Google云计算解决方案7_Google的其他计算技术

– New constraint: data validation must be performed by application layer!
© Spinnaker Labs, Inc.
Logical Data Representation
• Rows & columns identified by arbitrary strings • Multiple versions of a (row, col) cell can be accessed through timestamps
© Spinnaker Labs, Inc.
Google‘s Needs
• Data reliability • High speed retrieval • Storage of huge numbers of records (several TB of data) • (Multiple) past versions of records should be available
© Spinnaker Labs, Inc.
From Needs to Constraints
• Simplified data retrieval mechanism
– (row, col, timestamp) value lookup, only – No relational operators
• Determine which tablet server should hold a given (new) tablet • Interface with GFS to garbage collect stale SSTable files • Detect tablet server failures/resumption and load balance accordingly
© Spinnaker Labs, Inc.
Physical Data Representation
• SSTable file provides immutable keyvalue map with an index over all keys mapping keydisk block
– Index stored in RAM; value lookup involves only one disk seek to disk block
– Family name is a prefix on column name – e.g., ―fileattr:owning_group‖, ―fileattr:owning_user‖, etc.
• Permissions can be applied at family level to grant read/write access to different applications • Members of a family compressed together
– If a tablet grows beyond a certain size, it is split into two new tablets
© Spinnaker Labs, Inc.
Network Interface
• One master server
– Communicates only with tablet servers
• Any type of data can be stored in any column
– Bigtable sees only byte strings of arbitrary length
© Spinnaker Labs, Inc.
Consistency
• Multiple operations on a single row can be grouped together and applied atomically
BigTable
© Spinnaker Labs, Inc.
A Conventional Database…
• Data structure:
– arbitrary ## of rows – Fixed number and type of columns
• Supports search based on values in all cells • Supports synthesis of output reports based on multiple tables (relational operators)
Google Cluster Computing Faculty Training Workshop
Module VII: Other Google Technologies
© Spinnaker Labs, Inc.
Overview
• BigTable • Chubby
© Spinnaker Labs, Inc.
Version Control
• Cell versions stored most-recent first for faster access to more recent data • Two version expiration policies available:
– Retain last n copies – Retain data for n time units
© Spinnaker Labs, Inc.
No Data Validation
• Any number of columns can be stored in a row within the pre-defined families
– Database will not enforce existence of any minimum set of columns
GFS Chunkserver
Physical layout:
SSTable
SSTable
SSTable
SSTable
SSTable
SSTable
(replica) SSTable
(replica) SSTable
© Spinnaker Labs, Inc.
Master Responsibilities
• Atomic updates only possible at row level
© Spinnaker Labs, Inc.
But Some Additional Flexibility…
• Arbitrary number of columns per row • Arbitrary data type for each column
– No multi-row mutation operators available
• User can specify timestamp to apply to data or allow Bigtable to use ‗now()‘ function
© Spinnaker Labs, Inc.
© Spinnaker Labs, Inc.
Tablet Server Failure
Head node Chubby Server
Tablet server
Tablet server
Logical view:
Tablet
Tablet
Tablet
Tablet
Tablet
Tablet
GFS Chunkserver
• Several tablet servers
– Perform actual client accesses
• ―Chubby‖ lock server provides coordination and mutual exclusion • GFS servers provide underlying storage
© Spinnaker Labs, Inc.
Physical Representation (2)
• A logical ―table‖ is divided into multiple tablets
– Each tablet is one or more SSTable files
• Each tablet stores an interval of table rows
© Spinnaker Labs, Inc.
Reasonable Questions
• Are structured queries necessary? • Can data be organized such that related data is physically close by nature? • What is the minimum coordination required to retrieve data? • Can existing components be leveraged to provide reliability and abstraction?
© Spinnaker Labs, Inc.
Implementation
• Uses several other Google components:
– GFS provides reliable low-level storage for table files, metadata, and logs – Chubby provides distributed synchronization – Designed to easily interface with MapReduce
GFS Chunkserver
Physical layout: