Replication Multi-Master (Multi-Primary)

Replication Multi-Master (aussi Multi-Primary, ou leaderless) permet à plusieurs nodes d'accepter writes simultanément, avec synchronisation entre eux. Vs primary-replica où seul primary writes. Augmente write availability et permet active-active topologies cross-region, au prix de complexity (conflict resolution).

Systèmes multi-master :
(1) **Cassandra, ScyllaDB** — leaderless, any node writes, gossip protocol propagation, conflicts résolus via timestamps (last-write-wins) ou CRDTs.
(2) **DynamoDB Global Tables** — multi-region multi-master, eventual consistency, LWW.
(3) **Cosmos DB multi-region writes** — Azure offering, configurable consistency levels.
(4) **MySQL Galera Cluster, Group Replication** — synchronous multi-master, certification-based.
(5) **CockroachDB, YugabyteDB** — distributed SQL avec multi-region writes via Raft.
(6) **PostgreSQL BDR** (Bi-Directional Replication, EnterpriseDB).
(7) **MongoDB replica sets** — technically primary-secondary, mais failover automatique fait qu'any member can be primary.
(8) **Riak, CouchDB** — masterless designs.

Avantages : (1) **Write availability** — writes possible même si certains nodes down ; (2) **Geographic locality** — writes au plus proche node ; (3) **No single point failure** writes ; (4) **Active-active multi-DC** legitimement.

Défis :
(1) **Conflict resolution** — same record updated simultaneously on multiple nodes = conflict. Strategies : LWW (data loss possible), vector clocks (track causality), CRDTs (mathematical merge), application-level resolution.
(2) **Performance** — synchronous multi-master = higher latency (cross-node coordination).
(3) **Operational complexity** — debugging distributed state difficile.
(4) **Split brain** risk if partition.
(5) **Eventual consistency** typically (sauf distributed SQL avec consensus).

When worth complexity : (1) **multi-region writes** required for low latency ; (2) **always-on writes** critical (no maintenance windows); (3) **disaster recovery** active-active ; (4) **scale beyond single-master capacity**.

When avoid : transactional workloads requiring strong consistency, complex queries, small scale where primary-replica suffices. Compétences DP-300, DEA-C01.