LASS Reading Group Puru Kulkarni, UMass

1
LASS Reading GroupPuru Kulkarni, UMass

• Epidemic Algorithms for Replicated Database
  Maintenance
• A. Demers et.al
• Managing Update Conflicts in Bayou,
• a Weakly Connected Replicated Storage System
• D. Terry et.al
• Xerox Palo Alto Research Centre

2
Introduction

• replicated databases
• mutual consistency on updates to data
• consistency model
• algorithms analogous to epidemics
• eventual consistency
• randomized (server makes independent random
  choices)
• robust and scalable algorithms

3
Algorithms

• Direct Mail
• Anti-Entropy
• Rumor mongering
• Factors to compare
• Time for update propagation
• Network traffic generated for each update
• reliability

4
Direct Mail

• for each s in S do
• Postmailto s, msg("Update", x.valueof()
• endloop
• (traffic number of sites average distance
  between
• sites)
• On update send mail to all replicas
• Depends on reliability of underlying service
• Server needs to be aware of all sites

5
Epidemic Algorithms

• Types of sites Infective, Susceptible, Removed
• Anti-Entropy algorithm (executed at replica s)
• for some s in S do
• Resolve_Difference (s, s)
• endloop
• Resolve_Difference
• Push
• Pull
• Push-Pull

6
Anti-Entropy

• Properties
• site s chosen at random
• Periodic execution of anti-entropy
• compare entire dataset (expensive)
• some susceptible sites possible
• Always infective or susceptible
• Improvements
• database checksum
• bounded recent-updates list
• inverted index of database by timestamp

7
Complex Epidemics

• Maintain list of infective updates
• Replica tries to insert update and change its
  list
• When to remove update from list?
• Rumor mongering
• plant rumor (infect)
• Infected sites spread rumor randomly
• If recipient knows of rumor, with probability 1/k
  site goes to removed state
• probability of failure depends on k

8
Rumor mongering

• Criteria to judge epidemic
• Residue (sir 1)
• Traffic (ratio of total update traffic and
  number of sites)
• Delay (tarr, tlast)
• Variations of Rumor mongering
• Blind v/s Feedback
• Counter v/s Coin
• Push v/s Pull
• Connection limit, Hunting

9
Rumor mongering

• Less network traffic
• Non-zero probability of site not getting an
  update (depends on k)
• Anti-entropy for backup (infrequently)
• Direct mail for backup (traffic overhead)
• Rumor cycles more frequent than anti-entropy
• Less resources for update transfer

10
Epidemic algorithms

• Deletion of item
• Remove from database
• Resurrection problem
• Insert death certificate in replicas
• Spread using epidemic algorithms
• When to delete death certificates
• Tradeoff (risk of resurrection v/s space)
• Bounded time interval
• Distributed algorithm

11
Epidemic algorithms

• Dormant Death Certificates
• Hold death certificate till t1
• Create dormant after t1 till t2 (on some
  sites)
• Reactivate death certificate if obsolete update
  encounters dormant certificate
• If not, remove dormant after t2
• Reduces network traffic
• k and t2 control probability of failure
• Spatial Distribution
• Network not uniform, Favor neighbors?
• Effect of network topology

12
BAYOU

• Features
• Replicated data (relational database)
• Read/write anywhere model
• weakly consistent data (weak connectivity)
• Conflict detection/resolution (application-aware)
• Eventual consistency
• Example applications
• Meeting room scheduler
• Bibliography database
• Mobile email, Bayou project co-coordinator

13
Bayou

• System model
• Client Application Bayou API
• Read/write at any available server
• Log of writes and resulting data
• Write performed locally (execution, conflict
  detection/resolution)
• Effects of write immediately available
  (tentative/committed writes)
• System support for conflict detection
• Application specific conflict resolution

14
Bayou

• Update management
• Pairwise anti-entropy sessions
• Exchange of write-logs determine order of writes
  and conflicts
• Eventual consistency (theory of epidemics)
• Factors for write-convergence
• Network connectivity
• Frequency of anti-entropy
• Server policies (site selection)

15
Bayou

• Conflicts
• Application specific
• Meeting scheduler (timings overlap)
• File update (record level)
• Dependency checks
• How to detect conflicts?
• Application specific query (similar to SQL) and
  expected result
• Merge Procedure
• Steps to resolve conflict

16
Bayou

• Replica (eventual) Consistency
• Writes performed in the same, well-defined order
  at all servers (global-ordering)
• Conflict detection and merge procedures
  deterministic
• Tentative/committed writes
• Tentative writes ordered by timestamps of
  accepting server
• Support of undoing writes (use of write logs)
• No time-varying data used by merge procedures

17
Bayou

• Replica (eventual) Consistency
• Stable write, all the preceding writes are fixed
  in the log (do not need re-execution)
• Client can decide whether to access
  tentative/stable data
• How to determine stable write?
• Send accept-times for data on anti-entropy
• If server has time lower than all, then commit
• Partitions cause delays or commit problems
• Sending timestamps is overhead

18
Bayou

• Replica (eventual) Consistency
• Primary commit scheme
• Single server designated to commit writes
• Commits propagate during anti-entropy
• Different data items can have different primary
  servers
• Disconnected operation supported, quorum based or
  distributed algorithms for commit not a good idea
• Primary can be located near locus of activity, to
  increase rate of write commits

19
Bayou

• Replica (eventual) Consistency
• Non-transparent, application specific conflict
  detection/resolution
• Pair-wise anti-entropy
• Ordering of writes (tentative and global)
• Primary commit scheme

20
Bayou

• Implementation issues
• Space efficient write logging
• Efficient undo/redo
• Views of tentative/committed data
• Anti-entropy support
• Resurrection of discarded committed writes
• Authentication for access control

21
Bayou

• Conflict management conclusions
• Non-transparent
• Application specific resolution
• Per-write conflict re-solvers
• Partial and multi-object updates
• Tentative/stable resolutions
• Security
• Additional issues
• Rate of anti-entropy, server selection, partial
  replication, alternate data models