Consistency

• Consistency
  • Consistency Models vs Isolation Levels
    • Transaction Isolation
    • Distributed Consistency
  • Two Phase Commits
  • Fault Tolerant 2PC –> Take a look in more depth

Note that consistency may mean different things to different systems

System	Definition
ACID	Satisfying the database’s / application’s constraints
Read After Write Consistency	The ability to view recent changes (read data) directly afterward (write data)
Replication	Read operations return same result / is in same state

Consistency Models vs Isolation Levels

Transaction Isolation

• Avoid race conditions due to concurrently executing transactions
  • Serialization
  • Repeatable read
  • Read commits
  • Read uncommits

Distributed Consistency

• Coordination of the replica’s state, when delays or faults occur
  • Strict
  • Linearizability
  • Causal
  • Monotonic reads/writes
  • Read-your-writes

Two Phase Commits

• Client sends begin transaction
• Transaction executed as normally
• Command to commit is sent to coordinator
  • Coordinator asks nodes to prepare for commit
  • Only when all have given okay is the transaction commited
• Typically linearizable

What happens in the event of a crash?

• Coordinator decides when to write to disk
• Upon recovery, send decision to replicas
• Problem if crash occurs between prepare and commit (or revert)
• Algorithm is blocked until coordinator recovers (single point of failure)

Fault Tolerant 2PC --> Take a look in more depth

• Every node that participates in the transaction uses a total order broadcast to vote (commit or abort).
• If a node A suspects a Node B has crashed, it will vote abort for Node B
• Potential race condition if two conflicting votes are received from same Node
  • Count first vote to arrive

---