Multi-Primary (Multi-Master) Replication

In standard Primary-Replica systems, all writes must travel to a single geographic Primary node, forcing global users in Australia or Europe to endure high-latency cross-continental network hops (~150ms) for every write mutation.

Multi-Primary (Multi-Master / Active-Active) Replication deploys two or more independent Primary nodes located in different data centers or geographic regions:

• Every Primary node can process both Read and Write transactions locally in <5ms.
• Asynchronously or synchronously propagates writes bi-directionally across the global cluster.
• The Immense Challenge: What happens when two users concurrently update the exact same database row on different Primary nodes at the exact same millisecond?

Consider an account with balance $100:

• User A in New York withdraws 50 on US-Primary (sets balance =50).
• User B in London concurrently withdraws 60 on EU-Primary (sets balance =40).
• Both Primary nodes accept the writes locally and acknowledge success.
• When the asynchronous cross-region replication streams meet across the Atlantic, the database encounters a Write-Write Conflict. Overwriting one balance destroys money; keeping both creates an inconsistent split-brain.

Distributed systems resolve concurrent multi-primary conflicts using four distinct strategies:

1. Last-Write-Wins (LWW / Timestamp-Based)

The database attaches a physical wall-clock timestamp to every write. When a conflict occurs, the write with the highest timestamp overwrites the older write.

• The Danger: NTP Clock Drift. If Server A's clock is 100ms fast, its writes will silently overwrite and destroy legitimate updates from Server B, causing silent data loss.

2. Conflict-Free Replicated Data Types (CRDTs)

CRDTs are specialized mathematical data structures that can be replicated concurrently across multiple independent nodes without central coordination, guaranteed to converge to the exact same state regardless of the order or interleaving of received operations:

• PN-Counters (Positive-Negative Counters): Used for shopping cart quantities, likes, and voting counts.
• LWW-Element-Set / Observed-Remove Set (OR-Set): Used for collaborative real-time documents (Figma, Google Docs, Apple Notes).

3. Vector Clocks & Version Vectors

Attaches a tuple of counters [NodeA: 2, NodeB: 5] to each record. Allows the database to detect whether two updates are causally related (A → B) or concurrent (conflict). Concurrent branches are stored simultaneously (like a Git merge conflict), requiring application code to merge the values.

4. Conflict Avoidance (Geographic Data Partitioning)

Avoid conflicts altogether by ensuring that a specific customer's record is strictly owned by exactly one region:

• US users are pinned to US-Primary; EU users are pinned to EU-Primary. Cross-region writes are prohibited or coordinated via global locks.