Replication Implementations

A leader-based replication system needs a mechanism to get data changes from the leader to the followers. The choice of mechanism affects what you can do with the replicated data — version compatibility, cross-engine replication, change data capture, and zero-downtime upgrades all depend on it.

Statement-Based Replication

The leader logs every write statement (INSERT, UPDATE, DELETE) and sends them to each follower. The follower parses and executes each statement as if it received it from a client.

sequenceDiagram
    participant L as Leader
    participant F as Follower

    L->>L: Execute INSERT INTO orders VALUES (...)
    L->>F: Forward statement text
    F->>F: Parse and execute same statement
    Note over F: Result must match leader — but might not

Where it breaks:

Workaround: The leader can replace nondeterministic function calls with fixed return values in the statement log (MySQL did this in early versions).

Verdict: Fragile. Largely abandoned in favor of row-based replication. MySQL used statement-based replication by default before version 5.7.7.

Write-Ahead Log (WAL) Shipping

The leader’s WAL — the sequential log of every byte-level change to data pages — is streamed directly to the follower. The follower replays the same page-level changes to reconstruct an identical on-disk state.

sequenceDiagram
    participant L as Leader
    participant W as WAL (disk)
    participant F as Follower

    L->>W: Write WAL record (page 42, offset 128, old→new bytes)
    L->>F: Stream WAL record
    F->>F: Apply byte-level change to local page 42
    Note over L,F: Physical replication — follower is byte-identical to leader

PostgreSQL streaming replication uses this approach. The standby connects to the primary and continuously receives WAL records:

# postgresql.conf on primary
wal_level = replica
max_wal_senders = 5

# recovery.conf on standby (or standby.signal in PG 12+)
primary_conninfo = 'host=primary-host port=5432'

Limitation: Storage Engine Coupling

WAL records describe which bytes changed in which disk blocks. This ties replication to the exact storage format — the leader and follower must run:

• The same storage engine
• The same database version (or a compatible one)

This makes zero-downtime rolling upgrades difficult: you cannot run the leader on v15 and a follower on v16 if the on-disk page format changed between versions. The standard approach is to upgrade the follower first, promote it, then upgrade the old leader — but this requires the WAL format to be backward-compatible between those two versions.

Logical (Row-Based) Log Replication

Instead of shipping physical byte changes, the leader writes a logical log — a description of what changed at the row level, decoupled from the storage engine format.

Log Content

sequenceDiagram
    participant L as Leader
    participant Log as Logical Log
    participant F as Follower
    participant CDC as External Consumer

    L->>Log: Row-level change record (INSERT, id=7, name='Alice', ...)
    Log->>F: Replicate via logical decoding
    Log->>CDC: Stream to Kafka / data warehouse
    Note over F,CDC: Same log serves replication AND change data capture

MySQL binlog (in ROW format, default since 5.7.7) and PostgreSQL logical replication (via pgoutput plugin) both implement this approach:

-- PostgreSQL: create a logical replication publication
CREATE PUBLICATION my_pub FOR TABLE orders, customers;

-- On the subscriber
CREATE SUBSCRIPTION my_sub
  CONNECTION 'host=primary-host dbname=mydb'
  PUBLICATION my_pub;

-- MySQL: verify binlog format
SHOW VARIABLES LIKE 'binlog_format';  -- should return ROW

Why This Is the Standard Today

Trigger-Based Replication

When replication requirements go beyond what the database engine provides — replicating a subset of data, moving between different database types, or applying custom conflict resolution — the replication logic moves to the application layer.

Mechanisms:

-- Example: trigger-based change capture
CREATE TABLE orders_changelog (
    id SERIAL PRIMARY KEY,
    order_id INT,
    operation VARCHAR(10),
    changed_at TIMESTAMP DEFAULT NOW(),
    new_data JSONB
);

CREATE OR REPLACE FUNCTION capture_order_change() RETURNS TRIGGER AS $$
BEGIN
    INSERT INTO orders_changelog (order_id, operation, new_data)
    VALUES (NEW.id, TG_OP, row_to_json(NEW)::jsonb);
    RETURN NEW;
END;
$$ LANGUAGE plpgsql;

CREATE TRIGGER orders_audit
    AFTER INSERT OR UPDATE ON orders
    FOR EACH ROW EXECUTE FUNCTION capture_order_change();

Trade-off: higher overhead per write (trigger execution + changelog table insert) and more moving parts to maintain. Use this only when built-in replication cannot satisfy the requirement.

Comparison