Distributed systems force a more careful account of truth. Once clocks diverge and partitions become normal, every write carries a question about who can see it, who can confirm it, and how long disagreement is acceptable.
Software / Coordination Surfaces
Architecting Agreement Across Distance
How do we keep a system coherent when messages arrive late, nodes disappear, and no single machine can be assumed sovereign? This surface tracks the tradeoffs between (consensus), [replication], and <eventual agreement>.
Consensus Models
Consensus is the price paid for strong coordination. The more failures and delays you want to survive, the more explicit the protocol has to become about leadership, voting, and recovery.
Leader and Log
Raft keeps the model legible: a leader proposes entries, followers acknowledge them, and committed history becomes the shared source of truth.
Majority as Threshold
Quorums are brace-like thresholds for reality in a cluster. Below the line, a write is still proposed; above it, the system treats that write as instantiated.
Partition Tradeoffs
CAP is less a slogan than a boundary condition: when the network fractures, you choose which promise to weaken and how visibly to weaken it.
Replication and Convergence
Gossip and Anti-Entropy
Not every system needs a leader. Gossip spreads state through repeated local exchange, letting a cluster converge through cadence rather than command.
Vector Clocks and Causality
Once writes happen concurrently, timestamps alone stop being enough. Causal metadata preserves differentiation between “before,” “after,” and “simultaneous.”
CRDTs and Merge Safety
CRDTs trade stronger coordination for mergeable state. They are useful when the system values local action first and can afford delayed global settlement.
#>replica_cluster
#!runtime/distributed
^"write"{
value: "chapter_draft"
quorum: ?[majority]
replicate: ~"followers"
settle: !["commit when acknowledged"]
}