Introduction

A database transaction is a logical unit of database processing. It consists of a series of database access operations: insertion, deletions, selection, updating.

A transaction can be part of a user program or an interactive session.

Transactions can be read-only (selection) or updates (insertion, deletion, update).

Transactions have a clearly defined beginning and end. Example: an ATM transaction. There is an explicit login followed by a series of other transactions (deposit, withdrawl, transfer, balance) followed by an implicit logout. Each transaction must be finished before moving onto the next one.

At the physics level transactions involve a series of reads and writes to and from memory and disk. (We will worry about the physical layer in only a few special cases).

ACID properties. Transactions are:

• Atomic - all or nothing
• Consistency preservation - database state is consistent at beginning and end of transaction
  • database fulfills all constraints at both begining and end of transaction
• Isolation - transaction acts as though it is alone, and unaffected by concurrent transactions
  • solves problems of dirty reads, lost updates, etc.
• Durability - results in permanent changes to database

Database Concurrency is defined by the number of users who may access the database simultaneously:

• single-user: at most one user can access the database - mostly restricted to PC based databases
• multi-user: more than one user can access the database - 'real' databases

Concurrency Problems

The following are examples of concurrency problems.  A  and  B  are transactions.

• Lost Update

  Occurs when transactions are interleaved.

  A	B
  read(N) N = N - i
  	read(N) N = N + k
  write(N) (this update is lost)
  	write(N)

• Dirty Read

  A read is done on an updated item before the transaction that updated the item fails.

  A	B
  read(N) N = N - i write(N)
  	read(N)
  failure rollback(N)

  The N read by B is 'dirty data'

• Incorrect Summary

  When an aggregate function is applied to a number of records that are undergoing updating by another transaction

  A	B
  	sum = 0 sum = sum + N1 sum = sum + N2
  read(N5) N5 = N5 - i write(N5)	...
  	sum = sum + N5 sum = sum + N6
  read(N6) N6 = N6 - i write(N6)

  (assume reads of the appropriate values are done before each summation)

• Unrepeatable Read

  A value is changed between two reads, meaning that you do not get the same value twice

  A	B
  	read(N) display N
  read(N) N = n - i write(N)
  	read(N) N = N + j write(N)

  The update is done to the 'wrong' value of N.

Transaction States

• Begin Transaction
• Read
• Write
• End Transaction - reads and writes have ended, must choose one of next two
• Commit
• Rollback

System Log

• Log records are kept of every transaction, including:
  • Transaction ID
  • write( ID, old value, new value )
  • read( ID, value )
  • commit( ID )
  • rollback( ID )
  • checkpoint (more later) - when DBMS buffers are written to disk
• allows for UNDO or REDO of information
• reads are not necessarily always recorded (useful for auditing)
• logs are kept on disk, and in case of a system crash only transactions written to disk with commit records are considered committed. All others are rolled back.
• transaction log entries may interleave, if appropriate

Failures

System Failures

• database server is halted abnormally
• SQL commands are halted abnormally
• connections to clients are broken
• memory buffers are lost
• database files are not damaged with these failures

Media Failures

• database files are damaged
• may cause system failure

Log may recover from:

• system failure - determines active transactions when failure occurred, and undo uncommitted updates. Recreates lost committed updates.
• media failure - redo committed transactions lost since most recent backup
• aborted transaction - undo updates done by a rolled back transaction

Logs are scanned from tail to head to create a list of committed transactions, and then from head to tail to commit those transactions.

Checkpoints reduce the amount of log data that must be scanned after a failure. Tail to head scan goes only until last checkpoint (after system failure).

Checkpoint algorithm:

1. prevent new transactions from starting, and wait for active transactions to finish
2. copy memory buffers to disk
3. write a checkpoint record in the log
4. allow new transactions to start

Concurrency

For performance reasons, a database processing several transactions simultaneously. However, transactions appear to be processed serially.

Database Concurrency is defined by the number of users who may access the database simultaneously:

• single-user: at most one user can access the database - mostly restricted to PC based databases
• multi-user: more than one user can access the database - 'real' databases