Concurrency Control

● Lock-Based Protocols

● Timestamp-Based Protocols

● Validation-Based Protocols

● Multiple Granularity

● Multiversion Schemes

● Insert and Delete Operations

● Concurrency in Index Structures

Lock-Based Protocols

● A lock is a mechanism to control concurrent access to a data item

● Data items can be locked in two modes :

1. exclusive (X) mode. Data item can be both read as well as

written. X-lock is requested using lock-X instruction.

2. shared (S) mode. Data item can only be read. S-lock is

requested using lock-S instruction.

● Lock requests are made to concurrency-control manager. Transaction can proceed only after request is
granted.

● Lock-compatibility matrix

● A transaction may be granted a lock on an item if the requested lock is compatible with locks already held on
the item by other transactions.

● Any number of transactions can hold shared locks on an item,

● but if any transaction holds an exclusive on the item no other transaction may hold any lock on the
item.

● If a lock cannot be granted, the requesting transaction is made to wait till all incompatible locks held by
other transactions have been released. The lock is then granted.
● Example of a transaction performing locking:

T2: lock-S(A);

read (A);

unlock(A);

lock-S(B);

read (B);

unlock(B);

display(A+B)

● Locking as above is not sufficient to guarantee serializability — if A and B get updated in-between the read of
A and B, the displayed sum would be wrong.

● A locking protocol is a set of rules followed by all transactions while requesting and releasing locks. Locking
protocols restrict the set of possible schedules.
Pitfalls of Lock-Based Protocols

● Consider the partial schedule

● Neither T3 nor T4 can make progress — executing lock-S(B) causes T4 to wait for T3 to release its lock on B,
while executing lock-X(A) causes T3 to wait for T4 to release its lock on A.

● Such a situation is called a deadlock.

● To handle a deadlock one of T3 or T4 must be rolled back

and its locks released.

● The potential for deadlock exists in most locking protocols. Deadlocks are a necessary evil.

● Starvation is also possible if concurrency control manager is badly designed. For example:

● A transaction may be waiting for an X-lock on an item, while a sequence of other transactions
request and are granted an S-lock on the same item.

● The same transaction is repeatedly rolled back due to deadlocks.

● Concurrency control manager can be designed to prevent starvation.

The Two-Phase Locking Protocol

● This is a protocol which ensures conflict-serializable schedules.

● Phase 1: Growing Phase

● transaction may obtain locks

● transaction may not release locks

● Phase 2: Shrinking Phase

● transaction may release locks

● transaction may not obtain locks

● The protocol assures serializability. It can be proved that the transactions can be serialized in the order of
their lock points (i.e. the point where a transaction acquired its final lock).

● Two-phase locking does not ensure freedom from deadlocks

 ● Cascading roll-back is possible under two-phase locking. To avoid this, follow a modified protocol called strict
two-phase locking. Here a transaction must hold all its exclusive locks till it commits/aborts.

● Rigorous two-phase locking is even stricter: here all locks are held till commit/abort. In this protocol
transactions can be serialized in the order in which they commit.

Lock Conversions

● Two-phase locking with lock conversions:

– First Phase:

● can acquire a lock-S on item

● can acquire a lock-X on item

● can convert a lock-S to a lock-X (upgrade)

– Second Phase:

● can release a lock-S

● can release a lock-X

● can convert a lock-X to a lock-S (downgrade)

● This protocol assures serializability. But still relies on the programmer to insert the various locking
instructions.

Automatic Acquisition of Locks

● A transaction Ti issues the standard read/write instruction, without explicit locking calls.

● The operation read(D) is processed as:

if Ti has a lock on D

then

read(D)

else begin

if necessary wait until no other

transaction has a lock-X on D

grant Ti a lock-S on D;

read(D)

end

● write(D) is processed as:

if Ti has a lock-X on D

then

write(D)

else begin

if necessary wait until no other trans. has any lock on D,

if Ti has a lock-S on D
 then

upgrade lock on D to lock-X

else

grant Ti a lock-X on D

write(D)

end;

● All locks are released after commit or abort

Implementation of Locking

● A lock manager can be implemented as a separate process to which transactions send lock and unlock
requests

● The lock manager replies to a lock request by sending a lock grant messages (or a message asking the
transaction to roll back, in case of a deadlock)

● The requesting transaction waits until its request is answered

● The lock manager maintains a data-structure called a lock table to record granted locks and pending
requests

● The lock table is usually implemented as an in-memory hash table indexed on the name of the data item
being locked

Deadlock

Deadlock Handling

● System is deadlocked if there is a set of transactions such that every transaction in the set is waiting for
another transaction in the set.

● Deadlock prevention protocols ensure that the system will never enter into a deadlock state. Some
prevention strategies :

● Require that each transaction locks all its data items before it begins execution (predeclaration).

● Impose partial ordering of all data items and require that a transaction can lock data items only in
the order specified by the partial order (graph-based protocol).

