SRM

Institute of Science and

Technology
18CSC303J
Database Management System
Unit- V
 Topics covered in Unit 5

● Transaction concepts ● Recovery and Atomicity

● Properties of Transactions ● Log based recovery
● Serializability ● Concurrency problems
● Testing for Serializability ● Locking mechanism
● System Recovery ● Deadlock
● Concurrency control ● Two phase locking
● Two-phase commit protocol
protocol ● Isolation
● Intent locking
Transactions
 Transaction Concept
● A transaction is a unit of program execution that accesses and
possibly updates various data items.
● E.g., transaction to transfer $50 from account A to account B:
1. read(A)
2. A := A – 50
3. write(A)
4. read(B)
5. B := B + 50
6. write(B)
● Two main issues to deal with:
● Failures of various kinds, such as hardware failures and system
crashes
● Concurrent execution of multiple transactions
 Required Properties of a Transaction
● Consider a transaction to transfer $50 from account A to account B:
1. read(A)
2. A := A – 50
3. write(A)
4. read(B)
5. B := B + 50
6. write(B)
● Atomicity requirement
● If the transaction fails after step 3 and before step 6, money will be “lost”
leading to an inconsistent database state
4 Failure could be due to software or hardware
● The system should ensure that updates of a partially executed transaction are
not reflected in the database
● Durability requirement — once the user has been notified that the transaction has
completed (i.e., the transfer of the $50 has taken place), the updates to the database
by the transaction must persist even if there are software or hardware failures.
 Required Properties of a Transaction (Cont.)

● Consistency requirement in above example:

● The sum of A and B is unchanged by the execution of the transaction
● In general, consistency requirements include
4 Explicitly specified integrity constraints such as primary keys and
foreign keys
4 Implicit integrity constraints
– e.g., sum of balances of all accounts, minus sum of loan amounts
must equal value of cash-in-hand
● A transaction, when starting to execute, must see a consistent database.
● During transaction execution the database may be temporarily inconsistent.
● When the transaction completes successfully the database must be consistent
● Erroneous transaction logic can lead to inconsistency
 Required Properties of a Transaction (Cont.)

● Isolation requirement — if between steps 3 and 6 (of the fund transfer

transaction) , another transaction T2 is allowed to access the partially updated
database, it will see an inconsistent database (the sum A + B will be less than it
should be).

T1 T2
1. read(A)
2. A := A – 50
3. write(A)
read(A), read(B), print(A+B)
4. read(B)
5. B := B + 50
6. write(B)
● Isolation can be ensured trivially by running transactions serially
● That is, one after the other.
● However, executing multiple transactions concurrently has significant benefits,
as we will see later.
 ACID Properties

A transaction is a unit of program execution that accesses and possibly updates

various data items. To preserve the integrity of data the database system must ensure:

● Atomicity. Either all operations of the transaction are properly reflected in the
database or none are.
● Consistency. Execution of a transaction in isolation preserves the consistency
of the database.
● Isolation. Although multiple transactions may execute concurrently, each
transaction must be unaware of other concurrently executing transactions.
Intermediate transaction results must be hidden from other concurrently
executed transactions.
● That is, for every pair of transactions Ti and Tj, it appears to Ti that either Tj,
finished execution before Ti started, or Tj started execution after Ti finished.
● Durability. After a transaction completes successfully, the changes it has made
to the database persist, even if there are system failures.
 Transaction State
● Active – the initial state; the transaction stays in this state while it is
executing
● Partially committed – after the final statement has been executed.
● Failed -- after the discovery that normal execution can no longer proceed.
● Aborted – after the transaction has been rolled back and the database
restored to its state prior to the start of the transaction. Two options after it
has been aborted:
● Restart the transaction
4 can be done only if no internal logical error
● Kill the transaction
● Committed – after successful completion.
 Concurrent Executions
● Multiple transactions are allowed to run concurrently in the system.
Advantages are:
● Increased processor and disk utilization, leading to better
transaction throughput
4 E.g. one transaction can be using the CPU while another is
reading from or writing to the disk
● Reduced average response time for transactions: short
transactions need not wait behind long ones.
● Concurrency control schemes – mechanisms to achieve isolation
● That is, to control the interaction among the concurrent
transactions in order to prevent them from destroying the
consistency of the database
 Schedules
● Schedule – a sequences of instructions that specify the chronological
order in which instructions of concurrent transactions are executed
● A schedule for a set of transactions must consist of all instructions of
those transactions
● Must preserve the order in which the instructions appear in each
individual transaction.
● A transaction that successfully completes its execution will have a
commit instructions as the last statement
● By default transaction assumed to execute commit instruction as its
last step
● A transaction that fails to successfully complete its execution will have
an abort instruction as the last statement
 Schedule 1
● Let T1 transfer $50 from A to B, and T2 transfer 10% of the balance from A to B.
● An example of a serial schedule in which T1 is followed by T2 :
 Schedule 2
● A serial schedule in which T2 is followed by T1 :
 Schedule 3
● Let T1 and T2 be the transactions defined previously. The following
schedule is not a serial schedule, but it is equivalent to Schedule 1.

Note -- In schedules 1, 2 and 3, the sum “A + B” is preserved.

 Schedule 4
● The following concurrent schedule does not preserve the sum of “A
+ B”
 Serializability

● Basic Assumption – Each transaction preserves database

consistency.
● Thus, serial execution of a set of transactions preserves database
consistency.
● A (possibly concurrent) schedule is serializable if it is equivalent to a
serial schedule. Different forms of schedule equivalence give rise to
the notions of:
1. conflict serializability
2. view serializability
 Simplified view of transactions

● We ignore operations other than read and write instructions

● We assume that transactions may perform arbitrary computations
on data in local buffers in between reads and writes.
● Our simplified schedules consist of only read and write
instructions.
 Conflicting Instructions
● Let li and lj be two Instructions of transactions Ti and Tj respectively.
Instructions li and lj conflict if and only if there exists some item Q
accessed by both li and lj, and at least one of these instructions wrote Q.
1. li = read(Q), lj = read(Q). li and lj don’t conflict.
2. li = read(Q), lj = write(Q). They conflict.
3. li = write(Q), lj = read(Q). They conflict
4. li = write(Q), lj = write(Q). They conflict
● Intuitively, a conflict between li and lj forces a (logical) temporal order
between them.
● If li and lj are consecutive in a schedule and they do not conflict,
their results would remain the same even if they had been
interchanged in the schedule.
 Conflict Serializability

● If a schedule S can be transformed into a schedule S´ by a

series of swaps of non-conflicting instructions, we say that S
and S´ are conflict equivalent.
● We say that a schedule S is conflict serializable if it is conflict
equivalent to a serial schedule
 Conflict Serializability (Cont.)
● Schedule 3 can be transformed into Schedule 6 -- a serial schedule where T2
follows T1, by a series of swaps of non-conflicting instructions. Therefore,
Schedule 3 is conflict serializable.

Schedule 3 Schedule 6
 Conflict Serializability (Cont.)
● Example of a schedule that is not conflict serializable:

● We are unable to swap instructions in the above schedule to obtain

either the serial schedule < T3, T4 >, or the serial schedule < T4, T3 >.
 Precedence Graph
● Consider some schedule of a set of transactions T1, T2, ..., Tn
● Precedence graph — a direct graph where the vertices are the
transactions (names).
● We draw an arc from Ti to Tj if the two transaction conflict, and Ti
accessed the data item on which the conflict arose earlier.
● We may label the arc by the item that was accessed.
● Example
 Testing for Conflict Serializability
● A schedule is conflict serializable if and only if its
precedence graph is acyclic.
● Cycle-detection algorithms exist which take order n2
time, where n is the number of vertices in the graph.
● (Better algorithms take order n + e where e is the
number of edges.)
● If precedence graph is acyclic, the serializability order
can be obtained by a topological sorting of the graph.
● That is, a linear order consistent with the partial
order of the graph.
● For example, a serializability order for the
schedule (a) would be one of either (b) or (c)
 Recoverable Schedules
● Recoverable schedule — if a transaction Tj reads a data item previously
written by a transaction Ti , then the commit operation of Ti must appear before
the commit operation of Tj.
● The following schedule is not recoverable if T9 commits immediately after the
read(A) operation.

● If T8 should abort, T9 would have read (and possibly shown to the user) an
inconsistent database state. Hence, database must ensure that schedules are
recoverable.
 Cascading Rollbacks
● Cascading rollback – a single transaction failure leads to a series of
transaction rollbacks. Consider the following schedule where none of
the transactions has yet committed (so the schedule is recoverable)

If T10 fails, T11 and T12 must also be rolled back.

● Can lead to the undoing of a significant amount of work
 Cascadeless Schedules

● Cascadeless schedules — for each pair of transactions Ti and Tj such

that Tj reads a data item previously written by Ti, the commit operation
of Ti appears before the read operation of Tj.
● Every cascadeless schedule is also recoverable
● It is desirable to restrict the schedules to those that are cascadeless
● Example of a schedule that is NOT cascadeless
 Concurrency Control
● A database must provide a mechanism that will ensure that all possible
schedules are both:
● Conflict serializable.
● Recoverable and preferably cascadeless
● A policy in which only one transaction can execute at a time generates
serial schedules, but provides a poor degree of concurrency
● Concurrency-control schemes tradeoff between the amount of concurrency
they allow and the amount of overhead that they incur
● Testing a schedule for serializability after it has executed is a little too late!
● Tests for serializability help us understand why a concurrency control
protocol is correct
● Goal – to develop concurrency control protocols that will assure
serializability.
 Weak Levels of Consistency

● Some applications are willing to live with weak levels of consistency,

allowing schedules that are not serializable
● E.g., a read-only transaction that wants to get an approximate total
balance of all accounts
● E.g., database statistics computed for query optimization can be
approximate (why?)
● Such transactions need not be serializable with respect to other
transactions
● Tradeoff accuracy for performance
 Levels of Consistency in SQL-92
● Serializable — default
● Repeatable read — only committed records to be read, repeated reads of same
record must return same value. However, a transaction may not be serializable – it
may find some records inserted by a transaction but not find others.
● Read committed — only committed records can be read, but successive reads of
record may return different (but committed) values.
● Read uncommitted — even uncommitted records may be read.

● Lower degrees of consistency useful for gathering approximate

information about the database
● Warning: some database systems do not ensure serializable schedules by default
● E.g., Oracle and PostgreSQL by default support a level of consistency called
snapshot isolation (not part of the SQL standard)
 Transaction Definition in SQL
● Data manipulation language must include a construct for specifying
the set of actions that comprise a transaction.
● In SQL, a transaction begins implicitly.
● A transaction in SQL ends by:
● Commit work commits current transaction and begins a new
one.
● Rollback work causes current transaction to abort.
● In almost all database systems, by default, every SQL statement also
commits implicitly if it executes successfully
● Implicit commit can be turned off by a database directive
4 E.g. in JDBC, connection.setAutoCommit(false);
Other Notions of Serializability
 View Serializability

● Let S and S´ be two schedules with the same set of transactions. S and S´ are
view equivalent if the following three conditions are met, for each data item
Q,
1. If in schedule S, transaction Ti reads the initial value of Q, then in
schedule S’ also transaction Ti must read the initial value of Q.
2. If in schedule S transaction Ti executes read(Q), and that value was
produced by transaction Tj (if any), then in schedule S’ also transaction
Ti must read the value of Q that was produced by the same write(Q)
operation of transaction Tj .
3. The transaction (if any) that performs the final write(Q) operation in
schedule S must also perform the final write(Q) operation in schedule
S’.
● As can be seen, view equivalence is also based purely on reads and writes
alone.
 View Serializability (Cont.)
● A schedule S is view serializable if it is view equivalent to a serial schedule.
● Every conflict serializable schedule is also view serializable.
● Below is a schedule which is view-serializable but not conflict serializable.

● What serial schedule is above equivalent to?

● Every view serializable schedule that is not conflict serializable has blind
writes.
 Test for View Serializability
● The precedence graph test for conflict serializability cannot be used directly to
test for view serializability.
● Extension to test for view serializability has cost exponential in the size of
the precedence graph.
● The problem of checking if a schedule is view serializable falls in the class of
NP-complete problems.
● Thus, existence of an efficient algorithm is extremely unlikely.
● However ,practical algorithms that just check some sufficient conditions for
view serializability can still be used.
 More Complex Notions of Serializability
● The schedule below produces the same outcome as the serial schedule < T1, T5
>, yet is not conflict equivalent or view equivalent to it.

● If we start with A = 1000 and B = 2000, the final result is 960 and 2040
● Determining such equivalence requires analysis of operations other than read
and write.
Concurrency Control
 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 the concurrency-control manager by the
programmer. Transaction can proceed only after request is granted.
 Lock-Based Protocols (Cont.)
● 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.
 Lock-Based Protocols (Cont.)
● 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.
 The Two-Phase Locking Protocol

● This protocol 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).
 The Two-Phase Locking Protocol (Cont.)

● There can be conflict serializable schedules that cannot be obtained if

two-phase locking is used.
● However, in the absence of extra information (e.g., ordering of access
to data), two-phase locking is needed for conflict serializability in the
following sense:
● Given a transaction Ti that does not follow two-phase locking, we
can find a transaction Tj that uses two-phase locking, and a
schedule for Ti and Tj that is not conflict serializable.
 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
 Automatic Acquisition of Locks (Cont.)
● write(D) is processed as:
if Ti has a lock-X on D
then
write(D)
else begin
if necessary wait until no other transaction 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
 Deadlocks
● 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.
 Deadlocks (Cont.)

● Two-phase locking does not ensure freedom from deadlocks.

● In addition to deadlocks, there is a possibility of starvation.
● Starvation occurs if the 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.
 Deadlocks (Cont.)
● The potential for deadlock exists in most locking protocols.
Deadlocks are a necessary evil.
● When a deadlock occurs there is a possibility of cascading roll-backs.
● Cascading roll-back is possible under two-phase locking. To avoid
this, follow a modified protocol called strict two-phase locking -- 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.
 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
Lock Table
● Dark blue rectangles indicate granted locks;
light blue indicate waiting requests
● Lock table also records the type of lock
granted or requested
● New request is added to the end of the queue
of requests for the data item, and granted if it
is compatible with all earlier locks
● Unlock requests result in the request being
deleted, and later requests are checked to see
if they can now be granted
● If transaction aborts, all waiting or granted
requests of the transaction are deleted
● lock manager may keep a list of locks
held by each transaction, to implement
this efficiently
 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.
 More Deadlock Prevention Strategies
● Following schemes use transaction timestamps for the sake of deadlock
prevention alone.
● wait-die scheme — non-preemptive
● older transaction may wait for younger one to release data item. (older
means smaller timestamp) Younger transactions never Younger
transactions never wait for older ones; they are rolled back instead.
● a transaction may die several times before acquiring needed data item
● wound-wait scheme — preemptive
● older transaction wounds (forces rollback) of younger transaction instead
of waiting for it. Younger transactions may wait for older ones.
● may be fewer rollbacks than wait-die scheme.
 Deadlock prevention (Cont.)
● Both in wait-die and in wound-wait schemes, a rolled back transactions is
restarted with its original timestamp. Older transactions thus have precedence
over newer ones, and starvation is hence avoided.
● Timeout-Based Schemes:
● a transaction waits for a lock only for a specified amount of time. If the
lock has not been granted within that time, the transaction is rolled back
and restarted,
● Thus, deadlocks are not possible
● simple to implement; but starvation is possible. Also difficult to
determine good value of the timeout interval.
 Deadlock Detection
● Deadlocks can be described as a wait-for graph, which consists of a pair G =
(V,E),
● V is a set of vertices (all the transactions in the system)
● E is a set of edges; each element is an ordered pair Ti →Tj.
● If Ti → Tj is in E, then there is a directed edge from Ti to Tj, implying that Ti
is waiting for Tj to release a data item.
● When Ti requests a data item currently being held by Tj, then the edge Ti → Tj
is inserted in the wait-for graph. This edge is removed only when Tj is no
longer holding a data item needed by Ti.
● The system is in a deadlock state if and only if the wait-for graph has a cycle.
Must invoke a deadlock-detection algorithm periodically to look for cycles.
 Deadlock Detection (Cont.)

Wait-for graph without a cycle Wait-for graph with a cycle

 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
 Multiple Granularity
● Allow data items to be of various sizes and define a hierarchy of data
granularities, where the small granularities are nested within larger ones
● Can be represented graphically as a tree.
● When a transaction locks a node in the tree explicitly, it implicitly locks all
the node's descendents in the same mode.
● Granularity of locking (level in tree where locking is done):
● fine granularity (lower in tree): high concurrency, high locking overhead
● coarse granularity (higher in tree): low locking overhead, low
concurrency
 Example of Granularity Hierarchy

The levels, starting from the coarsest (top) level are

● database
● area
● file
● record
 Intention Lock Modes
● In addition to S and X lock modes, there are three additional lock modes with
multiple granularity:
● intention-shared (IS): indicates explicit locking at a lower level of the
tree but only with shared locks.
● intention-exclusive (IX): indicates explicit locking at a lower level with
exclusive or shared locks
● shared and intention-exclusive (SIX): the subtree rooted by that node is
locked explicitly in shared mode and explicit locking is being done at a
lower level with exclusive-mode locks.
● intention locks allow a higher level node to be locked in S or X mode without
having to check all descendent nodes.
 Compatibility Matrix with Intention Lock Modes

● The compatibility matrix for all lock modes is:

 Multiple Granularity Locking Scheme
● Transaction Ti can lock a node Q, using the following rules:
1. The lock compatibility matrix must be observed.
2. The root of the tree must be locked first, and may be locked in any mode.
3. A node Q can be locked by Ti in S or IS mode only if the parent of Q is
currently locked by Ti in either IX or IS mode.
4. A node Q can be locked by Ti in X, SIX, or IX mode only if the parent of Q
is currently locked by Ti in either IX or SIX mode.
5. Ti can lock a node only if it has not previously unlocked any node (that is, Ti
is two-phase).
6. Ti can unlock a node Q only if none of the children of Q are currently locked
by Ti.
● Observe that locks are acquired in root-to-leaf order, whereas they are released in
leaf-to-root order.
● Lock granularity escalation: in case there are too many locks at a particular
level, switch to higher granularity S or X lock
 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.
 Timestamp-Based Protocols (Cont.)
● 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)
1. 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.
2. 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)).
 Timestamp-Based Protocols (Cont.)
● Suppose that transaction Ti issues write(Q).
1. 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.
2. 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.
3. 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
 Thomas’ Write Rule
● Modified version of the timestamp-ordering protocol in which obsolete write
operations may be ignored under certain circumstances.
● When Ti attempts to write data item Q, if TS(Ti) < W-timestamp(Q), then Ti
is attempting to write an obsolete value of {Q}.
● Rather than rolling back Ti as the timestamp ordering protocol would
have done, this {write} operation can be ignored.
● Otherwise this protocol is the same as the timestamp ordering protocol.
● Thomas' Write Rule allows greater potential concurrency.
● Allows some view-serializable schedules that are not
conflict-serializable.
 Validation-Based Protocol
● 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-Based Protocol (Cont.)
● Each transaction Ti has 3 timestamps
● Start(Ti) : the time when Ti started its execution
● Validation(Ti): the time when Ti entered its validation phase
● Finish(Ti) : the time when Ti finished its write phase
● Serializability order is determined by timestamp given at validation time; this
is done to increase concurrency.
● Thus, TS(Ti) is given the value of Validation(Ti).
● This protocol is useful and gives greater degree of concurrency if probability
of conflicts is low.
● because the serializability order is not pre-decided, and
● relatively few transactions will have to be rolled back.
 Validation Test for Transaction Tj
● If for all Ti with TS (Ti) < TS (Tj) either one of the following condition
holds:
● finish(Ti) < start(Tj)
● start(Tj) < finish(Ti) < validation(Tj) and the set of data items
written by Ti does not intersect with the set of data items read by Tj.
then validation succeeds and Tj can be committed. Otherwise, validation
fails and Tj is aborted.
● Justification: Either the first condition is satisfied, and there is no
overlapped execution, or the second condition is satisfied and
● the writes of Tj do not affect reads of Ti since they occur after Ti has
finished its reads.
● the writes of Ti do not affect reads of Tj since Tj does not read any
item written by Ti.
 Schedule Produced by Validation
● Example of schedule produced using validation
 Multiversion Schemes
● Multiversion schemes keep old versions of data item to increase concurrency.
● Multiversion Timestamp Ordering
● Multiversion Two-Phase Locking
● Each successful write results in the creation of a new version of the data item
written.
● Use timestamps to label versions.
● When a read(Q) operation is issued, select an appropriate version of Q based
on the timestamp of the transaction, and return the value of the selected
version.
● reads never have to wait as an appropriate version is returned immediately.
 Multiversion Timestamp Ordering
● Each data item Q has a sequence of versions <Q1, Q2,...., Qm>. Each version
Qk contains three data fields:
● Content -- the value of version Qk.
● W-timestamp(Qk) -- timestamp of the transaction that created (wrote)
version Qk
● R-timestamp(Qk) -- largest timestamp of a transaction that successfully
read version Qk
● When a transaction Ti creates a new version Qk of Q, Qk's W-timestamp and
R-timestamp are initialized to TS(Ti).
● R-timestamp of Qk is updated whenever a transaction Tj reads Qk, and TS(Tj)
> R-timestamp(Qk).
 Multiversion Timestamp Ordering (Cont)
● Suppose that transaction Ti issues a read(Q) or write(Q) operation. Let Qk
denote the version of Q whose write timestamp is the largest write timestamp
less than or equal to TS(Ti).
1. If transaction Ti issues a read(Q), then the value returned is the content
of version Qk.
2. If transaction Ti issues a write(Q)
1. if TS(Ti) < R-timestamp(Qk), then transaction Ti is rolled back.
2. if TS(Ti) = W-timestamp(Qk), the contents of Qk are overwritten
3. else a new version of Q is created.
● Observe that
● Reads always succeed
● A write by Ti is rejected if some other transaction Tj that (in the
serialization order defined by the timestamp values) should read
Ti's write, has already read a version created by a transaction older than Ti.
● Protocol guarantees serializability
 Multiversion Two-Phase Locking
● Differentiates between read-only transactions and update transactions
● Update transactions acquire read and write locks, and hold all locks up to the
end of the transaction. That is, update transactions follow rigorous two-phase
locking.
● Each successful write results in the creation of a new version of the data
item written.
● Each version of a data item has a single timestamp whose value is obtained
from a counter ts-counter that is incremented during commit processing.
● Read-only transactions are assigned a timestamp by reading the current value
of ts-counter before they start execution; they follow the multiversion
timestamp-ordering protocol for performing reads.
 Multiversion Two-Phase Locking (Cont.)
● When an update transaction wants to read a data item:
● it obtains a shared lock on it, and reads the latest version.
● When it wants to write an item
● it obtains X lock on; it then creates a new version of the item and sets this
version's timestamp to ∞.
● When update transaction Ti completes, commit processing occurs:
● Ti sets timestamp on the versions it has created to ts-counter + 1
● Ti increments ts-counter by 1
● Read-only transactions that start after Ti increments ts-counter will see the
values updated by Ti.
● Read-only transactions that start before Ti increments the
ts-counter will see the value before the updates by Ti.
● Only serializable schedules are produced.
 MVCC: Implementation Issues
● Creation of multiple versions increases storage overhead
● Extra tuples
● Extra space in each tuple for storing version information
● Versions can, however, be garbage collected
● E.g. if Q has two versions Q5 and Q9, and the oldest active transaction
has timestamp > 9, than Q5 will never be required again
 Snapshot Isolation
● Motivation: Decision support queries that read large amounts of data have
concurrency conflicts with OLTP transactions that update a few rows
● Poor performance results
● Solution 1: Give logical “snapshot” of database state to read only transactions,
read-write transactions use normal locking
● Multiversion 2-phase locking
● Works well, but how does system know a transaction is read only?
● Solution 2: Give snapshot of database state to every transaction, updates alone
use 2-phase locking to guard against concurrent updates
● Problem: variety of anomalies such as lost update can result
● Partial solution: snapshot isolation level (next slide)
4 Proposed by Berenson et al, SIGMOD 1995
4 Variants implemented in many database systems
– E.g. Oracle, PostgreSQL, SQL Server 2005
 Snapshot Isolation
● A transaction T1 executing with Snapshot T1 T2 T3
Isolation
W(Y := 1)
● takes snapshot of committed data at start
Commit
● always reads/modifies data in its own
snapshot Start
● updates of concurrent transactions are not R(X) 0
visible to T1 R(Y) 1
● writes of T1 complete when it commits
● First-committer-wins rule: W(X:=2)
4 Commits only if no other concurrent W(Z:=3)
transaction has already written data that Commit
T1 intends to write.
R(Z) 0
R(Y) 1
W(X:=3)
Concurrent updates not visible
Commit-Req
Own updates are visible
Not first-committer of X Abort
Serialization error, T2 is rolled back
 Snapshot Read
● Concurrent updates invisible to snapshot read
Snapshot Write: First Committer Wins

