Replication changes have different representations
PostgreSQL physical replication reconstructs cluster state from physical changes, including WAL tied to blocks. Logical replication distributes changes interpreted at a logical level, such as changes to table rows. Both send data to another server, but receiver compatibility and selectable scope differ. This comparison focuses on PostgreSQL18 built-in mechanisms rather than treating the terms as identical across every database product. Physical versus logical representation is also separate from whether a commit waits synchronously for a configured remote milestone. Do not infer latency, durability, or freshness solely from the word logical or physical in a deployment diagram.
Physical replication fits a cluster-oriented replica
A physical standby is a candidate when the goal is a closely corresponding cluster-wide replica. The physical meaning of WAL requires compatible server versions and platform or configuration assumptions. It is not a mechanism for freely selecting one arbitrary table and transforming it into an unrelated target schema. A hot standby accepting reads still needs attention to replay conflicts and lag. An asynchronously replicated read need not immediately reflect a write whose success was observed on the primary. Failure standby and read-load distribution use overlapping infrastructure, but they evaluate different correctness and operational conditions even when the same replica serves both purposes.
Logical replication adds selection and mapping decisions
Publications and subscriptions let logical replication select tables and kinds of changes. Supported row filters and column lists can narrow the replicated content further. This does not automatically migrate arbitrary SQL or synchronize every database structure. Prepare appropriate target relations and compatible columns, and check replica identity requirements for locating UPDATE or DELETE targets. A table without a suitable identity can behave differently for inserts than for later updates. State which changes are required during initial configuration rather than treating a successful first copy as proof that the ongoing change stream will support every operation the application performs.
Plan schema and sequence handling separately
PostgreSQL logical replication does not automatically synchronize every DDL operation or sequence state in the same manner as table-row changes. Adding a source column requires a migration order that preserves target compatibility. Before making a subscriber the new writable primary, check sequence positions and objects outside the selected replication set. Matching table row counts does not prove that future inserts and every business operation will work after cutover. Read the restrictions for the product version and list what replication covers, what another procedure must synchronize, and which application operations verify the transition. This avoids discovering an unreplicated dependency only after writes have moved.
Monitor conflicts and retained log volume
Independent writes to a logical subscriber can create constraint failures or apply conflicts. Simply sending changes in two directions does not establish a complete conflict-resolution policy. Where replication slots retain required WAL, a stopped or delayed consumer can create storage pressure on the publisher. Do not assume the primary remains safe indefinitely while a subscriber is unavailable. Observe delivery position, receive and apply progress, errors, and retained volume on a common timeline. For both physical and logical replication, having a network connection is insufficient evidence that data is fresh enough or suitable for the application reads the deployment expects to serve.
Choose by purpose and verify recovery independently
A standby, selective data distribution, a version migration, and an analytics feed can require different scope and compatibility. Define selected tables, permitted write locations, tolerated lag, schema evolution, and cutover criteria before choosing a mechanism. Neither representation automatically substitutes for a backup: an erroneous DELETE can be replicated as faithfully as a correct one. Verification should include updates, deletes, restart catch-up, schema changes, and new writes after cutover, rather than only initial copying. The comparison describes design conditions, not benchmark results from a replication deployment. Recovery remains a separate requirement even when replication is healthy and current.