● Deadlock prevention by ordering usually ensured by careful programming of transactions

Deadlock Detection

● Deadlock detection algorithms used to detect deadlocks

Deadlock Recovery

● When deadlock is detected :

● Some transaction will have to rolled back (made a victim) to break deadlock. Select that transaction
as victim that will incur minimum cost.

● Rollback -- determine how far to roll back transaction

4 Total rollback: Abort the transaction and then restart it.

4 More effective to roll back transaction only as far as necessary to break deadlock.

● Starvation happens if same transaction is always chosen as victim. Include the number of rollbacks in
the cost factor to avoid starvation

Timestamp-Based Protocols

● Each transaction is issued a timestamp when it enters the system. If an old transaction Ti has time-stamp
TS(Ti), a new transaction Tj is assigned time-stamp TS(Tj) such that TS(Ti) <TS(Tj).

● The protocol manages concurrent execution such that the time-stamps determine the serializability order.

● In order to assure such behavior, the protocol maintains for each data Q two timestamp values:

● W-timestamp(Q) is the largest time-stamp of any transaction that executed write(Q) successfully.

● R-timestamp(Q) is the largest time-stamp of any transaction that executed read(Q) successfully.

● The timestamp ordering protocol ensures that any conflicting read and write operations are executed in
timestamp order.

● Suppose a transaction Ti issues a read(Q)

● If TS(Ti) ≤ W-timestamp(Q), then Ti needs to read a value of Q that was already overwritten.

● Hence, the read operation is rejected, and Ti is rolled back.

● If TS(Ti)≥ W-timestamp(Q), then the read operation is executed, and R-timestamp(Q) is set to max(R-
timestamp(Q), TS(Ti)).

● Suppose that transaction Ti issues write(Q).

● If TS(Ti) < R-timestamp(Q), then the value of Q that Ti is producing was needed previously, and the
system assumed that that value would never be produced.
 ● Hence, the write operation is rejected, and Ti is rolled back.

● If TS(Ti) < W-timestamp(Q), then Ti is attempting to write an obsolete value of Q.

● Hence, this write operation is rejected, and Ti is rolled back.

● Otherwise, the write operation is executed, and W-timestamp(Q) is set to TS(Ti).

Example Use of the Protocol

A partial schedule for several data items for transactions with

timestamps 1, 2, 3, 4, 5

Correctness of Timestamp-Ordering Protocol

● The timestamp-ordering protocol guarantees serializability since all the arcs in the precedence graph are of
the form:

Thus, there will be no cycles in the precedence graph

● Timestamp protocol ensures freedom from deadlock as no transaction ever waits.

● But the schedule may not be cascade-free, and may not even be recoverable.

Recoverability and Cascade Freedom

● Problem with timestamp-ordering protocol:

● Suppose Ti aborts, but Tj has read a data item written by Ti

 ● Then Tj must abort; if Tj had been allowed to commit earlier, the schedule is not recoverable.

● Further, any transaction that has read a data item written by Tj must abort

● This can lead to cascading rollback --- that is, a chain of rollbacks

● Solution 1:

● A transaction is structured such that its writes are all performed at the end of its processing

● All writes of a transaction form an atomic action; no transaction may execute while a transaction is
being written

● A transaction that aborts is restarted with a new timestamp

● Solution 2: Limited form of locking: wait for data to be committed before reading it

● Solution 3: Use commit dependencies to ensure recoverability

Validation-Based Protocols

● Execution of transaction Ti is done in three phases.

1. Read and execution phase: Transaction Ti writes only to

temporary local variables

2. Validation phase: Transaction Ti performs a ``validation test''

to determine if local variables can be written without violating

serializability.

3. Write phase: If Ti is validated, the updates are applied to the

database; otherwise, Ti is rolled back.

● The three phases of concurrently executing transactions can be interleaved, but each transaction must go
through the three phases in that order.

● Assume for simplicity that the validation and write phase occur together, atomically and serially

4 I.e., only one transaction executes validation/write at a time.

● Also called as optimistic concurrency control since transaction executes fully in the hope that all will go well
during validation

● Validation is based on timestamps, but with two timestamps:

● start time

● validation time

● Details in book

Phantom Problem

● Insertions, deletions and updates can lead to the phantom phenomenon.

● A transaction that scans a relation

4 (e.g., find sum of balances of all accounts in Perryridge)

and a transaction that inserts a tuple in the relation

4 (e.g., insert a new account at Perryridge)

(conceptually) conflict in spite of not accessing any tuple in common.