● Variant: “First-updater-wins”
4 Check for concurrent updates when write occurs by locking item
– But lock should be held till all concurrent transactions have finished
4 (Oracle uses this plus some extra features)
4 Differs only in when abort occurs, otherwise equivalent
 Benefits of Snapshot Isolation
● Reading is never blocked,
● and also doesn’t block other txns activities
● Performance similar to Read Committed
● Avoids the usual anomalies
● No dirty read
● No lost update
● No non-repeatable read
● Predicate based selects are repeatable (no phantoms)
● Problems with SI
● SI does not always give serializable executions
4 Serializable: among two concurrent txns, one sees the effects of the
other
4 In SI: neither sees the effects of the other
● Result: Integrity constraints can be violated
 Snapshot Isolation
● E.g. of problem with SI
● T1: x:=y
● T2: y:= x
● Initially x = 3 and y = 17
4 Serial execution: x = ??, y = ??
4 if both transactions start at the same time, with snapshot isolation: x
= ?? , y = ??
● Called skew write
● Skew also occurs with inserts
● E.g:
4 Find max order number among all orders
4 Create a new order with order number = previous max + 1
 Snapshot Isolation Anomalies
● SI breaks serializability when txns modify different items, each based on a previous
state of the item the other modified
● Not very common in practice
4 E.g., the TPC-C benchmark runs correctly under SI
4 when txns conflict due to modifying different data, there is usually also a
shared item they both modify too (like a total quantity) so SI will abort one of
them
● But does occur
4 Application developers should be careful about write skew
● SI can also cause a read-only transaction anomaly, where read-only transaction may see
an inconsistent state even if updaters are serializable
● We omit details
● Using snapshots to verify primary/foreign key integrity can lead to inconsistency
● Integrity constraint checking usually done outside of snapshot
 SI In Oracle and PostgreSQL
● Warning: SI used when isolation level is set to serializable, by Oracle, and
PostgreSQL versions prior to 9.1
● PostgreSQL’s implementation of SI (versions prior to 9.1) described in
Section 26.4.1.3
● Oracle implements “first updater wins” rule (variant of “first committer wins”)
4 concurrent writer check is done at time of write, not at commit time
4 Allows transactions to be rolled back earlier
4 Oracle and PostgreSQL < 9.1 do not support true serializable execution
● PostgreSQL 9.1 introduced new protocol called “Serializable Snapshot
Isolation” (SSI)
4 Which guarantees true serializabilty including handling predicate reads
(coming up)
 SI In Oracle and PostgreSQL
● Can sidestep SI for specific queries by using select .. for update in Oracle and
PostgreSQL
● E.g.,
1. select max(orderno) from orders for update
2. read value into local variable maxorder
3. insert into orders (maxorder+1, …)
● Select for update (SFU) treats all data read by the query as if it were also
updated, preventing concurrent updates
● Does not always ensure serializability since phantom phenomena can occur
(coming up)
● In PostgreSQL versions < 9.1, SFU locks the data item, but releases locks when the
transaction completes, even if other concurrent transactions are active
● Not quite same as SFU in Oracle, which keeps locks until all
● concurrent transactions have completed
 Insert and Delete Operations
● If two-phase locking is used :
● A delete operation may be performed only if the transaction deleting the
tuple has an exclusive lock on the tuple to be deleted.
● A transaction that inserts a new tuple into the database is given an
X-mode lock on the tuple
● Insertions and deletions 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
 Insert and Delete Operations (Cont.)
● The transaction scanning the relation is reading information that indicates what
tuples the relation contains, while a transaction inserting a tuple updates the same
information.
● The conflict should be detected, e.g. by locking the information.
● One solution:
● Associate a data item with the relation, to represent the information about what
tuples the relation contains.
● Transactions scanning the relation acquire a shared lock in the data item,
● Transactions inserting or deleting a tuple acquire an exclusive lock on the data
item. (Note: locks on the data item do not conflict with locks on individual
tuples.)
● Above protocol provides very low concurrency for insertions/deletions.
● Index locking protocols provide higher concurrency while
preventing the phantom phenomenon, by requiring locks
on certain index buckets.
 Index Locking Protocol
● Index locking protocol:
● Every relation must have at least one index.
● A transaction can access tuples only after finding them through one or more
indices on the relation
● A transaction Ti that performs a lookup must lock all the index leaf nodes that
it accesses, in S-mode
4 Even if the leaf node does not contain any tuple satisfying the index
lookup (e.g. for a range query, no tuple in a leaf is in the range)
● A transaction Ti that inserts, updates or deletes a tuple ti in a relation r
4 must update all indices to r
4 must obtain exclusive locks on all index leaf nodes affected by the
insert/update/delete
● The rules of the two-phase locking protocol must be observed
● Guarantees that phantom phenomenon won’t occur
 Next-Key Locking
● Index-locking protocol to prevent phantoms required locking entire leaf
● Can result in poor concurrency if there are many inserts
● Alternative: for an index lookup
● Lock all values that satisfy index lookup (match lookup value, or fall in
lookup range)
● Also lock next key value in index
● Lock mode: S for lookups, X for insert/delete/update
● Ensures that range queries will conflict with inserts/deletes/updates
● Regardless of which happens first, as long as both are concurrent
 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.
 Concurrency in Index Structures (Cont.)
● Example of index concurrency protocol:
● Use crabbing instead of two-phase locking on the nodes of the B+-tree, as follows.
During search/insertion/deletion:
● First lock the root node in shared mode.
● After locking all required children of a node in shared mode, release the lock on the
node.
● During insertion/deletion, upgrade leaf node locks to exclusive mode.
● When splitting or coalescing requires changes to a parent, lock the parent in
exclusive mode.
● Above protocol can cause excessive deadlocks
● Searches coming down the tree deadlock with updates going up the tree
● Can abort and restart search, without affecting transaction
● Better protocols are available; see Section 16.9 for one such protocol, the B-link tree
protocol
● Intuition: release lock on parent before acquiring lock on child
4 And deal with changes that may have happened between lock release and
acquire
 Weak Levels of Consistency
● Degree-two consistency: differs from two-phase locking in that S-locks may
be released at any time, and locks may be acquired at any time
● X-locks must be held till end of transaction
● Serializability is not guaranteed, programmer must ensure that no
erroneous database state will occur]
● Cursor stability:
● For reads, each tuple is locked, read, and lock is immediately released
● X-locks are held till end of transaction
● Special case of degree-two consistency
 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
 Transactions across User Interaction
● Many applications need transaction support across user interactions
● Can’t use locking
● Don’t want to reserve database connection per user
● Application level concurrency control
● Each tuple has a version number
● Transaction notes version number when reading tuple
4 select r.balance, r.version into :A, :version
from r where acctId =23
● When writing tuple, check that current version number is same as the version
when tuple was read
4 update r set r.balance = r.balance + :deposit
where acctId = 23 and r.version = :version
● Equivalent to optimistic concurrency control without validating read set
● Used internally in Hibernate ORM system, and manually in many applications
● Version numbering can also be used to support first committer wins check of snapshot
isolation
● Unlike SI, reads are not guaranteed to be from a single snapshot
 Deadlocks
● Consider the following two transactions:
T1: write (X) T2: write(Y)
write(Y) write(X)

● Schedule with deadlock

 Recovery System
● Failure Classification
● Storage Structure
● Recovery and Atomicity
● Log-Based Recovery
● Remote Backup Systems
 Failure Classification
● Transaction failure :
● Logical errors: transaction cannot complete due to some internal error
condition
● System errors: the database system must terminate an active transaction
due to an error condition (e.g., deadlock)
● System crash: a power failure or other hardware or software failure causes
the system to crash.
● Fail-stop assumption: non-volatile storage contents are assumed to not
be corrupted by system crash
4 Database systems have numerous integrity checks to prevent
corruption of disk data
● Disk failure: a head crash or similar disk failure destroys all or part of disk
storage
● Destruction is assumed to be detectable: disk drives use checksums to
detect failures
 Recovery Algorithms
● Consider transaction Ti that transfers $50 from account A to account B
● Two updates: subtract 50 from A and add 50 to B
● Transaction Ti requires updates to A and B to be output to the database.
● A failure may occur after one of these modifications have been made but
before both of them are made.
● Modifying the database without ensuring that the transaction will
commit may leave the database in an inconsistent state
● Not modifying the database may result in lost updates if failure occurs
just after transaction commits
● Recovery algorithms have two parts
1. Actions taken during normal transaction processing to ensure enough
information exists to recover from failures
2. Actions taken after a failure to recover the database contents to a state
that ensures atomicity, consistency and durability
 Storage Structure
● Volatile storage:
● does not survive system crashes
● examples: main memory, cache memory
● Nonvolatile storage:
● survives system crashes
● examples: disk, tape, flash memory,
non-volatile (battery backed up) RAM
● but may still fail, losing data
● Stable storage:
● a mythical form of storage that survives all failures
● approximated by maintaining multiple copies on distinct nonvolatile
media
● See book for more details on how to implement stable storage
 Stable-Storage Implementation
● Maintain multiple copies of each block on separate disks
● copies can be at remote sites to protect against disasters such as fire or
flooding.
● 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
● 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.
 Stable-Storage Implementation (Cont.)

