C-Store: Data Model and Data Organization Jianlin Feng School of Software SUN YAT-SEN UNIVERSITY May 17, 2010
Jan 05, 2016
C-Store: Data Model and Data Organization
Jianlin FengSchool of SoftwareSUN YAT-SEN UNIVERSITYMay 17, 2010
Data Model: standard relational model A database:
A set of named relations (tables) A relation
A set of named attributes (columns) Primary key
A set of attributes whose values uniquely identify a row (tuple) in a relation.
Foreign key References a primary key in another relation.
Primary Keys and Foreign Keys
How to Implement a logical relation? Given a logical relation R Row Store
R has a direct physical correspondence.
Column Store (C-Store) R may not have a direct physical correspondence. Covered by a set of C-Store Projections.
C-Store Projection
Is anchored on a given logical table T.
Contains one or more attributes from T.
Can also contain any number of attributes from other tables. Attributes from other tables must be referenced by
a sequence of n:1 (i.e., foreign keys) relationships from T.
How to Construct a Projection? Given a logical table T Extract the attributes of interest from T.
Retain any duplicate rows In standard Projection, duplicates are deleted.
Obtain the attributes of interest from other tables Perform foreign-key joins.
Attribute from DEPT table
Sort Key
Tuples in a projection are stored column-wise. K attributes K columns Each column is sorted on the same Sort Key.
Sort Key Can be any column or columns in a projection
Examples of Sort Key
Each Projection is horizontally partitioned into Segments
Horizontal Partition of a Projection Each projection is horizontally partitioned into 1 or
more segments.
SID > 0, Segment identifier
Value-based partitioning on the Sort Key of a projection Each segment has a key range of the sort key. The set of all key ranges partitions the whole space of sort
key.
How to Re-construct a Complete Tuple? To answer any SQL query
There must be a covering set of projections for every logical table T. Every column in T is stored in at least one projection.
To re-construct a tuple, Join segments from different projections Using Storage Keys and Join Indexes
Storage Keys
A segment, horizontal fragment of a projection May involve several coulmns Associate every data value of each column with a
Storage Key, SK. In the same segment, data values from different
columns with matching SK belong to the same logical tuple.
A SK is simply the ID of a logical tuple.
Architecture of C-Store (Vertica)On a Single Node
Storage Keys in Read Store (RS) SKs are numbered 1,2 ,3,...
SKs are not physically stored But are inferred from a tuple’s physical position in
a column.
Why SKs can be inferred? All the columns in the same segment are sorted
on the same Sort Key.
Storage Keys in Write Store (WS) SKs are physically stored
SKs are represented as integers.
Each SK in WS is larger than the largest SK for any segment in RS.
join indexes: A Mapping Table Suppose T1 and T2 are two projections that
cover a logical table T.
For each segment of T1 , build a join index to T2 is a Table with rows :
(s: SID in T2, k: Storage Key in Segment s)
Ordering of Data in Each Column Self-order
the column is ordered by values in itself. Foreign-order
the column is ordered by values in other column in the same projection. A projection may involve several columns
Example Projection EMP1(name, age| age) Column age is Self-order Column name is foreign order
Column Type in Read Store(RS) Type 1: Self-order, few distinct values
Type 2: Foreign-order, few distinct values
Type 3: Self-order, many distinct values
Type 4: Foreign-order, many distinct values
Compression of Columns in Type 1: Self-order, few distinct values Represented by a sequence of triples (v, f, n) v: a data value stored in the column f : the position in the column where v first
appears. n: the number of times v appears in the column
Example: 4’s appear in positions 12-18, is recorded as (4, 12, 7)
One triple is required for each distinct value in the column.
Fast Search over Columns in Type 1 Using Clustered B-Tree (i.e., Primary B-Tree)
Densepack the B-Tree No on-line updates in RS
Using large disk blocks 64-128K The height of this index can be kept small (2 or
less)
Compression of Columns in Type 2: Foreign-order, few distinct values
Represented by a sequence of tuples (v, b)
v: a data value stored in the column
b: a bitmap indicating the positions in which v is stored.
Use Run-Length Encoding to compress bitmaps.