● If only tuple locks are used, non-serializable schedules can result

4 E.g. the scan transaction does not see the new account, but reads some other tuple written
by the update transaction

● Index locking protocols used to prevent phantom phenomenon (see book for details)

Weak Levels of Consistency in SQL

● SQL allows non-serializable executions

● Serializable: is the default

● Repeatable read: allows only committed records to be read, and repeating a read should return the
same value (so read locks should be retained)

4 However, the phantom phenomenon need not be prevented

– T1 may see some records inserted by T2, but may not see others inserted by T2

● Read committed: same as degree two consistency, but most systems implement it as cursor-
stability

● Read uncommitted: allows even uncommitted data to be read

● In many database systems, read committed is the default consistency level

● has to be explicitly changed to serializable when required

4 set isolation level serializable

Concurrency in Index Structures

● Indices are unlike other database items in that their only job is to help in accessing data.

● Index-structures are typically accessed very often, much more than other database items.

● Treating index-structures like other database items, e.g. by 2-phase locking of index nodes can lead
to low concurrency.

● There are several index concurrency protocols where locks on internal nodes are released early, and not in a
two-phase fashion.

● It is acceptable to have nonserializable concurrent access to an index as long as the accuracy of the
index is maintained.

4 In particular, the exact values read in an internal node of a

B+-tree are irrelevant so long as we land up in the correct leaf node.
Recovery System

n Failure Classification

n Storage Structure

n Recovery and Atomicity

n Log-Based Recovery

n Shadow Paging

n Recovery With Concurrent Transactions

n Buffer Management

n Failure with Loss of Nonvolatile Storage

n Advanced Recovery Techniques

n ARIES Recovery Algorithm

n Remote Backup Systems

n Transaction failure :

n Logical errors: transaction cannot complete due to some internal error condition

n System errors: the database system must terminate an active transaction due to an error condition
(e.g., deadlock)

n System crash: a power failure or other hardware or software failure causes the system to crash.

n Fail-stop assumption: non-volatile storage contents are assumed to not be corrupted by system
crash

n Database systems have numerous integrity checks to prevent corruption of disk data

n Disk failure: a head crash or similar disk failure destroys all or part of disk storage

n Destruction is assumed to be detectable: disk drives use checksums to detect failures

Recovery Algorithms

n Recovery algorithms are techniques to ensure database consistency and transaction atomicity and durability
despite failures

 Focus of this chapter

n Recovery algorithms have two parts

 Actions taken during normal transaction processing to ensure enough information exists to recover
from failures

 Actions taken after a failure to recover the database contents to a state that ensures atomicity,
consistency and durability

Storage Structure

n Volatile storage:

 does not survive system crashes

 examples: main memory, cache memory

n Nonvolatile storage:
  survives system crashes

 examples: disk, tape, flash memory,

non-volatile (battery backed up) RAM

n Stable storage:

 a mythical form of storage that survives all failures

 approximated by maintaining multiple copies on distinct nonvolatile media

Stable-Storage Implementation

n Maintain multiple copies of each block on separate disks

 copies can be at remote sites to protect against disasters such as fire or flooding.

n Failure during data transfer can still result in inconsistent copies: Block transfer can result in

 Successful completion

 Partial failure: destination block has incorrect information

 Total failure: destination block was never updated

n Protecting storage media from failure during data transfer (one solution):

 Execute output operation as follows (assuming two copies of each block):

1. Write the information onto the first physical block.

2. When the first write successfully completes, write the same information onto the second
physical block.

3. The output is completed only after the second write successfully completes.

n Protecting storage media from failure during data transfer (cont.):

n Copies of a block may differ due to failure during output operation. To recover from failure:

 First find inconsistent blocks:

1. Expensive solution: Compare the two copies of every disk block.

2. Better solution:

n Record in-progress disk writes on non-volatile storage (Non-volatile RAM or special

area of disk).

n Use this information during recovery to find blocks that may be inconsistent, and
only compare copies of these.

n Used in hardware RAID systems

 If either copy of an inconsistent block is detected to have an error (bad checksum), overwrite it by
the other copy. If both have no error, but are different, overwrite the second block by the first
block.

Data Access

n Physical blocks are those blocks residing on the disk.

n Buffer blocks are the blocks residing temporarily in main memory.

n Block movements between disk and main memory are initiated through the following two operations:
  input(B) transfers the physical block B to main memory.

 output(B) transfers the buffer block B to the disk, and replaces the appropriate physical block there.

n Each transaction Ti has its private work-area in which local copies of all data items accessed and updated by
it are kept.

 Ti's local copy of a data item X is called xi.

n We assume, for simplicity, that each data item fits in, and is stored inside, a single block.

n Transaction transfers data items between system buffer blocks and its private work-area using the following
operations :

 read(X) assigns the value of data item X to the local variable xi.

 write(X) assigns the value of local variable xi to data item {X} in the buffer block.

 both these commands may necessitate the issue of an input(BX) instruction before the assignment, if
the block BX in which X resides is not already in memory.

n Transactions

 Perform read(X) while accessing X for the first time;

 All subsequent accesses are to the local copy.

 After last access, transaction executes write(X).

n output(BX) need not immediately follow write(X). System can perform the output operation when it deems
fit.