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

● Copies of a block may differ due to failure during output operation. To recover from
failure:
1. First find inconsistent blocks:
1. Expensive solution: Compare the two copies of every disk block.
2. Better solution:
● Record in-progress disk writes on non-volatile storage (Non-volatile
RAM or special area of disk).
● Use this information during recovery to find blocks that may be
inconsistent, and only compare copies of these.
● Used in hardware RAID systems
2. 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
● Physical blocks are those blocks residing on the disk.
● Buffer blocks are the blocks residing temporarily in main memory.
● 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.
● We assume, for simplicity, that each data item fits in, and is stored inside, a
single block.
 Example of Data Access
buffer
Buffer Block A input(A)
X A
Buffer Block B Y B
output(B)
read(X)
write(Y)

x2
x1
y1

work area work area

of T1 of T2

memory disk
 Data Access (Cont.)

● 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.
● Transferring data items between system buffer blocks and its private
work-area done by:
● 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.
● Note: output(BX) need not immediately follow write(X). System can
perform the output operation when it deems fit.
● Transactions
● Must perform read(X) before accessing X for the first time (subsequent
reads can be from local copy)
● write(X) can be executed at any time before the transaction commits
 Recovery and Atomicity
● To ensure atomicity despite failures, we first output information describing
the modifications to stable storage without modifying the database itself.
● We study log-based recovery mechanisms in detail
● We first present key concepts
● And then present the actual recovery algorithm
● Less used alternative: shadow-copy and shadow-paging (brief details in
book)

shadow-copy
 Log-Based Recovery
● 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.
● When transaction Ti starts, it registers itself by writing a
<Ti start>log record
● Before Ti executes write(X), a log record
<Ti, X, V1, V2>
is written, where V1 is the value of X before the write (the old value), and V2 is
the value to be written to X (the new value).
● When Ti finishes it last statement, the log record <Ti commit> is written.
● Two approaches using logs
● Deferred database modification
● Immediate database modification
 Immediate Database Modification
● The immediate-modification scheme allows updates of an uncommitted
transaction to be made to the buffer, or the disk itself, before the transaction
commits
● Update log record must be written before database item is written
● We assume that the log record is output directly to stable storage
● (Will see later that how to postpone log record output to some extent)
● Output of updated blocks to stable storage can take place at any time before
or after transaction commit
● Order in which blocks are output can be different from the order in which
they are written.
● The deferred-modification scheme performs updates to buffer/disk only at
the time of transaction commit
● Simplifies some aspects of recovery
● But has overhead of storing local copy
 Transaction Commit
● A transaction is said to have committed when its commit log record is output
to stable storage
● all previous log records of the transaction must have been output already
● Writes performed by a transaction may still be in the buffer when the
transaction commits, and may be output later
 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 BC output before T1
BB , BC commits
<T1 commit>
BA
BA output after T0
● Note: BX denotes block containing X.
commits
 Concurrency Control and Recovery
● With concurrent transactions, all transactions share a single disk buffer and a
single log
● A buffer block can have data items updated by one or more transactions
● We assume that if a transaction Ti has modified an item, no other transaction
can modify the same item until Ti has committed or aborted
● i.e. the updates of uncommitted transactions should not be visible to
other transactions
4 Otherwise how to perform undo if T1 updates A, then T2 updates A
and commits, and finally T1 has to abort?
● Can be ensured by obtaining exclusive locks on updated items and
holding the locks till end of transaction (strict two-phase locking)
● Log records of different transactions may be interspersed in the log.
 Undo and Redo Operations
● Undo of a log record <Ti, X, V1, V2> writes the old value V1 to X
● Redo of a log record <Ti, X, V1, V2> writes the new value V2 to X
● Undo and Redo of Transactions
● 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
4 each time a data item X is restored to its old value V a special log
record <Ti , X, V> is written out
4 when undo of a transaction is complete, a log record
<Ti abort> is written out.
● 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
4 No logging is done in this case
 Undo and Redo on Recovering from Failure

● When recovering after failure:

● Transaction Ti needs to be undone if the log
4 contains the record <Ti start>,
4 but does not contain either the record <Ti commit> or <Ti abort>.
● Transaction Ti needs to be redone if the log
4 contains the records <Ti start>
4 and contains the record <Ti commit> or <Ti abort>
● Note that If transaction Ti was undone earlier and the <Ti abort> record written to
the log, and then a failure occurs, on recovery from failure Ti is redone
● such a redo redoes all the original actions including the steps that restored
old values
4 Known as repeating history
4 Seems wasteful, but simplifies recovery greatly
Immediate DB Modification Recovery Example
Below we show the log as it appears at three instances of time.

Recovery actions in each case above are:

(a) undo (T0): B is restored to 2000 and A to 1000, and log records
<T0, B, 2000>, <T0, A, 1000>, <T0, abort> are written out
(b) redo (T0) and undo (T1): A and B are set to 950 and 2050 and C is restored to 700.
Log records <T1, C, 700>, <T1, abort> are written out.
(c) redo (T0) and redo (T1): A and B are set to 950 and 2050
respectively. Then C is set to 600
 Checkpoints
● Redoing/undoing all transactions recorded in the log can be very slow
1. processing the entire log is time-consuming if the system has run for a
long time
2. we might unnecessarily redo transactions which have already output
their updates to the database.
● 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 L> onto stable storage where L is a list
of all transactions active at the time of checkpoint.
● All updates are stopped while doing checkpointing
 Checkpoints (Cont.)
● 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 L>
record
● Only transactions that are in L or started after the checkpoint need to be
redone or undone
● Transactions that committed or aborted before the checkpoint already
have all their updates output to stable storage.
● Some earlier part of the log may be needed for undo operations
1. Continue scanning backwards till a record <Ti start> is found for every
transaction Ti in L.
● Parts of log prior to earliest <Ti start> record above are not needed for
recovery, and can be erased whenever desired.
 Example of Checkpoints

Tc Tf
T1
T2
T3
T4

checkpoint system failure

● T1 can be ignored (updates already output to disk due to checkpoint)
● T2 and T3 redone.
● T4 undone
 Recovery Algorithm

