
Hadoop Design Axioms
System Shall Manage and Heal Itself
Performance Shall Scale Linearly
Compute Should Move to Data
Simple Core, Modular and Extensible
Speculative Execution, Data rebalancing, Background Checksumming, etc.

10. HDFS: Hadoop Distributed File System
Block Size = 64MB
Replication Factor = 3
Pool commodity servers in a single hierarchical namespace.
Designed for large files that are written once and read many times.
Example here shows what happens with a replication factor of 3, each data block is present in at least 3 separate data nodes.
Typical Hadoop node is eight cores with 16GB ram and four 1TB SATA disks.
Default block size is 64MB, though most folks now set it to 128MB
Cost/GB is a few ¢/month vs $/month

11. MapReduce: Distributed Processing
Differentiate between MapReduce the platform and MapReduce the programming model. The analogy is similar to the RDBMs which executes the queries, and SQL which is the language for the queries.
MapReduce can run on top of HDFS or a selection of other storage systems
Intelligent scheduling algorithms for locality, sharing, and resource optimization.

12. MapReduce Example for Word Count
cat *.txt | mapper.pl | sort | reducer.pl > out.txt
Split 1
Split i
Split N
Map 1
(docid, text)
Map i
Map M
Reduce 1
Output File 1
(sorted words, sum of counts)
Reduce i
Output File i
Reduce R
Output File R
(words, counts)
(sorted words, counts)
Map(in_key, in_value) => list of (out_key, intermediate_value)
Reduce(out_key, list of intermediate_values) => out_value(s)
Shuffle
“To Be Or Not To Be?”
Be, 5
Be, 12
Be, 7
Be, 6
Be, 30
Think: SELECT word, count(*) FROM documents GROUP BY word
Checkout ParBASH:

13. Hadoop Is More Than Just Analytics/BI
Building the Web Search Index
Processing News/Content Feeds
Content/Ad Targeting Optimization
Fraud Detection and Fighting Spam
Facebook Lexicon: Trends of words on walls
Collaborative Filtering (you might like)
Batch Video/Image Transcoding
Gene Sequence Alignment
Other uses like face recognition, document discovery, OCR, gene sequence alignment, etc.
Data Mining:
** Search and Text Analytics
** Clustering/Categorization
** Modeling/Machine Learning
** Optimization/Operations Research
** Response Prediction/Forecasting
** Simulation, Monte-Carlo like.
** Random Walks of Connectivity Graphs

14. Apache Hadoop Ecosystem
ETL Tools
BI Reporting
RDBMS
Pig (Data Flow)
Hive (SQL)
Sqoop
MapReduce (Job Scheduling/Execution System)
Zookeepr (Coordination)
HBase (Key-Value store)
Avro (Serialization)
HDFS (Hadoop Distributed File System)
HBase: Low Latency Random-Access with per-row consistency for updates/inserts/deletes

15. Hadoop Development Languages
Java MapReduce
Gives the most flexibility and performance, but with a potentially longer development cycle
Streaming MapReduce
Allows you to develop in any language of your choice, but slightly slower performance
Pig
A relatively new data-flow language contributed by Yahoo, suitable for ETL like workloads (procedural multi-stage jobs)
Hive
A SQL warehouse on top of MapReduce (contributed by Facebook). It has two main components:
A meta-store which keeps the schema for files, and
An interpreter which converts the SQL query into MapReduce
First bullet is like assembly, then it gets higher level from there.