Recovery and Atomicity

n Modifying the database without ensuring that the transaction will commit may leave the database in an
inconsistent state.

n Consider transaction Ti that transfers $50 from account A to account B; goal is either to perform all database
modifications made by Ti or none at all.
 n Several output operations may be required for Ti (to output A and B). A failure may occur after one of these
modifications have been made but before all of them are made.

n To ensure atomicity despite failures, we first output information describing the modifications to stable
storage without modifying the database itself.

n We study two approaches:

n log-based recovery, and

n shadow-paging

n We assume (initially) that transactions run serially, that is, one after the other.

Log-Based Recovery

n A log is kept on stable storage.

 The log is a sequence of log records, and maintains a record of update activities on the database.

n When transaction Ti starts, it registers itself by writing a

<Ti start>log record

n Before Ti executes write(X), a log record <Ti, X, V1, V2> is written, where V1 is the value of X before the
write, and V2 is the value to be written to X.

 Log record notes that Ti has performed a write on data item Xj Xj had value V1 before the write, and
will have value V2 after the write.

n When Ti finishes it last statement, the log record <Ti commit> is written.

n We assume for now that log records are written directly to stable storage (that is, they are not buffered)

n Two approaches using logs

 Deferred database modification

 Immediate database modification

Deferred Database Modification

n The deferred database modification scheme records all modifications to the log, but defers all the writes to
after partial commit.

n Assume that transactions execute serially

n Transaction starts by writing <Ti start> record to log.

n A write(X) operation results in a log record <Ti, X, V> being written, where V is the new value for X

 Note: old value is not needed for this scheme

n The write is not performed on X at this time, but is deferred.

n When Ti partially commits, <Ti commit> is written to the log

n Finally, the log records are read and used to actually execute the previously deferred writes.

n During recovery after a crash, a transaction needs to be redone if and only if both <Ti start> and<Ti commit>
are there in the log.

n Redoing a transaction Ti ( redoTi) sets the value of all data items updated by the transaction to the new
values.

n Crashes can occur while

  the transaction is executing the original updates, or

 while recovery action is being taken

n example transactions T0 and T1 (T0 executes before T1):

T0: read (A) T1 : read (C)

A: - A - 50 C:- C- 100

Write (A) write (C)

read (B)

B:- B + 50

write (B)

n Below we show the log as it appears at three instances of time.

n
n If log on stable storage at time of crash is as in case:

(a) No redo actions need to be taken

(b) redo(T0) must be performed since <T0 commit> is present

(c) redo(T0) must be performed followed by redo(T1) since

<T0 commit> and <Ti commit> are present

n The immediate database modification scheme allows database updates of an uncommitted transaction to
be made as the writes are issued

 since undoing may be needed, update logs must have both old value and new value

n Update log record must be written before database item is written

 We assume that the log record is output directly to stable storage

 Can be extended to postpone log record output, so long as prior to execution of an output(B)
operation for a data block B, all log records corresponding to items B must be flushed to stable
storage

n Output of updated blocks can take place at any time before or after transaction commit

n Order in which blocks are output can be different from the order in which they are written.
Immediate Database Modification Example

Log Write Output

<T0 start>

<T0, A, 1000, 950>

To, B, 2000, 2050

A = 950

B = 2050

<T0 commit>

<T1 start>

<T1, C, 700, 600>

C = 600

BB, BC

<T1 commit>

BA

n Note: BX denotes block containing X.

n Recovery procedure has two operations instead of one:

n undo(Ti) restores the value of all data items updated by Ti to their old values, going backwards from
the last log record for Ti

n redo(Ti) sets the value of all data items updated by Ti to the new values, going forward from the first
log record for Ti

n Both operations must be idempotent

n That is, even if the operation is executed multiple times the effect is the same as if it is executed
once

n Needed since operations may get re-executed during recovery

n When recovering after failure:

n Transaction Ti needs to be undone if the log contains the record

<Ti start>, but does not contain the record <Ti commit>.

n Transaction Ti needs to be redone if the log contains both the record <Ti start> and the record <Ti
commit>.

n Undo operations are performed first, then redo operations.

n Below we show the log as it appears at three instances of time.

 n Recovery actions in each case above are:

n (a) undo (T0): B is restored to 2000 and A to 1000.

n (b) undo (T1) and redo (T0): C is restored to 700, and then A and B are

n set to 950 and 2050 respectively.

n (c) redo (T0) and redo (T1): A and B are set to 950 and 2050

n respectively. Then C is set to 600

Checkpoints

n Problems in recovery procedure as discussed earlier :

1. searching the entire log is time-consuming

2. we might unnecessarily redo transactions which have already

3. output their updates to the database.

n Streamline recovery procedure by periodically performing checkpointing

1. Output all log records currently residing in main memory onto stable storage.

2. Output all modified buffer blocks to the disk.

3. Write a log record < checkpoint> onto stable storage.

n During recovery we need to consider only the most recent transaction Ti that started before the checkpoint,
and transactions that started after Ti.

