Distributed Systems · Cluster
Replication
Consistency models and replica placement.
- 01 Replication Topologies and What Each One Tolerates How single-leader, multi-leader, and leaderless replication differ in where writes are accepted, which node losses they survive, and what conflict handling each one forces.
- 02 Synchronous and Asynchronous Replication: Where the Commit Point Sits How the moment a write is acknowledged determines durability, latency, and the amount of data a failover can lose, and why semi-synchronous replication is the usual compromise.
- 03 Leader Leases and Fencing Tokens in Failover Why a demoted leader can still issue writes, how time-based leases bound that window, and why fencing tokens rather than timeouts are what actually protects shared storage.
- 04 Version Vectors and Conflict Detection in Multi-Writer Replication How per-replica counters encode causal history, why vector comparison identifies concurrency that timestamps hide, and what the vectors cost in metadata and pruning.
- 05 Read Repair and Hinted Handoff on the Write Path How leaderless stores bring lagging replicas forward without a full scan, why hinted handoff is a temporary buffer rather than a repair, and where each mechanism fails under sustained failure.
- 06 Quorum Intersection and the Arithmetic of Stale Reads Why R + W > N guarantees a read sees the latest acknowledged write, what that inequality does not guarantee, and how sloppy quorums and read-your-writes routing break it.
- 07 Session Guarantees: Read-Your-Writes and Monotonic Reads How four per-session guarantees are implemented with version tracking, why they are cheaper than global consistency, and where client redirection breaks them.
- 08 Merkle Tree Anti-Entropy: Finding Divergence Without Comparing Everything How hash trees over key ranges reduce replica comparison from a full data exchange to a walk over the subtrees that actually differ, and what write churn does to that saving.
- 09 Failure-Domain Placement: Keeping Replicas Out of the Same Blast Radius How rack, host, and power topology constrain where the third replica goes, why capacity-weighted placement and CRUSH avoid a central allocator, and how correlated loss is measured.
- 10 Bootstrapping a Replica from a Snapshot and a Log Tail Why a running replica cannot simply be copied file by file, how a consistent cut is captured while writes continue, and how the log position inside the snapshot stitches it to the stream.