● So far: we covered key concepts

● Now: we present the components of the basic recovery algorithm
● Later: we present extensions to allow more concurrency
 Recovery Algorithm
● Logging (during normal operation):
● <Ti start> at transaction start
● <Ti, Xj, V1, V2> for each update, and
● <Ti commit> at transaction end
● Transaction rollback (during normal operation)
● Let Ti be the transaction to be rolled back
● Scan log backwards from the end, and for each log record of Ti of the form
<Ti, Xj, V1, V2>
4 perform the undo by writing V1 to Xj,
4 write a log record <Ti , Xj, V1>
– such log records are called compensation log records
● Once the record <Ti start> is found stop the scan and write the log record
<Ti abort>
 Recovery Algorithm (Cont.)
● Recovery from failure: Two phases
● Redo phase: replay updates of all transactions, whether they
committed, aborted, or are incomplete
● Undo phase: undo all incomplete transactions
● Redo phase:
1. Find last <checkpoint L> record, and set undo-list to L.
2. Scan forward from above <checkpoint L> record
1. Whenever a record <Ti, Xj, V1, V2> or <Ti, Xj, V2> is found, redo
it by writing V2 to Xj
2. Whenever a log record <Ti start> is found, add Ti to undo-list
3. Whenever a log record <Ti commit> or <Ti abort> is found,
remove Ti from undo-list
 Recovery Algorithm (Cont.)
● Undo phase:
1. Scan log backwards from end
1. Whenever a log record <Ti, Xj, V1, V2> is found where Ti is in
undo-list perform same actions as for transaction rollback:
1. perform undo by writing V1 to Xj.
2. write a log record <Ti , Xj, V1>
2. Whenever a log record <Ti start> is found where Ti is in undo-list,
1. Write a log record <Ti abort>
2. Remove Ti from undo-list
3. Stop when undo-list is empty
● i.e. <Ti start> has been found for every transaction in undo-list
● After undo phase completes, normal transaction processing can commence
Example of Recovery
 Log Record Buffering
● 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.
● Log force is performed to commit a transaction by forcing all its log records
(including the commit record) to stable storage.
● Several log records can thus be output using a single output operation, reducing
the I/O cost.
 Log Record Buffering (Cont.)
● 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.
4 This rule is called the write-ahead logging or WAL rule
– Strictly speaking WAL only requires undo information to be
output
 Database Buffering
● 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
● The recovery algorithm supports the no-force policy: i.e., updated blocks need
not be written to disk when transaction commits
● force policy: requires updated blocks to be written at commit
4 More expensive commit
● The recovery algorithm supports the steal policy:i.e., blocks containing updates of
uncommitted transactions can be written to disk, even before the transaction
commits
 Database Buffering (Cont.)
● 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
● (Write ahead logging)
● 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.
4 Such locks held for short duration are called latches.
● To output a block to disk
1. First acquire an exclusive latch on the block
1. Ensures no update can be in progress on the block
2. Then perform a log flush
3. Then output the block to disk
4. Finally release the latch on the block
 Buffer Management (Cont.)
● Database buffer can be implemented either
● in an area of real main-memory reserved for the database, or
● in virtual memory
● 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.
 Buffer Management (Cont.)
● 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, 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!
4 Known as dual paging problem.
● Ideally when OS needs to evict a page from the buffer, it should pass
control to database, which in turn should
1. Output the page to database instead of to swap space (making sure to
output log records first), if it is modified
2. Release the page from the buffer, for the OS to use
Dual paging can thus be avoided, but common operating systems do not
support such functionality.
 Fuzzy Checkpointing
● To avoid long interruption of normal processing during checkpointing,
allow updates to happen during checkpointing
● Fuzzy checkpointing is done as follows:
1. Temporarily stop all updates by transactions
2. Write a <checkpoint L> log record and force log to stable storage
3. Note list M of modified buffer blocks
4. Now permit transactions to proceed with their actions
5. Output to disk all modified buffer blocks in list M
● blocks should not be updated while being output
● Follow WAL: all log records pertaining to a block must be
output before the block is output
6. Store a pointer to the checkpoint record in a fixed position
last_checkpoint on disk
 Fuzzy Checkpointing (Cont.)
● 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

……
<checkpoint L>
…..
<checkpoint L>
last_checkpoint
…..

Log
 Failure with Loss of Nonvolatile Storage
● So far we assumed no loss of non-volatile storage
● 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
4 Output all log records currently residing in main memory onto
stable storage.
4 Output all buffer blocks onto the disk.
4 Copy the contents of the database to stable storage.
4 Output a record <dump> to log on stable storage.
 Recovering from Failure of Non-Volatile Storage
● To recover from disk failure
● restore database from most recent dump.
● Consult the log and redo all transactions that committed after the
dump
● Can be extended to allow transactions to be active during dump;
known as fuzzy dump or online dump
● Similar to fuzzy checkpointing
Recovery with Early Lock Release and
Logical Undo Operations
 Recovery with Early Lock Release
● Support for high-concurrency locking techniques, such as those used for
B+-tree concurrency control, which release locks early
● Supports “logical undo”
● Recovery based on “repeating history”, whereby recovery executes exactly
the same actions as normal processing
 Logical Undo Logging
● 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).
● For such operations, undo log records should contain the undo operation to be
executed
● Such logging is called logical undo logging, in contrast to physical undo
logging
4 Operations are called logical operations
● Other examples:
4 delete of tuple, to undo insert of tuple
– allows early lock release on space allocation information
4 subtract amount deposited, to undo deposit
– allows early lock release on bank balance
 Physical Redo
● Redo information is logged physically (that is, new value for each write)
even for operations with logical undo
● Logical redo is very complicated since database state on disk may not be
“operation consistent” when recovery starts
● Physical redo logging does not conflict with early lock release
 Operation Logging
● Operation logging is done as follows:
1. When operation starts, log <Ti, Oj, operation-begin>. Here Oj is a unique
identifier of the operation instance.
2. While operation is executing, normal log records with physical redo and
physical undo information are logged.
3. When operation completes, <Ti, Oj, operation-end, U> is logged, where U
contains information needed to perform a logical undo information.
Example: insert of (key, record-id) pair (K5, RID7) into index I9

<T1, O1, operation-begin>

….
<T1, X, 10, K5> Physical redo of steps in insert
<T1, Y, 45, RID7>
<T1, O1, operation-end, (delete I9, K5, RID7)>
 Operation Logging (Cont.)
● 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.
● 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.
● Redo of operation (after crash) still uses physical redo information.
 Transaction Rollback with Logical Undo
Rollback of transaction Ti is done as follows:
● Scan the log backwards
1. If a log record <Ti, X, V1, V2> is found, perform the undo and log a al <Ti,
X, V1>.
2. If a <Ti, Oj, operation-end, U> record is found
4 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>.
4 Skip all preceding log records for Ti until the record
<Ti, Oj operation-begin> is found
 Transaction Rollback with Logical Undo (Cont.)

● Transaction rollback, scanning the log backwards (cont.):

3. If a redo-only record is found ignore it
4. If a <Ti, Oj, operation-abort> record is found:
● 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:
● Cases 3 and 4 above can occur only if the database crashes while a
transaction is being rolled back.
● Skipping of log records as in case 4 is important to prevent multiple rollback
of the same operation.
 Transaction Rollback with Logical Undo

