Bigtable: A Distributed Storage System for Structured Data

1
Bigtable A Distributed Storage System for
Structured Data

• by Fay Chang, Jeffrey Dean, Sanjay Ghemawat,
  Wilson C. Hsieh, Deborah A. Wallach, Mike
  Burrows, Tushar Chandra, Andrew Fikes, and Robert
  E. Gruber
• Google, Inc.

2
Bigtable Defined

• Bigtable is a data model that is
• Scalable
• Distributed
• Flexible
• High-performance
• Highly available

3
Bigtable as a database

• Multidimensional sorted map
• Indexed row, column, timestamp
• eg. languageID
• Data as uninterpreted byte strings
• Bigtable API

Figure from paper
4
Bigtable API

• Clients can
• read rows and row subsets
• write/delete (atomic single-row transactions)
• run scripts in server space (Sawzall)
• Create/delete tables/column families
• Metadata for clusters, tables, column families

5
Google Infrastructure

• Cluster management
• Google File System (GFS)
• Chubby
• Google SSTable file format

6
Chubby

• A highly available, persistent distributed lock
  service
• five replicas with one master
• live as long as a majority can communicate
• Tasks
• ensures that there only one active server master
• discovers data servers finalize deaths
• stores Bigtable schema information

7
Bigtable SSTable

8
Bigtable Components
master

• Library linked into client
• Tablet Server
• manages tablets
• handles client read/writes
• Master Server
• manages tablet servers
• tablet assignment

M
1
2
n
tablet servers
9
Bigtable Tablet
Figure from paper

SSTable
10
Tablet Assignment
master

Chubby
M
TS1 TS2 TSn M
1
2
n
tablet servers
11
Tablet Serving
Figure from paper

12
Compactions
Figure from paper

• Minor Compaction
• Merging Compaction

13
Refinements

• Locality groups
• Compression
• Caching
• Bloom filters
• Commit-log implementation
• Tablet Recovery
• Exploiting SSTable immutability

14
Google Earth

15
Google Earth

• One table to preprocesses data
• handles raw images
• holds 70 terabytes of data
• 1 row per geographic segment
• Set of tables to serve client data
• provides image location
• small 500 GB
• in memory columns to reduce latency

16
Lessons

• Large distributed systems many types of failure
• Delay adding new features
• Need for system-level monitoring
• Value of simple design
• code and design clarity are of immense help in
  code maintenance and debugging."

17
Bigtable Development Notes

• As of August 2006
• 388 production Bigtable clusters
• 24,500 combined total tablet servers
• 2.5-year effort
• 7 person years of design implementation
• 100K lines of code
• Used by 60 Google projects

18
References

• http//www.usenix.org/events/osdi06/
• http//www.usenix.org/events/osdi06/tech/chang.htm
  l
• https//www2.blogger.com/comment.g?blogID36203843
  postID116234368988171763