Column C1
V=0 V=1 V=2
0 1 0 0
0 1 0 0
1 0 1 0
1 0 1 0
2 0 0 1
1 0 1 0
0 1 0 0
2 0 0 1
1 0 1 0
Fast Search over Columns in Type 2 To efficiently find the i-th value in a column, Use Offset indexes
B-Trees that map positions in a column to the values contained in that column.
Compression of Columns in Type 3: Self-order, many distinct values Basic Idea
Represent each value in the column as a delta from the previous value.
Use block (64-128K) as the unit of compression.
Like compressing an Inverted Index
Column C1 Delta representation
1 1 (starting value)
4 3
7 3 (delta)
7 0 (delta)
8 1 (delta)
12 4 (delta)
Fast Search over Columns in Type 3 Use densepacked B-Tree at the block-level
A block is viewed as a tuple.
Compression of Columns in Type 4: Foreign-order, many distinct values Two choices at the moment
Do not compress Use densepacked B-Tree
Open for further research
Write Store (WS)
WS is also a Column Store To avoid writing two query optimizers
Implements the identical physical DBMS design as RS.
The Major Difference due to Updates in WS Storage representation
How to Do Updates?
Update Insert a new tuple Modify an existing tuple Delete an (existing) tuple
Update depends on first doing a query Insert: need to check Primary Key Constraint
etc. Modify: need to first find the specified tuple Delete: need to first find the specified tuple
Queries in C-Store
Use some known values of some columns as conditions to search values on other columns Example table:
EMP1(name, age| age) Example query:
use age values to query name values.
Tuples in C-Store are stored column-wise We need to keep track of all column values of the same
logical tuple The Idea: Storage Key as the tuple ID.
How to Speed-up Queries in C-Store? The Basic Idea:
Pre-sort every projection in some SORT KEY order Keep several copies of the same projection, each copy is
pre-sorted in different sort key order. For a given query, use a copy of the projection whose order
matches with the query conditions.
Therefore queries in C-Store usually equal to searches via sort key.
Note: Each Projection in WS has only one copy
Storage Key in WS
The Storage Key, SK, for each tuple is explicitly stored in each WS segment. Columns in WS only keep a logical SORT KEY order via
SKs. Delay physical sorts on SORT KEY order when tuples are
moved from WS to RS.
A unique SK is given to each insert of a logical tuple r in a table T. SK is a serial number generated by system. The SK of r must be recorded in each projection that stores
data for r.
Storage Representation of Columnsin WS Every column in a Projection
Represented as a collection of (v, sk) pairs v : a data value in the column sk : the storage key (explicitly stored)
Build a B-Tree over each column Use the second field of each pair, sk, as the KEY
Sort Keys of Each Projection in WS Represented as a collection of (s, sk) pairs
s : a sort key value sk : the storage key describing where s first
appears.
Build a B-Tree over the (s, sk) pairs Use the first field of each pair, s, as the KEY
Searches via Sort Key: Two Steps 1st Step,use the B-Tree over the (s, sk) pairs
Search condition: known sort key values Search result: corresponding storage
keys
2nd Step, use the B-Tree over the each column (i.e., the (v, sk) pairs) Search condition: storage keys found in 1st Step. Search result: data values in the column.
Partition of a Projection in WS Partitioning a Projection in the same way as in
RS Equals to partitioning the space of sort key 1:1 mapping between RS segments and WS ones.
A tuple is identified by a (sid, storage_key) pair in either RS or WS sid: segment ID Storage_key: the SK of the tuple
General Picture of WS
Usually is very small, can be fully stored in Memory Data processing is fast in memory
Do not compress data, represent data directly
Each projection uses B-Tree indexing to maintain a logical sort-key order. The B-Trees are secondary B-Tree.
References
1. Mike Stonebraker, Daniel Abadi, Adam Batkin, Xuedong Chen, Mitch Cherniack, Miguel Ferreira, Edmond Lau, Amerson Lin, Sam Madden, Elizabeth O'Neil, Pat O'Neil, Alex Rasin, Nga Tran and Stan Zdonik. C-Store: A Column Oriented DBMS , VLDB, 2005. (http://db.csail.mit.edu/projects/cstore/vldb.pdf)
2. VERTICA DATABASE TECHNICAL OVERVIEW WHITE PAPER. http://www.vertica.com/php/pdfgateway?file=VerticaArchitectureWhitePaper.pdf