Replication topologies: where is the serialization point?

Where does each replication topology put the serialization point, and what do quorums really require?

The whole classification follows from one question. A mutable item that accepts concurrent writes needs some place where those writes are put in an order. That place is the serialization point. There are exactly three places to put it.

Single-leader

The serialization point is a designated node. A writer pays one RTT to that specific node: not the nearest, not the median of a group, a specific one. So the write latency inherits that node's full tail. Waiting on one node exposes you to its p99, and there is no order statistic to hide behind. In exchange, ordering is free, conflicts cannot exist, and followers replay a single stream.

The price is a single point of write availability. The system therefore needs failure detection, election and fencing. The entire second half of this mechanism exists to service this one topology.

The compression that matters: single-leader replication converts a distributed-systems problem into a failure-detection problem. While the leader is up, nothing is hard. All the difficulty is concentrated into a rare transition. Multi-leader and leaderless topologies spread the same difficulty over every single operation. You choose whether your hard problem is rare and acute or constant and chronic. This is the idea "costs are relocated, never removed" in its cleanest topological form.

Multi-leader

There is a serialization point at each site. Writers pay only local latency, which is the entire point, and there is no global order, which is the entire cost. Two sites accepting writes to the same key produce two versions with no ordering between them.

Note what this means precisely: the conflicts are manufactured by the topology, not discovered in the workload. The rules for resolving conflicts belong to the study of consistency models. What matters here is that choosing multi-leader means choosing to generate a class of events that did not exist before. A multi-leader deployment survives only if the workload naturally partitions by site (each user's data is written in one region). In other words, it survives only if it is secretly single-leader per key.

Leaderless

The serialization point is established per operation, in the client, by contacting a quorum. The writer waits for W of N acknowledgements and the reader for R, with R + W > N guaranteeing that the read set intersects the write set. This is pure pigeonhole. It uses no clock and no timing assumption, which is why it survives arbitrary delay.

The under-appreciated property: a quorum is the only fan-out that does not amplify the tail. A scatter-gather over N nodes waits for the maximum of N response times, so its p99 is far worse than any single node's p99. A majority quorum waits for the ⌈(N+1)/2⌉-th response, an order statistic near the middle of the distribution. Fan-out is normally a tail hazard; quorums are the case where fan-out truncates the tail instead. The difference is k-of-n versus n-of-n.

Quorums beyond the majority: a correction

The folklore says "you need a majority." What you actually need is that the quorums which must intersect, do. Flexible Paxos showed that only the leader-election quorum and the replication quorum need to intersect, not the replication quorums with each other. So on N = 5 you can run W = 4 with an election quorum of 2 (4 + 2 = 6 > 5), or W = 3 with an election quorum of 3. This turns one dial into two: larger write quorums buy cheaper elections, and the reverse.

Witness replicas use the same idea from the other side. A witness is a member that votes but stores no data. It supplies intersection cardinality without paying R× storage or receiving the write payload. A deployment with 2 data replicas and 1 witness gets majority safety while storing only two copies of the bytes.

Learner (non-voting) replicas are the exact inverse: they hold data and take no part in intersection. That is what a read replica or a cross-region follower actually is.

The three roles, vote, store, serve, are independent. Most systems' "replication factor" collapses them into one number, which then cannot be tuned separately.

Compression: choose a topology by asking where the order is established and how many round trips the writer pays to get there. A fixed node exposes its tail, no node manufactures conflicts, and a quorum truncates the tail but supplies no order.