1. Scan backwards from end of log to find the most recent <checkpoint> record

2. Continue scanning backwards till a record <Ti start> is found.

3. Need only consider the part of log following above start record. Earlier part of log can be ignored
during recovery, and can be erased whenever desired.

4. For all transactions (starting from Ti or later) with no <Ti commit>, execute undo(Ti). (Done only in
case of immediate modification.)

5. Scanning forward in the log, for all transactions starting from Ti or later with a <Ti commit>,
execute redo(Ti).
Example of Checkpoints

n T1 can be ignored (updates already output to disk due to checkpoint)

n T2 and T3 redone.

n T4 undone

Shadow Paging

n Shadow paging is an alternative to log-based recovery; this scheme is useful if transactions execute serially

n Idea: maintain two page tables during the lifetime of a transaction –the current page table, and the shadow
page table

n Store the shadow page table in nonvolatile storage, such that state of the database prior to transaction
execution may be recovered.

 Shadow page table is never modified during execution

n To start with, both the page tables are identical. Only current page table is used for data item accesses
during execution of the transaction.

n Whenever any page is about to be written for the first time

 A copy of this page is made onto an unused page.

 The current page table is then made to point to the copy

 The update is performed on the copy

 n To commit a transaction :

1. Flush all modified pages in main memory to disk

2. Output current page table to disk

3. Make the current page table the new shadow page table, as follows:

 keep a pointer to the shadow page table at a fixed (known) location on disk.

 to make the current page table the new shadow page table, simply update the pointer to point to
current page table on disk

n Once pointer to shadow page table has been written, transaction is committed.

n No recovery is needed after a crash — new transactions can start right away, using the shadow page table.

n Pages not pointed to from current/shadow page table should be freed (garbage collected).

Recovery With Concurrent Transactions (Cont.)

n Checkpoints are performed as before, except that the checkpoint log record is now of the form
< checkpoint L>
where L is the list of transactions active at the time of the checkpoint

 We assume no updates are in progress while the checkpoint is carried out (will relax this later)

n When the system recovers from a crash, it first does the following:

 Initialize undo-list and redo-list to empty

 Scan the log backwards from the end, stopping when the first <checkpoint L> record is found.
For each record found during the backward scan:

H if the record is <Ti commit>, add Ti to redo-list

H if the record is <Ti start>, then if Ti is not in redo-list, add Ti to undo-list

  For every Ti in L, if Ti is not in redo-list, add Ti to undo-list

n At this point undo-list consists of incomplete transactions which must be undone, and redo-list consists of
finished transactions that must be redone.

n Recovery now continues as follows:

 Scan log backwards from most recent record, stopping when

<Ti start> records have been encountered for every Ti in undo-list.

H During the scan, perform undo for each log record that belongs to a transaction in undo-list.

 Locate the most recent <checkpoint L> record.

 Scan log forwards from the <checkpoint L> record till the end of the log.

H During the scan, perform redo for each log record that belongs to a transaction on redo-list

Example of Recovery

n Go over the steps of the recovery algorithm on the following log:

<T0 start>

<T0, A, 0, 10>

<T0 commit>

<T1 start>

<T1, B, 0, 10>

<T2 start> /* Scan in Step 4 stops here */

<T2, C, 0, 10>

<T2, C, 10, 20>

<checkpoint {T1, T2}>

<T3 start>

<T3, A, 10, 20>

<T3, D, 0, 10>

<T3 commit>

Log Record Buffering

n Log record buffering: log records are buffered in main memory, instead of of being output directly to stable
storage.

 Log records are output to stable storage when a block of log records in the buffer is full, or a log
force operation is executed.

n Log force is performed to commit a transaction by forcing all its log records (including the commit record) to
stable storage.

n Several log records can thus be output using a single output operation, reducing the I/O cost.

n The rules below must be followed if log records are buffered:

 Log records are output to stable storage in the order in which they are created.
  Transaction Ti enters the commit state only when the log record
<Ti commit> has been output to stable storage.

 Before a block of data in main memory is output to the database, all log records pertaining to data in
that block must have been output to stable storage.

n This rule is called the write-ahead logging or WAL rule

n Strictly speaking WAL only requires undo information to be output

Database Buffering

n Database maintains an in-memory buffer of data blocks

 When a new block is needed, if buffer is full an existing block needs to be removed from buffer

 If the block chosen for removal has been updated, it must be output to disk

n As a result of the write-ahead logging rule, if a block with uncommitted updates is output to disk, log records
with undo information for the updates are output to the log on stable storage first.

n No updates should be in progress on a block when it is output to disk. Can be ensured as follows.

 Before writing a data item, transaction acquires exclusive lock on block containing the data item

 Lock can be released once the write is completed.

➢ Such locks held for short duration are called latches.