16. Hive Features A subset of SQL covering the most common statements
Agile data types: Array, Map, Struct, and JSON objects
User Defined Functions and Aggregates
Regular Expression support
MapReduce support
JDBC support
Partitions and Buckets (for performance optimization)
In The Works: Indices, Columnar Storage, Views, Microstrategy compatibility, Explode/Collect
More details:
Query: SELECT, FROM, WHERE, JOIN, GROUP BY, SORT BY, LIMIT, DISTINCT, UNION ALL
Join: LEFT, RIGHT, FULL, OUTER, INNER
DDL: CREATE TABLE, ALTER TABLE, DROP TABLE, DROP PARTITION, SHOW TABLES, SHOW PARTITIONS
DML: LOAD DATA INTO, FROM INSERT
Types: TINYINT, INT, BIGINT, BOOLEAN, DOUBLE, STRING, ARRAY, MAP, STRUCT, JSON OBJECT
Query: Subqueries in FROM, User Defined Functions, User Defined Aggregates, Sampling (TABLESAMPLE)
Relational: IS NULL, IS NOT NULL, LIKE, REGEXP
Built in aggregates: COUNT, MAX, MIN, AVG, SUM
Built in functions: CAST, IF, REGEXP_REPLACE, …
Other: EXPLAIN, MAP, REDUCE, DISTRIBUTE BY
List and Map operators: array[i], map[k], struct.field

17. Hadoop vs. Relational Databases
A data grid operating system
Stores Files (Unstructured)
Stores 10s of petabytes
Processes 10s of PB/job
Weak Consistency
Scan all blocks in all files
Queries & Data Processing
Batch response (>1sec)
Relational Databases:
An ACID Database system
Stores Tables (Schema)
Stores 100s of terabytes
Processes 10s of TB/query
Transactional Consistency
Lookup rows using index
Mostly queries
Interactive response
Hadoop is good for storing and processing large amounts of unstructured or structured data in batch form (i.e. full table scans)
Hadoop with HBASE (or Hypertable) can do inserts/updates/deletes with reasonable interactive response times (also see Cassandra).

18. Use The Right Tool For The Right Job
Hadoop:
Relational Databases:
Sports car is refined, accelerates very fast, and has a lot of addons/features. But it is pricey on a per bit basis and is expensive to maintain.
Cargo train is rough, missing a lot of functionality, slow to start, but once it gets going it can carry a lot of stuff very economically.

19. Hadoop Criticisms (part 1)
Hadoop MapReduce requires Rocket Scientists
Hadoop has the benefit of both worlds, the simplicity of SQL and the power of Java (or any other language for that matter)
Hadoop is not very efficient hardware wise
Hadoop optimizes for scalability, stability and flexibility versus squeezing every tiny bit of hardware performance
It is cost efficient to throw more “pizza box” servers to gain performance than hire more engineers to manage, configure, and optimize the system or pay 10x the hardware cost in software
Hadoop can’t do quick random lookups
HBase enables low-latency key-value pair lookups (no fast joins)
Hadoop doesn’t support updates/inserts/deletes
Not for multi-row transactions, but HBase enables transactions with row-level consistency semantics
Hadoop is efficient on a cost basis.
Security: Need better integration with systems like LDAP or Kerberos. Also need better isolation against malicious users, though auditing can potentially catch those.

20. Hadoop Criticisms (part 2)
Hadoop isn’t highly available
Though Hadoop rarely loses data, it can suffer from down-time if the master NameNode goes down. This issue is currently being addressed, and there are HW/OS/VM solutions for it
Hadoop can’t be backed-up/recovered quickly
HDFS, like other file systems, can copy files very quickly. It also has utilities to copy data between HDFS clusters
Hadoop doesn’t have security
Hadoop has Unix style user/group permissions, and the community is working on improving its security model
Hadoop can’t talk to other systems
Hadoop can talk to BI tools using JDBC, to RDBMSes using Sqoop, and to other systems using FUSE, WebDAV & FTP

21. Conclusion
Hadoop is a data grid operating system which augments current BI systems and improves their agility by providing an economically scalable solution for storing and processing large amounts of unstructured data over long periods of time

22. twitter.com/awadallah
Contact Information
If you have further questions or comments:
Amr Awadallah
CTO, Cloudera Inc.
twitter.com/awadallah
twitter.com/cloudera

23. Appendix

24. Hadoop High-Level Architecture
Hadoop Client
Contacts Name Node for data or Job Tracker to submit jobs
Name Node
Maintains mapping of file blocks to data node slaves
Job Tracker
Schedules jobs across task tracker slaves
Data Node
Stores and serves blocks of data
The Data Node slave and the Task Tracker slave can, and should, share the same server instance to leverage data locality whenever possible.
Task Tracker
Runs tasks (work units) within a job
Share Physical Node