Distributed Transaction Management and Consistency

Lecture 25 introduces distributed transaction management, the central mechanism for preserving database consistency across networked processing nodes. A transaction is formally defined as an atomic, logical unit of database processing that includes one or more database access operations (read, computation, and write). Students analyze the formal transaction finite state machine, tracing transitions between active, partially committed, committed, failed, and aborted states. The lecture examines database consistency, distinguishing internal consistency (enforcing structural integrity constraints and schema rules) from external consistency (reflecting real-world domain facts). The complexities introduced by distribution—such as local transactions versus global transactions, network message latency, and independent site execution coordinators—are systematically explored. Distributed deadlock detection and resolution algorithms examine wait-for graphs, probe messages, and priority-based prevention strategies. Timestamp-ordering concurrency control mechanisms illustrate how physical and logical timestamps enforce serial ordering without deadlock vulnerability. Multi-version and conservative timestamp ordering algorithms are analyzed for read-heavy workloads and long-duration distributed transactions. Distributed query processing lifecycles trace query decomposition, algebraic normalization, semantic analysis, and elimination of redundant operations. Data localization algorithms transform global algebraic operator trees into localized fragment queries using relational reduction rules. Heuristic query optimization techniques illustrate how selection and projection operations are pushed down close to local physical fragments. Cost-based query optimization models analyze CPU overhead, disk page access costs, and communication network transmission delays. Distributed join optimization strategies contrast broadcast joins, partitioned hash joins, and semi-join reducers like the SDD-1 algorithm. Distributed reliability architectures examine failure classifications, including transaction aborts, site crashes, and network partitioning anomalies. Atomic commitment protocols, notably Two-Phase Commit (2PC) and Three-Phase Commit (3PC), are demonstrated with formal state transition diagrams. Parallel database system architectures, spanning Shared-Memory, Shared-Disk, and Shared-Nothing topologies, are contrasted for speedup and scaleup performance. Parallel data placement techniques, including round-robin, hash partitioning, and range partitioning, illustrate workload balancing across processing nodes. Parallel join algorithms evaluate parallel nested-loop, associative, and parallel hash join executions in high-throughput query environments. Distributed Object-Oriented Database Management Systems (OODBMS) explore object identity, encapsulation, complex object hierarchies, and pointer swizzling.