 Before a block is output to disk, the system acquires an exclusive latch on the block

➢ Ensures no update can be in progress on the block

Buffer Management (Cont.)

n Database buffer can be implemented either

 in an area of real main-memory reserved for the database, or

 in virtual memory

n Implementing buffer in reserved main-memory has drawbacks:

 Memory is partitioned before-hand between database buffer and applications, limiting flexibility.

 Needs may change, and although operating system knows best how memory should be divided up at
any time, it cannot change the partitioning of memory.

n Database buffers are generally implemented in virtual memory in spite of some drawbacks:

 When operating system needs to evict a page that has been modified, to make space for another
page, the page is written to swap space on disk.

 When database decides to write buffer page to disk, buffer page may be in swap space, and may
have to be read from swap space on disk and output to the database on disk, resulting in extra I/O!

n Known as dual paging problem.

 Ideally when swapping out a database buffer page, operating system should pass control to
database, which in turn outputs page to database instead of to swap space (making sure to output
log records first)

n Dual paging can thus be avoided, but common operating systems do not support such
functionality.
Failure with Loss of Nonvolatile Storage

n So far we assumed no loss of non-volatile storage

n Technique similar to checkpointing used to deal with loss of non-volatile storage

 Periodically dump the entire content of the database to stable storage

 No transaction may be active during the dump procedure; a procedure similar to checkpointing must
take place

➢ Output all log records currently residing in main memory onto stable storage.

➢ Output all buffer blocks onto the disk.

➢ Copy the contents of the database to stable storage.

➢ Output a record <dump> to log on stable storage.

 To recover from disk failure

➢ restore database from most recent dump.

➢ Consult the log and redo all transactions that committed after the dump

n Can be extended to allow transactions to be active during dump;

known as fuzzy dump or online dump

 Will study fuzzy checkpointing later

Advanced Recovery Techniques

n Support high-concurrency locking techniques, such as those used for B+-tree concurrency control

n Operations like B+-tree insertions and deletions release locks early.

 They cannot be undone by restoring old values (physical undo), since once a lock is released, other
transactions may have updated the B+-tree.

 Instead, insertions (resp. deletions) are undone by executing a deletion (resp. insertion) operation
(known as logical undo).

n For such operations, undo log records should contain the undo operation to be executed

 called logical undo logging, in contrast to physical undo logging.

n Redo information is logged physically (that is, new value for each write) even for such operations

 Logical redo is very complicated since database state on disk may not be “operation consistent”

n Operation logging is done as follows:

 When operation starts, log <Ti, Oj, operation-begin>. Here Oj is a unique identifier of the operation
instance.

 While operation is executing, normal log records with physical redo and physical undo information
are logged.

 When operation completes, <Ti, Oj, operation-end, U> is logged, where U contains information
needed to perform a logical undo information.

n If crash/rollback occurs before operation completes:

 the operation-end log record is not found, and

 the physical undo information is used to undo operation.

 n If crash/rollback occurs after the operation completes:

 the operation-end log record is found, and in this case

 logical undo is performed using U; the physical undo information for the operation is ignored.

n Redo of operation (after crash) still uses physical redo information.

Rollback of transaction Ti is done as follows:

n Scan the log backwards

1. If a log record <Ti, X, V1, V2> is found, perform the undo and log a special redo-only log record <Ti, X,
V1>.

2. If a <Ti, Oj, operation-end, U> record is found

➢ Rollback the operation logically using the undo information U.

– Updates performed during roll back are logged just like during normal operation
execution.

– At the end of the operation rollback, instead of logging an operation-end record,

generate a record

<Ti, Oj, operation-abort>.

➢ Skip all preceding log records for Ti until the record <Ti, Oj operation-begin> is found

n Scan the log backwards (cont.):

3. If a redo-only record is found ignore it

4. If a <Ti, Oj, operation-abort> record is found:

H skip all preceding log records for Ti until the record

<Ti, Oj, operation-begin> is found.

5. Stop the scan when the record <Ti, start> is found

6. Add a <Ti, abort> record to the log

Some points to note:

n Cases 3 and 4 above can occur only if the database crashes while a transaction is being rolled back.

n Skipping of log records as in case 4 is important to prevent multiple rollback of the same operation.

The following actions are taken when recovering from system crash

1. Scan log forward from last < checkpoint L> record

1. Repeat history by physically redoing all updates of all transactions,

2. Create an undo-list during the scan as follows

➢ undo-list is set to L initially

➢ Whenever <Ti start> is found Ti is added to undo-list

➢ Whenever <Ti commit> or <Ti abort> is found, Ti is deleted from undo-list

This brings database to state as of crash, with committed as well as uncommitted transactions having been
redone.
 Now undo-list contains transactions that are incomplete, that is, have neither committed nor been fully
rolled back.

Recovery from system crash (cont.)

2. Scan log backwards, performing undo on log records of transactions found in undo-list.

 Transactions are rolled back as described earlier.

 When <Ti start> is found for a transaction Ti in undo-list, write a <Ti abort> log record.

