PostgreSQL 14 → 18: an upgrade decision record
An illustrative readiness report based on a reproducible self-managed lab. Generated data only; this is not a customer engagement or an RDS/Aurora rehearsal.
An illustrative readiness report based on a reproducible self-managed lab. Generated data only; this is not a customer engagement or an RDS/Aurora rehearsal.
Synthetic example. The recorded observations below come from the published lab script and output run on 21 October 2026. No client data or client outcome is represented.
| Item | Lab evidence | Decision use |
|---|---|---|
| Engine | PostgreSQL 14.24 → 18.6, self-managed Docker/PGDG packages | Path applies to this lab only |
| Data | Generated orders table, 500,000 rows | Representative of this test, not a client workload |
| Extensions | Upgrade check logged “Checking for extension updates … ok”; a complete extension version matrix was not captured | Real review requires pg_extension inventory and target availability |
| Replication | pgoutput logical slot debezium_orders and publication cdc_pub | CDC consumer restart and reconciliation need an owner |
| Checksums | PG14 cluster off; PG18 default initialization on | Cluster creation settings must match for pg_upgrade |
| Check | Observed | Action before go |
|---|---|---|
| Data checksums | Initial check failed with mismatched settings | Initialize target consistently, or choose and test another migration path |
| Prepared transaction | Check failed until open transaction was rolled back | Resolve with application owner; prevent recreation during cutover |
regproc column | Check failed until the lab column was converted to text | Review application semantics and schema migration |
| Logical slot / CDC | Check passed with the slot present; slot was absent after upgrade | Plan CDC position, recreation and downstream reconciliation |
| Planner statistics | Lab ran staged analyze after upgrade | Include statistics rebuild and workload comparison in runbook |
The lab records a 500,000-row orders table but does not record pre/post query latency or plans. Therefore it cannot establish performance readiness. A client report would include agreed query fingerprints, plans, p95/p99 latency, call volume, waits and error rate from comparable windows. The go decision stays pending until that evidence exists.
pg_upgrade --link has different recovery constraints from a copy-based path. Owner: operations. Validation: rehearse restore/switchover and set a cutover decision time.pg_upgrade --check passes on the final rehearsal copy.Illustrative decision: no-go until the unmeasured workload and CDC recovery criteria are documented and tested. Passing the mechanical upgrade check alone is insufficient.
See the full upgrade checklist and review scope.