View Updates

1
View Updates Temporal Correctness in Soft
Real-Time Database Systems
Oral Defence by Cheng Chun Kong (MPhil.,
HKUCSIS) Supervised by Dr. B.C.M. Kao 10th
August, 2000.
2
Contents

• Problems Studied
• Updates Recomputations
• Temporal Correctness
• Transaction Scheduling
• Performance Study
• Future Work

3
I Problems Studied
4
Real-Time Database Systems (RTDB)

• A database system with time constraints
• Transactions can be associated with time
  constraints e.g., deadlines
• Refresh database items to reflect the external
  world.

5
Examples of RTDB

• Cellular Phone Network
• Threat Analysis System
• Programmed Stock Trading

6
Example Programmed Stock Trading

• Monitoring stocks financial instruments
• On discovering an opportunity, a trading
  transaction is performed.
• The transaction becomes useless if it misses the
  deadline.

7
System Model
Monitor
Dynamic Environment
Base Data
Views
8
Timing Requirements of Application Transactions

• Transaction Timeliness How fast a transaction is
  completed
• Can a transaction finish before its deadline?
• Data Timeliness How fresh the data is
• Stale data is less useful to a transaction

9
The 2 Requirements Conflict!

• Updates Recomputations applied promptly to
  satisfy data timeliness.
• They are very heavy loads.
• Transactions get little share of system
  resources.
• This hurts transaction timeliness.

10
Problems Studied in this Thesis

• Understand Updates, Recoms Transactions
• Balancing the 2 timing requirements by redefining
  Temporal Correctness
• Design scheduling algorithms so that temporal
  correctness is enforced

11
II Updates Recomputations
12
Properties of Updates

• Based on Adelberg, Garcia-Molina Kaos work
  Applying Update Streams in a Soft RTDB in ACM
  SIGMOD 1995.

13
Properties of Updates

• Updates arrive at a very high rate
• In the US market, 500 stock updates per second.
• An update need not be given full transactional
  support.
• Use a single update process to install updates

14
Properties of Recomputations

• Based on Adelberg, Garcia-Molina Kaos work,
  Database Support for Efficiently Maintaining
  Derived Data in EDBT 1996.

15
Fan-in and Fan-out
High Fan-out
High Fan-in
16
Batching

• A technique for alleviating the heavy recom load
• A recom request is not served immediately.
• It sleeps for a small amount of time.
• During the sleep period, any recoms for the same
  item are batched into a single recom.

17
Implications

• Updates Recoms are heavy burdens.
• Need special techniques e.g., batching
• These techniques need to be integrated into the
  scheduling protocols.

18
III Temporal Correctness
19
Temporal Correctness

• Absolute Consistency (AC)
• A data item timely reflects the state of an
  external object.
• Relative Consistency (RC)
• A set of data items with values reflecting the
  states of the external objects at the same time
  instant.

20
AC and RC An Example
A1
A2
time
A
B1
B2
time
B
Current Time

• A2 and B2 are AC. They are also RC.
• (A1 , B1) and (A1 , B2) are RC pairs.
• A2 and B1 are NOT consistent pairs.

21
Temporal Correctness from Transactions Point of
View

• Keeping an AC/RC database is difficult.
• We can extend temporal consistency from the
  transactions perspective.
• Can much reduce the update/recomputation
  required.

22
Defining AC/RC from the Perspective of a
Transaction
Database containing 1000 items
23
Instantaneous System

• An ideal system which applies updates/recoms as
  soon as they arrive, taking zero time to do it.
• A measure of correctness of real systems.

24
Real System - ACS

• A transaction is given the values of all objects
  it reads such that
• these values can be found in an instantaneous
  system at the commit time.
• No transactions can read old or inconsistent data.

25
Real System - RCS

• Can we relax data timeliness?
• Allow transactions to read slightly old data.
• The set of values read by a transaction can be
  found in an instantaneous system not older than
  transactions start time - ?.

26
ACS vs RCS An Example
A1
time
A
time
B
time
T
commit
start

• Readset of T in ACS (A2, B1)
• Readset of T in RCS (A1, B1), (A2, B1)

27
Advantages of RCS over ACS

• In ACS, an incoming update/recom can abort a
  transaction due to data conflicts
• In RCS, a transaction will not suffer from this
  problem.
• Freshness is sacrificed for fewer transaction
  aborts.

28
IV Transaction Scheduling
29
Schedulers for ACS and RCS

• URT
• Updates and Recoms first, Transactions later
• OD
• On Demand Triggering of Updates Recoms
• OD-H
• No transactions URT
• Otherwise, OD

30
Implementing the Schedulers

• Concurrency control
• ACS HP-2PL
• RCS Multi-version Concurrency Control
• OD manager for OD
• Version manager for RCS

31
Implementing the Schedulers

• Batching
• No transaction should read the view while its
  recomputation is sleeping.
• Lock the view before sleeping.

32
V Performance Study
33
Effect of Transaction Arrival Rate (ACS)
34
Comparing URT, OD OD-H
35
Effect of Transaction Arrival Rate (ACS RCS)
36
Effect of Transaction Arrival Rate(OD-H/ACS vs
OD-H/RCS)
37
Comparing ACS and RCS

• All 3 policies under RCS perform significantly
  better than ACS.
• 2 drawbacks of RCS
• Use of older versions.
• More complex to implement.

38
VI Conclusions
39
Conclusions

• We studied how temporal correctness can be
  incorporated into an RTDB.
• We defined temporal consistency from the
  perspective of transactions.

40
Conclusions

• The data timeliness of ACS is too strict.
• It has a high transaction miss rate.
• RCS is less strict.
• It has a smaller transaction miss rate.

41
Conclusions

• We studied how URT, OD and OD-H should be
  implemented for ACS.
• OD-H when applied to an RCS results in the
  smallest transaction miss rate.

42
Logical vs Temporal Correctness
Logical Correctness
Temporal Correctness
43
Future Work

• Version selection problem in an RCS
• If the readset of a transaction is known before
  its execution, the performance of RCS may be
  improved.
• The simulation model can be expanded from single
  processor and single disk, to multiple processors
  and multiple disks.

44
Q A Session