 Stop scan when <Ti start> records have been found for all Ti in undo-list

n This undoes the effects of incomplete transactions (those with neither commit nor abort log records).
Recovery is now complete.

n Checkpointing is done as follows:

1. Output all log records in memory to stable storage

2. Output to disk all modified buffer blocks

3. Output to log on stable storage a < checkpoint L> record.

Transactions are not allowed to perform any actions while checkpointing is in progress.

n Fuzzy checkpointing allows transactions to progress while the most time consuming parts of checkpointing
are in progress

 Performed as described on next slide

n Fuzzy checkpointing is done as follows:

 Temporarily stop all updates by transactions

 Write a <checkpoint L> log record and force log to stable storage

 Note list M of modified buffer blocks

 Now permit transactions to proceed with their actions

 Output to disk all modified buffer blocks in list M

n blocks should not be updated while being output

n Follow WAL: all log records pertaining to a block must be output before the block is output

 Store a pointer to the checkpoint record in a fixed position last_checkpoint on disk

n When recovering using a fuzzy checkpoint, start scan from the checkpoint record pointed to by
last_checkpoint

 Log records before last_checkpoint have their updates reflected in database on disk, and need not
be redone.

 Incomplete checkpoints, where system had crashed while performing

checkpoint, are handled safely

ARIES Recovery Algorithm

n ARIES is a state of the art recovery method

 Incorporates numerous optimizations to reduce overheads during normal processing and to speed
up recovery
  The “advanced recovery algorithm” we studied earlier is modeled after ARIES, but greatly simplified
by removing optimizations

n Unlike the advanced recovery algorithm, ARIES

 Uses log sequence number (LSN) to identify log records

➢ Stores LSNs in pages to identify what updates have already been applied to a database page

 Physiological redo

 Dirty page table to avoid unnecessary redos during recovery

 Fuzzy checkpointing that only records information about dirty pages, and does not require dirty
pages to be written out at checkpoint time

➢ More coming up on each of the above …

n Physiological redo

 Affected page is physically identified, action within page can be logical

➢ Used to reduce logging overheads

n e.g. when a record is deleted and all other records have to be moved to fill hole

n Physiological redo can log just the record deletion

n Physical redo would require logging of old and new values for much of the
page

➢ Requires page to be output to disk atomically

n Easy to achieve with hardware RAID, also supported by some disk systems

n Incomplete page output can be detected by checksum techniques,

n But extra actions are required for recovery

n Treated as a media failure

n Log sequence number (LSN) identifies each log record

 Must be sequentially increasing

 Typically an offset from beginning of log file to allow fast access

➢ Easily extended to handle multiple log files

n Each page contains a PageLSN which is the LSN of the last log record whose effects are reflected on the page

 To update a page:

➢ X-latch the pag, and write the log record

➢ Update the page

➢ Record the LSN of the log record in PageLSN

➢ Unlock page

 Page flush to disk S-latches page

➢ Thus page state on disk is operation consistent

n Required to support physiological redo

  PageLSN is used during recovery to prevent repeated redo

➢ Thus ensuring idempotence

n DirtyPageTable

 List of pages in the buffer that have been updated

 Contains, for each such page

➢ PageLSN of the page

➢ RecLSN is an LSN such that log records before this LSN have already been applied to the
page version on disk

– Set to current end of log when a page is inserted into dirty page table (just before
being updated)

– Recorded in checkpoints, helps to minimize redo work

n Checkpoint log record

 Contains:

➢ DirtyPageTable and list of active transactions

➢ For each active transaction, LastLSN, the LSN of the last log record written by the transaction

 Fixed position on disk notes LSN of last completed

checkpoint log record

ARIES recovery involves three passes

n Analysis pass: Determines

 Which transactions to undo

 Which pages were dirty (disk version not up to date) at time of crash

 RedoLSN: LSN from which redo should start

n Redo pass:

 Repeats history, redoing all actions from RedoLSN

 ➢ RecLSN and PageLSNs are used to avoid redoing actions already reflected on page

n Undo pass:

 Rolls back all incomplete transactions

➢ Transactions whose abort was complete earlier are not undone

– Key idea: no need to undo these transactions: earlier undo actions were logged, and
are redone as required

Analysis pass

n Starts from last complete checkpoint log record

 Reads in DirtyPageTable from log record

 Sets RedoLSN = min of RecLSNs of all pages in DirtyPageTable

➢ In case no pages are dirty, RedoLSN = checkpoint record’s LSN

 Sets undo-list = list of transactions in checkpoint log record

 Reads LSN of last log record for each transaction in undo-list from checkpoint log record

n Scans forward from checkpoint

 .. On next page …

Analysis pass (cont.)

n Scans forward from checkpoint

 If any log record found for transaction not in undo-list, adds transaction to undo-list

 Whenever an update log record is found

➢ If page is not in DirtyPageTable, it is added with RecLSN set to LSN of the update log record

