[tianzhou]

Fixes #653

Problem

PARTITION OF copies a parent's inline PRIMARY KEY and UNIQUE constraints onto the new partition under generated names. Since #643 a dump creates partitions before the standalone indexes, so when a standalone index holds such a name (for example events_leaf_pkey), the copy takes it:

• CREATE INDEX IF NOT EXISTS events_leaf_pkey is skipped with a notice and the following ALTER INDEX ... ATTACH PARTITION fails with SQLSTATE 42809.
• When a child's key index was renamed and its original name reused for an index on another column, replay succeeds but silently drops that index.
• pgschema apply of its own dump fails the same way, because the desired state is loaded from the same SQL.

All three were reproduced on main (1.13.3) against PostgreSQL 17.6 with the SQL from the issue.

Fix

Follow pg_dump. A PRIMARY KEY or UNIQUE constraint on a partitioned table that has partitions in the output is no longer written inline. After all tables exist, the dump emits one step per constraint:

ALTER TABLE ONLY events ADD CONSTRAINT events_pkey PRIMARY KEY (id);
ALTER TABLE ONLY events_leaf ADD CONSTRAINT events_leaf_pkey1 PRIMARY KEY (id);
ALTER INDEX events_pkey ATTACH PARTITION events_leaf_pkey1;

ONLY stops PostgreSQL from creating partition copies under generated names; each copy is added under its catalog name and attached explicitly, nested partitions included. Key columns carry NOT NULL explicitly, as pg_dump writes them. Foreign keys referencing the table or any of its partitions wait for these constraints, and the standalone partitioned indexes from #643 follow them (they may attach to constraint-backed children).

• The inspector records the partition copies' names (pg_constraint.conparentid, contype in p/u) in Table.InheritedConstraints. It is inspection metadata, excluded from serialization and fingerprints. If a name is missing, the constraint stays inline as before.
• Migration generation applies the same ordering to complete new hierarchies, which planPartitionIndexCreation already selects for explicit index attachment. Existing hierarchies are unchanged.
• Multi-file dumps write each constraint family under constraints/, like postponed foreign keys. Docs updated.

Compared with #645, this needs no runtime name lookups, no DO blocks, and no changes to the desired-state loader.

Tests

• testdata/diff/create_table/issue_653_inherited_constraint_index_name: plan and apply of a new hierarchy with a colliding standalone index, a nested partition, a partitioned UNIQUE, and foreign keys to the parent and to a partition.
• testdata/dump/issue_653_inherited_constraint_index_name: dump fixture from pg_dump output, including the renamed-key variant.
• TestDumpCommand_PartitionIndexIdentity gains a partitioned table with PRIMARY KEY and UNIQUE whose generated names are taken, nested partitions, a renamed copy, and a referencing foreign key. It replays every output mode and compares the index and foreign-key catalogs with the source.
• Refreshed sakila, issue_409_partitioned_fk and issue_472_partition_clone_trigger dump fixtures (key columns gain NOT NULL, constraints move after the tables).
• Passed locally: go test ./internal/diff ./internal/dump ./internal/plan ./ir/..., go test ./cmd/dump, go test ./cmd on PostgreSQL 18, and the new cases plus Sakila and PartitionIndexIdentity on PostgreSQL 14.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The changes affect partitioned-table dumps and migrations that create complete new hierarchies.

Fix the SQL-function ordering before merging; valid schemas can otherwise fail during replay or apply.

Findings

1.  Functions precede required primary keys ▶

Summary

Preserves inherited primary-key and unique-constraint names when dumping partitioned tables.

• New partition hierarchies keep the names of their key indexes.
• Partitioned table dumps keep key names and add constraints after tables.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Inspect inherited constraint names] --> B[Create tables without postponed keys]
  B --> C[Create SQL functions]
  C --> D[Create named constraint copies]
  D --> E[Attach constraint indexes]
  E --> F[Create standalone partitioned indexes]
  F --> G[Add postponed foreign keys]
  C -. Query needs missing primary key .-> H[Function creation fails]

Loading

_{Reviews (1) · Last reviewed commit: "fix: keep inherited constraint index nam..." · Reviewed by Greptile}

[Copilot]

🟡 Changes recommended

Constraint-dependent SQL routines can be emitted before deferred keys exist, causing dump replay and new-hierarchy application failures.

1 open issue

• Defer constraint-dependent SQL routines until their keys are created · New

1 file not reviewed

• ir/queries/queries.sql.go: Generated file

What changed in this PR

This PR addresses #653 by preserving inherited primary-key and unique-constraint index names in partitioned schema dumps and complete new-hierarchy migrations, extending the explicit index attachment approach from #643.

