Publikasi, Langganan, dan Identitas Replika

Physical (streaming) replication ships the entire WAL byte-for-byte to identical replicas. Logical replication instead decodes the WAL into row-level change events (INSERT/UPDATE/DELETE) and streams only the tables you choose.

• Selective: replicate a hot subset of tables, not the whole cluster.
• Cross-version & cross-schema: publisher and subscriber can differ in major version and have extra columns or indexes.
• Performance use cases: offload reporting queries to a read replica, build a slimmer OLAP copy, or shard write traffic.

The two building blocks are a PUBLICATION on the source and a SUBSCRIPTION on the target.

Logical replication requires the WAL to carry enough information to reconstruct rows. Set wal_level = logical on the publisher (a server restart is required).

• max_wal_senders must allow one slot per subscription plus headroom.
• max_replication_slots bounds the number of logical slots.

Each subscription consumes one replication slot, which pins WAL on the publisher until the subscriber confirms it. An inactive subscriber can therefore cause unbounded WAL growth.

A publication is a named set of changes from one or more tables. You decide which tables and which operations are published.

• FOR TABLE lists specific tables.
• FOR ALL TABLES publishes every current and future table (superuser only).
• publish controls which DML types stream: insert, update, delete, truncate.

Publishing only the operations you need reduces WAL decoding work and network traffic.

PostgreSQL 15 added row filters and column lists to publications, letting you replicate only the slice of data the subscriber needs. This shrinks the replicated dataset and the decoding cost.

• A WHERE clause filters rows; it may only reference replicated columns.
• A column list replicates a subset of columns and must include the replica identity columns.

This is ideal for building a lean reporting copy that excludes archived rows or sensitive columns.

On the target server, a SUBSCRIPTION connects to the publisher, creates a replication slot, and starts applying changes. The target tables must already exist with compatible columns.

• copy_data = true (default) snapshots existing rows first, then streams live changes.
• Each subscription spawns an apply worker on the subscriber.

The connection string points back to the publisher with a role that has REPLICATION (or is superuser) and can read the published tables.

An INSERT carries the full new row, so the subscriber can simply insert it. But an UPDATE or DELETE must tell the subscriber which existing row to change. That requires the WAL to contain an identifying "old" image of the row.

This identifying image is the replica identity. Without it, the publisher cannot encode the old key values, and the change either fails to apply or is silently dropped.

• INSERT: needs no replica identity.
• UPDATE / DELETE: requires a usable replica identity on the published table.

Every table has a REPLICA IDENTITY setting that controls what old-row data is written to the WAL for UPDATE and DELETE:

• DEFAULT: logs the columns of the primary key. The usual, efficient choice.
• USING INDEX: logs the columns of a chosen unique, non-partial, NOT NULL index.
• FULL: logs the entire old row; the subscriber matches on all columns. Correct but expensive.
• NOTHING: logs no old image; UPDATE/DELETE on the table then fail to replicate.

With the DEFAULT replica identity but no primary key, a table effectively behaves like NOTHING: UPDATE and DELETE cannot be replicated. PostgreSQL raises an error such as "cannot update table because it does not have a replica identity and publishes updates".

You have three fixes, in order of preference:

• Add a primary key (best for performance).
• Add a unique, NOT NULL index and set REPLICA IDENTITY USING INDEX.
• Set REPLICA IDENTITY FULL as a last resort.

REPLICA IDENTITY FULL writes every column of the old row into the WAL on each UPDATE and DELETE. On wide or high-churn tables this inflates WAL volume and slows the publisher.

It also hurts the subscriber: to find the matching row it must compare all columns. Before PostgreSQL 16 this meant a sequential scan per change; PG 16+ can use a usable index on the subscriber, but full-row matching is still costlier than a key lookup.

Use FULL only when no unique key exists and you cannot add one.

The replica identity is stored in pg_class.relreplident: d = default (primary key), i = using index, f = full, n = nothing. Audit your published tables before going live so no UPDATE/DELETE silently fails to replicate.

• Tables published with update or delete must not have relreplident = 'n' unless they also lack a usable key.
• psql's \d+ tablename also prints the replica identity.

For a performance architecture you must watch replication health. The publisher exposes apply progress via pg_stat_replication, and slot retention via pg_replication_slots.

• pg_replication_slots.active false plus growing restart_lsn distance means WAL is piling up because a subscriber is stuck.
• On the subscriber, pg_stat_subscription.latest_end_lsn versus the publisher's current LSN shows apply lag.

Set max_slot_wal_keep_size so a dead subscriber cannot fill the publisher's disk.

You configured selective logical replication and the replica identity that makes UPDATE/DELETE streaming correct.

• Publisher: wal_level = logical, then CREATE PUBLICATION with chosen tables, operations, and (PG 15+) row/column filters.
• Subscriber: CREATE SUBSCRIPTION spawns an apply worker and a replication slot; copy_data seeds existing rows.
• Replica identity: INSERT needs none; UPDATE/DELETE need DEFAULT (primary key), USING INDEX (unique NOT NULL index), or FULL. NOTHING and PK-less DEFAULT break UPDATE/DELETE.
• Performance: prefer a key-based identity over FULL; monitor slots and lag, and cap retained WAL with max_slot_wal_keep_size.