Trade-off Summary: Raft trades some throughput for understandability — it’s designed to be genuinely comprehensible by practitioners, not just theorists. It excels in primary-backup replication scenarios where a clear leader simplifies coordination. However, it inherits Raft’s leader bottleneck and doesn’t scale writes horizontally. For anything beyond moderate write throughput, consider layered approaches or different consensus algorithms.

Trade-off Summary: Two-Phase Commit provides atomic commitment across distributed participants, but it blocks during the commit phase and suffers from the coordinator crash problem — if the coordinator fails after participants have prepared but before commit, those participants are stuck indefinitely. 2PC is only suitable for short, fast transactions with highly available participants. For anything complex, use the Saga pattern with compensations.

Trade-off Summary: CRDTs enable coordination-free synchronization — any two replicas can merge independently without coordination, making them ideal for partition-prone, multi-leader scenarios. However, they impose constraints on data types, can grow unboundedly without compaction, and require careful selection of CRDT type for each piece of application state. They’re not a replacement for all data structures — they’re a specific tool for specific consistency problems.

Trade-off Summary: Circuit breakers prevent cascading failures and enable graceful degradation by failing fast when downstream services are unhealthy. However, they require careful threshold tuning (failure count, timeout window, half-open probe size) and introduce additional state machine complexity. They work best for protecting against genuinely unreliable dependencies — for stable internal services, a well-tuned timeout is often sufficient.

Trade-off Summary: Gossip protocols scale gracefully in large, dynamic systems by leveraging randomness to disseminate information in O(log n) rounds rather than O(n) broadcasts. They’re the foundation of Cassandra, Consul, and Dynamo-style systems, but they provide only probabilistic guarantees about state convergence. Use them when scale and partition tolerance matter more than immediacy — not when millisecond-level consistency is required.

Trade-off Summary: Vector clocks capture causality across distributed processes, enabling precise conflict detection and “happened-before” reasoning. They’re more powerful than logical timestamps for distributed debugging and causally consistent databases, but they consume O(n) space per key where n is the number of nodes. Many systems use hybrid logical clocks or version vectors as a space-efficient compromise that captures enough causality for practical use.

Trade-off Summary: Exactly-once delivery is a lie — the Two Generals and FLP impossibility results prove you can’t have it over unreliable networks. What systems actually implement is “effectively-once”: at-least-once delivery with idempotent consumers that deduplicate at the destination. The real cost is not in the broker but in your consumer logic — every message must carry a unique ID, and your processor must track which IDs it has successfully processed. Only pursue exactly-once when duplicate messages have unacceptable consequences.