1 Transaction Management Database recovery Concurrency control.

1

Transaction Management

Database recovery

Concurrency control

2

Transactions

A transaction is a sequence of operations that perform a single logical task.

The database must be in a consistent state before and after a transaction but may become inconsistent during execution.

Transaction management ensures that the database can be restored to a consistent state if something goes wrong during the transaction.

3

ACID Properties of a Transaction

Atomicity either all or none of the operations of a transaction are

performed. Consistency

db is in a consistent state before and after a executing a transaction (may be inconsistent during execution).

Isolation Transactions are isolated from one another.

Durability Once transaction commits, the changes it has made to db

persist, even if system fails.

4

Atomic Transactions

Either all or none of the operations of the transaction must be performed.

If all transactions were atomic, the database would always be in a consistent state. Unfortunately, they are not.

5

Example

Transfer $500 from savings account to checking account.

Begin transaction Read Sav_Amt Sav_Amt := Sav-Amt - 500 Write Sav_Amt Read Chk_Amt Chk_Amt := Chk_Amt + 500 Write Chk-AmtEnd transaction

6

Structure of a Transaction

Begin transaction

Read input message

Perform processing

against database

If successful

send output message(s)

and COMMIT

else

send error message

and ROLLBACK

End transaction

7

Commit

A commit Signals successful completion of a transaction to the

DBMS Frees any locks, created for example to avoid another

user from accessing the same data Makes all changes permanent and visible to other

users A commit does not mean that data has been written to

disk. DBMS keeps track of which changes have been

saved.

8

Rollback

A rollback Signals unsuccessful completion of a transaction

to the DBMS Undoes all changes made by the transaction

9

Database Recovery

After failure of some kind, database must be restored to some state known to be correct.

Recovery should be done Quickly With minimal transaction loss

To ensure possibility of recovery, one needs to perform Database backup Database logging Checkpointing

10

Database Backup

Periodically, the entire database should be copied to archival storage (e.g., tape, CD-ROM).

This copy should be stored in a safe place, preferably off-site.

11

Database Logging

Whenever a change is made to the database, write a record to a special log file or journal.

Record in log consists of Transaction name Data item name New value of data item Old value of item.

12

Checkpointing

When failure occurs, need to determine which transactions must be redone or undone.

Too time-consuming to search the entire log. Solution: Use checkpointing

Synchronize log and the database by performing all pending writes

Once checkpointing has been done, write a checkpoint to the log.

Do recovery from last checkpoint.

13

Types of Failure and the Recovery Mechanism I

Transaction - local failure As only one transaction is affected, perform a ROLLBACK.

System-wide, no damage to DB All transactions in progress are affected, and hence undo

changes made by transactions in progress. Redo every committed transaction for which it is not known

whether all changes have been physically written to database. If possible, restart transactions that were rolled back.

14

Types of Failure and the Recovery Mechanism II System-wide failure with damage to database

Restore database from latest backup, and redo all committed transactions from the log file.

Clearly, this is a very slow process.

15

Concurrency Control

Concurrently executed transactions may interleave with each other in such a way that they produce an incorrect overall result, even though each transaction is correct.

Consider the following example: User1 wants to transfer $50 from A to B(T1). User

2 wants to transfer 10% from A to B (T2).

16

Serial Execution I

T1 T2

read(A)

A := A - 50

write(A)

read(B)

B := B + 50

write(B)

read(A)

temp = A *.1

A := A - temp

write(A)

read(B)

B := B + temp

write(B)

17

Serial Execution II

T1 T2

read(A)

temp = A *.1

A := A - temp

write(A)

read(B)

B := B + temp

write(B)

read(A)

A := A - 50

write(A)

read(B)

B := B + 50

write(B)

18

Good Interleaving

T1 T2

read(A)

A := A - 50

write(A)

read(A)

temp = A *.1

A := A - temp

write(A)

read(B)

B := B + 50

write(B)

read(B)

B := B + temp

write(B)

19

Bad Interleaving

T1 T2

read(A)

A := A - 50

read(A)

temp = A *.1

A := A - temp

write(A)

read(B)

write(A)

read(B)

B := B + 50

write(B)

B := B + temp

write(B)

