05 - Consistency, Transactions and Coordination

<aside> 💡 Generated by BigBrain. Repeated material was merged while distinct trade-offs, caveats, examples, and operational details were retained.

</aside>

05 - Consistency, Transactions, and Coordination

This reference page explores the foundational concepts, mechanisms, and trade-offs involved in consistency models, distributed transactions, coordination, and correctness in distributed systems. It covers consistency guarantees, isolation levels, consensus algorithms, replication techniques, leader election, and operational considerations. The material is organized from fundamental principles to design choices and real-world lessons.

Consistency Models

Consistency models define the guarantees a distributed system provides about the visibility and ordering of updates across replicas. They are critical for reasoning about system behavior under concurrent operations.

Linearizability

Linearizability ensures that operations on a single data object appear instantaneous and respect real-time order. It creates the illusion of a single copy of the data, even in replicated systems. Key properties:

Atomic Execution: Operations are indivisible and not interrupted by others.
Real-Time Order: Operations are ordered based on a global clock.

Implications:

All replicas apply state changes in the same order, often requiring consensus.
Linearizable reads ensure that all clients observe the same state at a given time.

Serializability

Serializability is the strongest isolation level for transactions, ensuring that the outcome of concurrent transactions is equivalent to some serial execution. Unlike linearizability, it does not enforce real-time ordering. Key points:

Prevents anomalies like dirty reads, non-repeatable reads, and phantom reads.
In distributed systems, serializability can lead to inconsistencies if transactions are applied in different orders across replicas.

Strict Serializability

Combining linearizability and serializability, strict serializability (or external consistency) ensures a single global serialization order across all replicas. This is the gold standard for distributed databases, as seen in systems like Google Spanner.