 If transaction end log record found, delete transaction from undo-list

 Keeps track of last log record for each transaction in undo-list

➢ May be needed for later undo

n At end of analysis pass:

 RedoLSN determines where to start redo pass

 RecLSN for each page in DirtyPageTable used to minimize redo work

 All transactions in undo-list need to be rolled back

Redo Pass: Repeats history by replaying every action not already reflected in the page on disk, as follows:

n Scans forward from RedoLSN. Whenever an update log record is found:

1. If the page is not in DirtyPageTable or the LSN of the log record is less than the RecLSN of the page in
DirtyPageTable, then skip the log record

2. Otherwise fetch the page from disk. If the PageLSN of the page fetched from disk is less than the
LSN of the log record, redo the log record

NOTE: if either test is negative the effects of the log record have already appeared on the page. First test avoids
even fetching the page from disk!
ARIES Undo Actions

n When an undo is performed for an update log record

 Generate a CLR containing the undo action performed (actions performed during undo are logged
physicaly or physiologically).

➢ CLR for record n noted as n’ in figure below

 Set UndoNextLSN of the CLR to the PrevLSN value of the update log record

➢ Arrows indicate UndoNextLSN value

n ARIES supports partial rollback

 Used e.g. to handle deadlocks by rolling back just enough to release reqd. locks

 Figure indicates forward actions after partial rollbacks

➢ records 3 and 4 initially, later 5 and 6, then full rollback

1 2 3 4 4 3 5 6 6 52 1

ARIES: Undo Pass

Undo pass

n Performs backward scan on log undoing all transaction in undo-list

 Backward scan optimized by skipping unneeded log records as follows:

➢ Next LSN to be undone for each transaction set to LSN of last log record for transaction
found by analysis pass.

➢ At each step pick largest of these LSNs to undo, skip back to it and undo it

➢ After undoing a log record

– For ordinary log records, set next LSN to be undone for transaction to PrevLSN noted
in the log record

– For compensation log records (CLRs) set next LSN to be undo to UndoNextLSN noted
in the log record

» All intervening records are skipped since they would have been undo
already

n Undos performed as described earlier

Other ARIES Features

n Recovery Independence

 Pages can be recovered independently of others

➢ E.g. if some disk pages fail they can be recovered from a backup while other pages are being
used

n Savepoints:
  Transactions can record savepoints and roll back to a savepoint

➢ Useful for complex transactions

➢ Also used to rollback just enough to release locks on deadlock

n Fine-grained locking:

 Index concurrency algorithms that permit tuple level locking on indices can be used

➢ These require logical undo, rather than physical undo, as in advanced recovery algorithm

n Recovery optimizations: For example:

 Dirty page table can be used to prefetch pages during redo

 Out of order redo is possible:

➢ redo can be postponed on a page being fetched from disk, and

performed when page is fetched.

➢ Meanwhile other log records can continue to be processed

Remote Backup Systems

n Remote backup systems provide high availability by allowing transaction processing to continue even if the
primary site is destroyed.

n Detection of failure: Backup site must detect when primary site has failed

 to distinguish primary site failure from link failure maintain several communication links between
the primary and the remote backup.

n Transfer of control:

 To take over control backup site first perform recovery using its copy of the database and all the long
records it has received from the primary.

➢ Thus, completed transactions are redone and incomplete transactions are rolled back.

 When the backup site takes over processing it becomes the new primary

 To transfer control back to old primary when it recovers, old primary must receive redo logs from
the old backup and apply all updates locally.

n Time to recover: To reduce delay in takeover, backup site periodically proceses the redo log records (in
effect, performing recovery from previous database state), performs a checkpoint, and can then delete
earlier parts of the log.
 n Hot-Spare configuration permits very fast takeover:

 Backup continually processes redo log record as they arrive, applying the updates locally.

 When failure of the primary is detected the backup rolls back incomplete transactions, and is ready
to process new transactions.

n Alternative to remote backup: distributed database with replicated data

 Remote backup is faster and cheaper, but less tolerant to failure

➢ more on this in Chapter 19

n Ensure durability of updates by delaying transaction commit until update is logged at backup; avoid this
delay by permitting lower degrees of durability.

n One-safe: commit as soon as transaction’s commit log record is written at primary

 Problem: updates may not arrive at backup before it takes over.

n Two-very-safe: commit when transaction’s commit log record is written at primary and backup

 Reduces availability since transactions cannot commit if either site fails.

n Two-safe: proceed as in two-very-safe if both primary and backup are active. If only the primary is active, the
transaction commits as soon as is commit log record is written at the primary.

 Better availability than two-very-safe; avoids problem of lost transactions in one-safe.

Block Storage Operations

Portion of the Database Log Corresponding to T0 and T1

State of the Log and Database Corresponding to T0 and T1

Portion of the System Log Corresponding to T0 and T1

State of System Log and Database Corresponding to T0 and T1