Chapter 4. Consistency and Replication

1
Chapter 4. Consistency and Replication
2
Replication and Consistency

• Reasons for Replication
• Reliability
• Performance e.g. scalability
• Drawbacks
• Performance and Scalability to maintain
  consistency
• When updates take place more often than simple
  accesses
• Update propagation to maintain consistency

Replica
Replica
Update
Replica
3
Example Shared Objects

• Handling concurrent accesses of a shared object

4
Example Replicated Object
Distributed system is responsible for replica
management
Replication-Aware Distributed Objects
5
Approach

• Goal
• Scalability by replication
• Without loss of consistency
• Problem
• Updates take long time to propagate them to
  replicas
• Approach A compromise
• Loosen consistency constraint
• Copies are not same at all time
• Classification
• Data Centric Approach Global Consistency
• Client Centric Approach Consistency to each
  client

6
Data-Centric Approach vs. Client-Centric
Approach
7
Data-Centric Approach Strict Consistency

• Strict Consistency
• Any read on a data item x returns a value
  corresponding to the result of the most recent
  write on x.
• But it is impossible to realize strict
  consistency. Why ?
• History order of executions (interleaved or not)

time
Strict Consistent
Not Strict Consistent
8
Data-Centric Approach Sequential Consistency

• No matter what operations really take place
• cf. Strict Consistency
• Orders seen by different machine should be same
• Weaker than Strict Consistency
• What problems in realizing sequential consistency
  ?
• Comparison with serializability

Sequentially Consistent
Sequentially Inconsistent
9
Sequential Consistency Example

• 6!(720) orders wrt execution orders
• 64 results wrt print output
• 000000 not allowed, since no respect on
  execution order
• 001001 not allowed, since no respect on
  sequential consistency

10
Data-Centric Approach Causal Consistency

• Distinction between
• Concurrent operations not causally related
• Causally related operations
• Example. P1 W(x), P2 R(x), W(y)
• W(y) in P2 is causally related with R(x) in P2
• Causally Related Operations Should respect
  order
• Weaker than Sequential Consistency

Causally Consistent
11
Data-Centric Approach FIFO Consistency

• Two types of writing orders
• Writing order in a machine
• Writing order in several machine
• FIFO Consistency
• Only respect writing order in a machine
• Weaker than Causal Consistency

12
Data-Centric Approach Weak Consistency

• No consistency except Explicit Synchronization
  Request
• Synchronize Operation
• All writes on a machine should be propagated to
  all copies
• No operation on a synchronized variable is
  allowed until all previous writes have completed
  everywhere.
• No read or write operations are allowed until
  all previous operation to synchronization
  variables have been completed

Weak Consistency
Not Weak Consistency
13
Release Consistency and Entry Consistency

• Release Consistency
• Explicit Specification on Critical Section
• Acquire and Release
• Like Enter_CS and Leave_CS
• Programmer should specify
• Entry Consistency
• Like Release Consistency
• But explicit specification for each data item

14
Summary of Data-Centric Consistency
15
Eventual Consistency and Client-Centric
Consistency

• Rare Updates
• For a certain period of time, inconsistency can
  be tolerated.
• Client-Centric Consistency
• Lazy update propagation

16
Client-Centric Consistency

• Client-Centric Consistency
• Not all copies must be consistent
• Consistency from the viewpoint of client (or
  process)
• Lazy update propagation
• Example Web pages
• Four types of Consistency
• Monotonic Read Consistency
• Monotonic Write Consistency
• Read-your-writes Consistency
• Writes-follow-reads Consistency

17
Monotonic Read Consistency

• Guarantee
• If a read by a process returns a value, any
  other reads by this process at a later time
  should notreturn an old value
• Suppose a mobile client, the order of reads by
  this clientshould be respected the order
  wherever it is.
• Update Propagation for Read Order
• Mobile Client has consistent Read

Monotonic-Read Consistent
Not Monotonic-Read Consistent
18
Monotonic Write Consistency

• Guarantee
• Order of writes by a same process wherever they
  are performed
• Suppose a mobile client, the order of writes by
  this clientshould be respected the order
  wherever it is.
• Update Propagation for Write-Order
• Mobile Client has consistent Writes

Monotonic-Write Consistent
Not Monotonic-Write Consistent
19
Implementation Issues Distribution Protocol

• Replication
• Deciding Where, When, and by Whom to duplicate
• Update Propagation
• A Conceptual Classification

20
Server-Initiated Replication

• Server-Initiated Replicas
• Replicas made by server to enhance performance
• How to decide when and where ?
• Keep Access Count
• Del(S,F) Threshold for deleting a replica
• Rep(S,F) Threshold for making a replica

21
Client-Initiated Replication

• Client caches
• on the client machine but not necessarily
• on a machine in its LAN
• Good when no update happens
• For a short time
• Cache Replacement
• Hit Ratio

22
Update Propagation

• State vs. Operation What to propagate ?
• Notification of update
• Transfer new data
• Update Operation
• Pull vs. Push Protocol
• Push Protocol (Server-based Protocol)
• Updates are propagated to other copies WITHOUT
  requests
• Often between Permanent and Server-initiated
  Copies
• When High Degree of Consistency is required
• Pull Protocol (Client-based Protocol)
• Update propagation when requested
• Often client cache
• When Read-to-update ratio is low

23
Comparison of Push and Pull Protocol

• Comparison
• Uni-casting vs. Multicasting
• Multicasting is convenient for push protocol