Hybrid Replication Protocols for Trading Constraint Consistency against Availability

1
Hybrid Replication Protocols for Trading
Constraint Consistency against Availability

• Johannes Osrael

2
Content

• Motivation
• Contribution
• Related Work
• System Characteristics
• Replication in DeDiSys
• Future Work

3
Motivation

• DeDiSys objective Enhancing availability of
  data-centric, distributed systems by relaxing
  data integrity (constraint consistency)
• Replication is used as a means to achieve
  fault-tolerance
• A plethora of replication protocols exist but
  they do not support the trade-off constraint
  consistency / availability

4
My Contribution

• New replication protocols that support the
  trade-off
• constraint consistency / availability in
  object-based,
• data-centric distributed systems

5
Consistency Types

• Replica consistency? defines correctness of
  replicas
• Concurrency consistency ? criterion for
  concurrent access to a particular data item
  (isolation)
• Constraint consistency ? defines correctness of
  the system state with respect to a set of data
  integrity rules

6
Related Work

• Replication protocols in general
• Replication protocols that support
• Trade-off replica consistency/availability
• Trade-off concurrency consistency/availability
• Other trade-offs e.g., performance/availability

7
Related Work TACT

• H. Yu, A. Vahdat Design and Evaluation of a
  Conit-Based Continuous Consistency Model for
  Replicated Services.
• TACT Tunable Availability/Consistency Tradeoffs
  
• Contribution continuous consistency model
• Replication system based on consistency units
  (conits)
• Conit logical consistency unit
• Applications impose consistency requirements on
  conits
• Consistency level of each conit is defined using
  a simple set of metrics

8
Related Work TACT (cont.)

• Numerical error limits difference (nr. of
  writes) between observed value in a replica and
  the ideal value
• Order error limits amount of out-of-order
  updates on a given replica
• Staleness real-time bound on the delay of write
  propagation among replicas

9
Related Work TACT (cont.)

• Example Replicated bulletin board service
• (one conit covers all news messages)
• Numerical error bounds total number of messages
  posted system-wide but not seen locally
• Order error limits number of out of order
  messages on a given replica
• Staleness limits delay of messages

10
Related Work TACT (cont.)

• Continuous scale of consistency values
• Replication system enforces that specified limits
  are not exceeded
• Similarity to DeDiSys fine-grained trade-off
  consistency/availability
• However, we aim at relaxing constraint
  consistency rather than replica consistency

11
Related Work Update Window Approach

• C. Zhang, Z. Zhang Trading Replication
  Consistency for Performance and Availability An
  Adaptive Approach.
• Updates are locally applied and buffered till
  update window is full
• Update window size of queue of locally applied
  updates
• Updates are pushed to other replicas when queue
  is full
• Local replica is blocked until updates are
  applied at all replicas

12
Related Work Update Window Approach (cont.)

• Small size strong consistency, reduced
  availability
• Large size loose consistency, enhanced
  availability
• Contribution algorithm that adapts size of
  update windows to optimize availability
• Similarity to DeDiSys fine-grained trade-off
  between consistency and availability
• However, replica consistency is traded and not
  constraint consistency

13
System Characteristics

• Data-centric, object-based distributed systems
• Node and link failures, partitions
• Pause-crash behaviour
• LAN WAN, but leased lines with guaranteed
  bandwidth (about 2MBit)
• Number of nodes max. 30

14
System Characteristics (cont.)

• Typical object size 5-12 kB
• Large number of objects (10000 1 million)
• Up to 20 objects accessed per second
• Write/read ratio 1/10
• Nested invocations
• Up to 10 of object invocations are remote

15
Replication

• Distributed systems (objects) vs. databases (data
  items)
• Different replication models for both worlds
• Mapping not always possible
• Synchronous vs. asynchronous behaviour
• State transfer vs. operation transfer

16
Replication Models for Distributed Systems

• Active replication

• Passive Replication

Intermediate models coordinator-cohort,
semi-active, ...
17
Passive Replication Synchronous vs. Asynchronous

• Synchronous primary replies to client after all
  backup replicas have been updated (blocking)
• Asynchronous primary replies to client
  immediately after request processing and
  propagates updates afterwards (non-blocking)

18
Operation Transfer vs. State Transfer

• Updates can be propagated using either operation
  transfer or state transfer
• State transfer if update message contains all
  information to update replica unconditionally
• Operation transfer if receiving replica has to
  be read to perform update

19
Operation Transfer vs. State Transfer
20
Operation vs. State Transfer in DS

• Employee e new Employee()
• State transfer e.setSalary(2500)
• Operation transfer e.setSalary(e.getSalary()100)
  

21
DeDiSys Approach

• Object replication
• Passive model
• Either operation transfer or state transfer
• (criterion state transfer time vs. time to
  process operation)
• Healthy periods synchronous replication
• Degraded periods asynchronous replication
• Reconciliation after rejoining of partitions

22
Protocol Phases
Phase 1 healthy system
Phase 2 degradation
Phase 3 reconciliation
23
Future Work

• Design of new protocols
• Simulations
• Proof of concept implementation (EJB)
• Assessment (comparison of protocols, technology
  recommendations)
• Publication (target conferences DOA, SRDS, DSN,
  Middleware, )
• Writing the thesis