Versioning Maintenance Algo'

1
Versioning Maintenance Algo.

• CS2002
• Spring 2004

2
View Holder
3
Data Structure

• Tuple Version List(TVL)

4
Example 1
5
Example 1
6
(No Transcript)
7
(No Transcript)
8
(No Transcript)
9
View Holder Roles

• Holder-as-Proxy
• View holder only stores the latest version
• Holder-as-Buffer
• View holder has a full replication of the views
  currently in use
• Holder-as-Cache
• View holder maintains a superview of the info
  required by MH
• Our example is in the Holder-as-Cache role

10
Garbage Collection

• Proposition 1
• If a tuple does not belong to a version i then it
  will not belong to any version j for all j gt i
• We keep a record of MaximumVersionNumber.
• If MaximumVersionNumber lt VN and the tuple is
  not a part of the version VN
• Then we can delete the tuple

11
Delete a Version
If MH doesnt need version 2, then version 4 will
be the oldest available version. Although
MaximumVersionNumber 1 and VN 4,
MaximumVersionNumber lt VN, however, we cannot
delete version 1 since it is a part of version 4.
If version 1 is not required by version 4, we
can remove it safely.
12
Life cycle of a tuple
MVNL support multiple versions by using Time
Travel
deletion timestamp
insertion timestamp
13
Example 2
1. Insert an order with orderkey 12345
2. Update to this entry from pending to completed
3. Delete the entry
14
Mapping logical to physical operations
is a private version number counter that is used
by the maintenance process
15
Garbage collection

• Delete the tuples that have been marked as
  deleted, but do not belong to a version that is
  currently being used by a reader
• Delete from R where Tmax lt Tkill
• Tkill is the highest version number that is not
  being accessed by any reader

16
From 2VNL to nVNL

• They are all horizontal redundancy
• Modification of relation schema is similar
• Allocate space for 1 vs n-1 versions in every
  tuple

17
From Horizontal to Vertical

• MVNL is vertical redundancy
• Both enable user queries to run while the
  warehouse is being updated
• MVNL, no limit to the number of maintenance
  transactions a query can overlap with during its
  execution

Concurrent Execution MVNL
18
What can we do with MVNL?
We are here!
19

• In view holder, one has to have knowledge of the
  application domain and decide on the set of
  updateable attributes
• CREATE TABLE Fixed_Attributes(tupleID, city,
  product)
• CREATE TABLE Updatable_Attributes(tupleID, VN,
  total_sales)
• CREATE TABLE Latest(operation, city, product,
  data, total_sales)

20

• From 2VNL to MVNL?
• In 2VNL
• When a maintenance transaction is updating the
  tuples, only the current version held in
  total_sales is available to readers
• When there is no maintenance transaction
  executing, both the previous and current versions
  can be read

21

• Can we just send modified attribute to clients?
• Keep LatestInUsed flag
• Storage overhead
• Concurrency performance
• Garbage collection