Paper

•  commit protocol: preserving atomicity across multiple sites
• termination protocol: invoked when the occurence of site failures render the continued execution of the commit protocol impossible
• recovery protocols: used by failed sites to resume transaction processing
• global state: state of all local sites and outstanding network messages
• committable: if all sites have voted yes on the transaction
• concurrency set: given state of site i, you can derive the states of other sites
• nonblocking protocols always have more than one committable state?
• central site model
  • there is a single coordinator
  • all other participants are slaves
  • slaves only communicate with coordinator
  • coordinator sends same messages to all slaves, they respond
  • synchronous within one state transition
• decentralized
  • each site is equal
  • every site uses same protocol
  • every sites communicates with every other
• non blocking if
  • there exists no state such that it's concurrency set contains both and abort and commit
  • there existsno noncommittable state whose concurrency set contains a commit
•  

Lecture

• problem: multi-site T: substransaction on each site, sites can reply DONE, user says COMMIT, how does everyone see the same thing?
• Mohan model: xact manager: does 2PC during normal operation, recover process when things go wrong
• recover from: in doubt transactions (after a subordinate crash), termination protocol (COMMIT/ABORT for coordinator)
• classic 2PC:
  • user -> coord: COMMIT
  • coord -> subord: PREPARE
  • subord -> coord: send YES or NO vote after forcing that decision to disk
  • cord -> log: force the decision given votes. This commits the transaction
  • cord -> subord: decision to subord (COMMIT/ABORT)
  • subord -> cord: ACK
  • cord -> disk: end You can forget about the transaction
• failurehandling in basic 2PC
  • coordinator hands T to recover process if a subord fails while in a prepared state
  • subord hands T to recovery process if coord fails after saying PREPARE
  • crash recovery also hands in-flight T to recovery process
  • recovery process periodicly tries to finish
    • ask coord the outcome of Ts
    • notify non-ACK'd subords about the outcome
    • if no info found, you can send abort
• goals for commit protocol:
  • atomicity in with failures
  • "forget" T fairly quickly
  • minimizing logging (esp. forces)
  • maximize opportunity for independent decision making (unilateral aborts)
  • good behavior for read-only T
  • tree of processes
•  hierarchical tree:
  • coor -> sub -> leaf: PREPARE
  • leaf -> sub: READ
  • sub -> coord: YES
  • coord: force write commit
  • coord -> sub: COMMIT
  • sub: force write commit
  • sub -> coord: ACK
  • no info = ABORT
• coordinator logs:
  • commit (forced)
  • end
• subordinates log:
  • prepare with leafs (force)
  • decision after respponses (force)
  • commit (force)
  • end if it has any update leafs
• read only participants:
  • no logging at all
  • not do RO T (including subs and coord)
• aborts are streamlined
  • no-info assumtion
  • abort records are not forced, don't need participant list
  • abort msgs aren't ACKed
  • if there's a failure in abort case, don't need to hand off T
• read paper for PC - because of the problems with read only, not used in practice
•  other topology: centralized, tree, linear,
• 3PC (skeen)
  • coord -> sub: PREPARE
  • sub: force write PREPARE
  • sub -> coord: VOTE YES
  • coord: forcce write ALL_PREPARED
  • coord -> sub: ALL_PREPARED
  • sub: force write ALL_PREPARED
  • sub -> coord: ACK
  • coord: force write COMMIT (commits T)
  • coord -> sub: COMMIT
  • sub: force write commit  (TODO is this needed?)
  • sub -> coord: ACK2
  • coord: end
• termination - pick
  • move to my current state
  • commit/abort (chosen based on my state)
  • if you're in the all-prepared state => you can move into commit
  • else you can abort
  •  
• backup coordinator (subordinate), used in skeen to recover from coord failure
•