20

The Lost Update Problem

Transaction A time Transaction B

- -

- -

RETRIEVE t -

- -

- RETRIEVE t

- -

UPDATE t -

- -

- UPDATE t

- -

t1

t2

t3

t4

21

Uncommitted Dependency Problem

Transaction A time Transaction B

- -

- -

- UPDATE t

- -

RETREIVE t -

- -

- ROLLBACK

-

t1

t2

t3

22

Uncommitted Dependency Problem

Transaction A time Transaction B

- -

- -

- UPDATE t

- -

UPDATE t -

- -

- ROLLBACK

-

t1

t2

t3

23

The Inconsistent Analysis Problem

Example: Acc 1 = 40; Acc 2 = 50; Acc 3 = 30 Transaction A : Sum all account balances Transaction B : Transfer 10 from 3 to 1

24

The Inconsistent Analysis Problem

Transaction A time Transaction B

RETRIEVE Acc 1 : - Sum = 40 -

- - RETRIEVE Acc 2 : -

Sum = 90 -- -

- RETRIEVE Acc 3:- -- UPDATE Acc3:- 30 -> 20 - -- RETRIEVE Acc 1:- -- UPDATE Acc 1:- 40 -> 50- -- COMMIT-

RETRIEVE Acc 3 : - Sum = 110 (not 120) -

t3

t1

t2

t4

t5

t6

t7

t8

25

Serializable Executions

An interleaved execution of some transactions is correct if it produces the same result as some serial execution of the transactions.

Such an execution is called serializable. A concurrency control scheme must prevent non-

serializable execution from occurring. One possibility is locking.

26

Locks

While one transaction accesses a data item, no other transaction should modify it.

Locking ensures that a data item can be updated only if the transaction holds a lock on the data item.

Two types of lock: Shared locks Exclusive locks

27

Shared locks

If a transaction holds a shared lock (S-lock) on an object, other transactions can also request S-locks.

However, a transaction cannot acquire an exclusive lock on the object.

If a transaction has the only shared lock on an object, it can be promoted to an exclusive lock.

28

Exclusive Locks

If a transaction holds an exclusive lock (X-lock) on an object, no other transaction can acquire a lock on the object, or access it.

To update a record R through some transaction T T must obtain an X-lock on R The X-lock must be retained until the end of T,

either through COMMIT or ROLLBACK.

29

Granting of Locks

Care must be taken to ensure that a transaction will not be starved.

S X

S T F

X F F

Lock Compatibility Matrix

A lock request should never get blocked by a lock request that is made later.

30

Good Interleaving with X-locks I

T1 T2

lock-X(A)

read(A)

A := A - 50

write(A)

unlock(A)

lock-X(A)

read(A)

temp = A *.1

A := A - temp

write(A)

unlock(A)

31

Good Interleaving with X-locks II

T1 T2

lock-X(B)

read(B)

B := B + 50

write(B)

unlock(B)

lock-X(B)

read(B)

B := B = temp

write(B)

unlock(X)

32

X-locks Block Bad Interleaving

T1 T2

lock-X(A)

read(A)

A := A - 50

read(A)

The read(A) of T2 cannot be executed because T1 has an X-lock on A.

33

Deadlocks

X-locks may lead to deadlocks. This arises when two transactions are mutually

excluded from accessing the next record required to computer their transaction.

34

Our Example Slightly Changed

T1 T2read(B)B := B + 50write(B)

read(A)temp = A *.1A := A - tempwrite(A)

read(A)A := A - 50write(A)

read(B)B := B = tempwrite(B)

35

Example of Deadlock

T1 T2

lock-X(B)

update B

lock-X(A)

update A

request lock(A)

wait for T2 to

release lock on A

request lock(B)

waiting for T1 to

release lock on B

36

Deadlocks

Deadlocks can be overcome by: Prevention Detection and Recovery

37

Technique 1 - Preventing Deadlocks

A transaction must lock all the data items it needs before execution begins.

Data-item utilization may be slow.

38

Technique 2 -Preventing Deadlocks

Ordering If a transaction requires access to two records,

make sure that they are always accessed in the same order.

In our case, always access A before B. May slow down operations considerably.

39

Detection and Recovery

Periodically, let DBMS check to determine if line waiting for resource exceeds some limit.

Use graph manipulation to detect deadlocks: Draw arrow from transaction to record being

sought, and from record to transaction using it. If graph has cycles, then we have deadlock.

If deadlock detected, cancel one transaction and advance other.

40

Example of Deadlock Detection

T1 T2

Record A Record B