Réplication multimaître (BDR)

Traditional PostgreSQL replication involves one primary (master) and one or more standbys (replicas). Writes only happen on the primary, and reads can be offloaded to replicas.

Multi-master replication (MMR) is different. It allows writes to occur on multiple database servers simultaneously, with all changes synchronized across the cluster.

MMR offers significant advantages over traditional single-master setups:

• High Availability: If one master node fails, others can continue accepting writes, ensuring continuous operation.
• Write Scalability: Distribute write operations across multiple nodes, potentially improving performance for write-heavy applications.
• Geographic Distribution: Place active database nodes closer to users in different regions, reducing latency for local operations.

The biggest challenge in multi-master replication is handling data conflicts.

A conflict arises when the same piece of data (e.g., a specific row) is modified independently on two different master nodes before those changes have replicated to each other. For example:

• User A updates a row on Node 1.
• User B updates the same row on Node 2.

Which change should ultimately prevail?

BDR stands for Bi-Directional Replication. It's an open-source extension for PostgreSQL developed to enable true multi-master replication.

BDR allows multiple PostgreSQL instances to accept writes concurrently while keeping their data synchronized. It builds upon PostgreSQL's native logical replication features but adds crucial capabilities for conflict management.

In a BDR cluster, each participating node acts as both a publisher and a subscriber:

• When a change (INSERT, UPDATE, DELETE) occurs on Node A, it publishes that change.
• Node B (and all other nodes) receives this change as a subscriber and applies it to its local database.
• The same process happens in reverse: Node B publishes its changes, and Node A subscribes to them.

This peer-to-peer communication keeps all nodes synchronized.

BDR employs robust mechanisms to detect when conflicts occur:

• It tracks changes using a combination of row identifiers and version numbers.
• When a replicated change arrives from another node, BDR checks if the local row has been modified independently since the last common state.

Common conflict types include UPDATE vs. UPDATE, UPDATE vs. DELETE, and INSERT vs. INSERT (on unique keys).

Once a conflict is detected, BDR applies a configurable resolution strategy to decide which version of the data prevails:

• Last-Updater-Wins (LUW): The change with the most recent commit timestamp is applied, overwriting older conflicting changes.
• First-Updater-Wins (FUW): The first committed change prevails, discarding subsequent conflicting updates.
• Custom Conflict Handlers: Developers can define their own logic using SQL functions to resolve conflicts based on specific business rules.

Setting up a BDR cluster involves several key steps on each participating PostgreSQL instance:

• Install the BDR extension.
• Configure postgresql.conf for logical replication and BDR-specific settings.
• Initialize the BDR cluster and add each node as a peer.
• Create publications for the tables or schemas you want to replicate.
• Create subscriptions on other nodes to receive changes from the publications.

This ensures all nodes are aware of each other and can exchange data.

BDR is ideal for scenarios requiring active-active databases and high write availability:

• Geographically Distributed Applications: Provide local write access with global data consistency.
• High Availability Clusters: Minimize downtime during individual node failures by allowing writes to continue elsewhere.
• Disaster Recovery: Implement robust disaster recovery plans with multiple active sites.

It's often used in complex, mission-critical environments where downtime and data loss are unacceptable.

While powerful, BDR introduces operational complexity:

• Conflict Management: Requires careful planning of resolution strategies to avoid unexpected data outcomes.
• Monitoring: You need to actively monitor replication lag, conflict rates, and node health.
• Schema Changes: Requires coordinated DDL (Data Definition Language) changes across all nodes to maintain consistency.

BDR is not a drop-in solution and demands a deeper understanding of its architecture and behavior.

In the context of multi-master replication, what is the *primary* challenge that conflict resolution strategies aim to address?

Multi-master replication allows writes on multiple database nodes, significantly boosting availability and write scalability compared to single-master setups.

The main hurdle in MMR is resolving data conflicts, which occur when the same data is independently changed on different nodes.

PostgreSQL's BDR (Bi-Directional Replication) extension provides robust multi-master capabilities, including advanced conflict detection and customizable resolution strategies like Last-Updater-Wins.

While powerful for high availability and distributed systems, BDR requires careful design, monitoring, and management due to its inherent complexity.