Changes:

• Records inherited constraint names as inspection-only metadata.
• Creates constraint families explicitly after tables, before standalone partitioned indexes and dependent foreign keys.
• Adds collision and nested-partition regressions, refreshes dump snapshots, and documents the output format.

File	Description
testdata/​dump/​sakila/​pgschema.sql	Refreshes deferred payment keys and explicit nullability.
testdata/​dump/​issue_653_inherited_constraint_index_name/​raw.sql	Defines collision and renamed-key scenarios.
testdata/​dump/​issue_653_inherited_constraint_index_name/​pgschema.sql	Adds expected dump output.
testdata/​dump/​issue_653_inherited_constraint_index_name/​pgdump.sql	Supplies the PostgreSQL reference dump.
testdata/​dump/​issue_653_inherited_constraint_index_name/​manifest.json	Describes the regression fixture.
testdata/​dump/​issue_472_partition_clone_trigger/​pgschema.sql	Refreshes deferred ledger key output.
testdata/​dump/​issue_409_partitioned_fk/​pgschema.sql	Refreshes key and foreign-key ordering.
testdata/​diff/​create_table/​issue_653_inherited_constraint_index_name/​plan.txt	Adds expected readable migration plan.
testdata/​diff/​create_table/​issue_653_inherited_constraint_index_name/​plan.sql	Adds expected executable plan.
testdata/​diff/​create_table/​issue_653_inherited_constraint_index_name/​plan.json	Adds expected structured plan.
testdata/​diff/​create_table/​issue_653_inherited_constraint_index_name/​old.sql	Sets an empty starting schema.
testdata/​diff/​create_table/​issue_653_inherited_constraint_index_name/​new.sql	Defines the target hierarchy and referencing tables.
testdata/​diff/​create_table/​issue_653_inherited_constraint_index_name/​diff.sql	Adds expected migration statements.
ir/​queries/​queries.sql.go	Adds the inherited-constraint query binding.
ir/​queries/​queries.sql	Queries inherited constraint names and parent relationships.
ir/​ir.go	Adds transient inherited-constraint metadata.
ir/​inspector.go	Populates inherited constraint names.
ir/​index_test.go	Checks metadata stays out of serialization.
internal/​diff/​table.go	Omits deferred inline keys and preserves NOT NULL.
internal/​diff/​partition_index.go	Plans and emits named constraint families and attachments.
internal/​diff/​diff.go	Schedules deferred families before partitioned indexes.
docs/​cli/​dump.mdx	Documents explicit constraint creation and output files.
cmd/​dump/​partition_index_integration_test.go	Extends replay coverage for constraint-index collisions.
cmd/​dump/​dump_integration_test.go	Registers the new exact-match dump regression.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Dependency ordering can create partitions before their required functions and policies before the primary keys needed to parse them.

2 open issues

• Preserve parent, function, and partition dependency order · New
• Defer policies that depend on postponed table keys · New

1 resolved since last review

• Defer constraint-dependent SQL routines until their keys are created

1 file not reviewed

• ir/queries/queries.sql.go: Generated file

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Remaining routine-dependency and cross-table policy ordering defects can break replay or apply for valid schemas.

4 open issues

• Schedule key-dependent procedures after constraint creation · New
• Miss transitive dependencies in partition default wrappers · New
• Create parent key before key-dependent partition default functions · New
• Preserve deferred policy ordering across table dependencies · New

2 resolved since last review

• Defer policies that depend on postponed table keys
• Preserve parent, function, and partition dependency order

1 file not reviewed

• ir/queries/queries.sql.go: Generated file

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Unresolved routine and consuming-table dependency ordering can still cause dump replay and new-hierarchy apply failures.

1 open issue

• Schedule wrapper procedures after transitive callees · New

1 previously missed issue

In code that hasn't changed since last review

Defer unrelated default consumer until function dependencies

internal/​diff/​diff.go:3128

In the new default-function-chain fixture, adding a SQL-standard get_payload() with SELECT payload FROM events GROUP BY id LIMIT 1 as the default of an unrelated settings table marks the function as needed and keeps it ahead of events_pkey, causing replay/apply to fail. Unlike a key-dependent default on the partition itself, this dependency chain is acyclic: partition → key → get_payload() → settings. Reschedule the unrelated consuming table and its function after the key, and add a regression for this case.

4 resolved since last review

• Preserve deferred policy ordering across table dependencies
• Create parent key before key-dependent partition default functions
• Miss transitive dependencies in partition default wrappers
• Schedule key-dependent procedures after constraint creation

1 file not reviewed

• ir/queries/queries.sql.go: Generated file

_{Review effort: Balanced}