● Transaction rollback during normal

operation
Failure Recovery with Logical Undo
 Transaction Rollback: Another Example
● Example with a complete and an incomplete operation
<T1, start>
<T1, O1, operation-begin>
….
<T1, X, 10, K5>
<T1, Y, 45, RID7>
<T1, O1, operation-end, (delete I9, K5, RID7)>
<T1, O2, operation-begin>
<T1, Z, 45, 70>
T1 Rollback begins here
<T1, Z, 45> redo-only log record during physical undo (of incomplete O2)
<T1, Y, .., ..> Normal redo records for logical undo of O1
…
<T1, O1, operation-abort> What if crash occurred immediately after this?
<T1, abort>
 Recovery Algorithm with Logical Undo
Basically same as earlier algorithm, except for changes described earlier for
transaction rollback
1. (Redo phase): Scan log forward from last < checkpoint L> record till end of
log
1. Repeat history by physically redoing all updates of all transactions,
2. Create an undo-list during the scan as follows
4 undo-list is set to L initially
4 Whenever <T start> is found T is added to undo-list
i i
4 Whenever <T commit> or <T abort> is found, T is deleted from
i i i
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 with Logical Undo (Cont.)
Recovery from system crash (cont.)
2. (Undo phase): Scan log backwards, performing undo on log records of
transactions found in undo-list.
● Log records of transactions being rolled back are processed as described
earlier, as they are found
4 Single shared scan for all transactions being undone
● 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
● This undoes the effects of incomplete transactions (those with neither
commit nor abort log records). Recovery is now complete.
ARIES Recovery Algorithm
 ARIES
● ARIES is a state of the art recovery method
● Incorporates numerous optimizations to reduce overheads during normal
processing and to speed up recovery
● The recovery algorithm we studied earlier is modeled after ARIES, but
greatly simplified by removing optimizations
● Unlike the recovery algorithm described earlier, ARIES
1. Uses log sequence number (LSN) to identify log records
4 Stores LSNs in pages to identify what updates have already been
applied to a database page
2. Physiological redo
3. Dirty page table to avoid unnecessary redos during recovery
4. Fuzzy checkpointing that only records information about dirty pages,
and does not require dirty pages to be written out at checkpoint time
4 More coming up on each of the above …
 ARIES Optimizations
● Physiological redo
● Affected page is physically identified, action within page can be logical
4 Used to reduce logging overheads
– e.g. when a record is deleted and all other records have to be moved
to fill hole
» Physiological redo can log just the record deletion
» Physical redo would require logging of old and new values for
much of the page
4 Requires page to be output to disk atomically
– Easy to achieve with hardware RAID, also supported by some disk
systems
– Incomplete page output can be detected by checksum techniques,
» But extra actions are required for recovery
» Treated as a media failure
 ARIES Data Structures
● ARIES uses several data structures
● Log sequence number (LSN) identifies each log record
4 Must be sequentially increasing
4 Typically an offset from beginning of log file to allow fast access
– Easily extended to handle multiple log files
● Page LSN
● Log records of several different types
● Dirty page table
 ARIES Data Structures: Page LSN
● 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:
4 X-latch the page, and write the log record
4 Update the page
4 Record the LSN of the log record in PageLSN
4 Unlock page
● To flush page to disk, must first S-latch page
4 Thus page state on disk is operation consistent
– Required to support physiological redo
● PageLSN is used during recovery to prevent repeated redo
4 Thus ensuring idempotence
 ARIES Data Structures: Log Record
● Each log record contains LSN of previous log record of the same transaction

LSN TransID PrevLSN RedoInfo UndoInfo

● LSN in log record may be implicit
● Special redo-only log record called compensation log record (CLR) used to log
actions taken during recovery that never need to be undone
● Serves the role of operation-abort log records used in earlier recovery algorithm
● Has a field UndoNextLSN to note next (earlier) record to be undone
4 Records in between would have already been undone
4 Required to avoid repeated undo of already undone actions

LSN TransID UndoNextLSN RedoInfo

1 2 3 4 4' 3'
2' 1'
 ARIES Data Structures: DirtyPage Table
● DirtyPageTable
● List of pages in the buffer that have been updated
● Contains, for each such page
4 PageLSN of the page
4 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
ARIES Data Structures
 ARIES Data Structures: Checkpoint Log
● Checkpoint log record
● Contains:
4 DirtyPageTable and list of active transactions
4 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
● Dirty pages are not written out at checkpoint time
4 Instead, they are flushed out continuously, in the background
● Checkpoint is thus very low overhead
● can be done frequently
 ARIES Recovery Algorithm
ARIES recovery involves three passes
● 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
● Redo pass:
● Repeats history, redoing all actions from RedoLSN
4 RecLSN and PageLSNs are used to avoid redoing actions already
reflected on page
● Undo pass:
● Rolls back all incomplete transactions
4 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
 Aries Recovery: 3 Passes
● Analysis, redo and undo passes
● Analysis determines where redo should start
● Undo has to go back till start of earliest incomplete transaction

Last checkpoint End of Log

Time

Log Analysis pass

Redo pass

Undo pass
 ARIES Recovery: Analysis

Analysis pass
● Starts from last complete checkpoint log record
● Reads DirtyPageTable from log record
● Sets RedoLSN = min of RecLSNs of all pages in DirtyPageTable
4 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
● Scans forward from checkpoint
● .. Cont. on next page …
 ARIES Recovery: Analysis (Cont.)
Analysis pass (cont.)
● 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
4 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
4 May be needed for later undo
● 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
 ARIES Redo Pass
Redo Pass: Repeats history by replaying every action not already reflected in the
page on disk, as follows:
● 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
● 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).
4 CLR for record n noted as n’ in figure below
● Set UndoNextLSN of the CLR to the PrevLSN value of the update log record
4 Arrows indicate UndoNextLSN value
● 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
4 records 3 and 4 initially, later 5 and 6, then full rollback

1 2 3 4 4' 3' 5 6 6' 5' 2' 1'

 ARIES: Undo Pass
Undo pass:
● Performs backward scan on log undoing all transaction in undo-list
● Backward scan optimized by skipping unneeded log records as follows:
4 Next LSN to be undone for each transaction set to LSN of last log record
for transaction found by analysis pass.
4 At each step pick largest of these LSNs to undo, skip back to it and undo it
4 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
undone already
● Undos performed as described earlier
Recovery Actions in ARIES
 Other ARIES Features
● Recovery Independence
● Pages can be recovered independently of others
4 E.g. if some disk pages fail they can be recovered from a backup while other
pages are being used
● Savepoints:
● Transactions can record savepoints and roll back to a savepoint
4 Useful for complex transactions
4 Also used to rollback just enough to release locks on deadlock
 Other ARIES Features (Cont.)
● Fine-grained locking:
● Index concurrency algorithms that permit tuple level locking on indices
can be used
4 These require logical undo, rather than physical undo, as in earlier
recovery algorithm
● Recovery optimizations: For example:
● Dirty page table can be used to prefetch pages during redo
● Out of order redo is possible:
4 redo can be postponed on a page being fetched from disk, and
performed when page is fetched.
4 Meanwhile other log records can continue to be processed
 Remote Backup Systems

● Remote backup systems provide high availability by allowing transaction processing to

continue even if the primary site is destroyed.
 Remote Backup Systems (Cont.)
● 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.
● Heart-beat messages
● 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.
4 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.
 Remote Backup Systems (Cont.)

● 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.
● 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.
● Alternative to remote backup: distributed database with replicated data
● Remote backup is faster and cheaper, but less tolerant to failure
4 more on this in Chapter 19
 Remote Backup Systems (Cont.)

● Ensure durability of updates by delaying transaction commit until update is logged

at backup; avoid this delay by permitting lower degrees of durability.
● 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.
● 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.
● 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.