[zikolach]

Summary

Both single-file and multi-file dumps can create extra indexes when attached partition indexes have custom names. The dump generator creates parent indexes recursively, so PostgreSQL creates or selects child indexes before the exporter creates the named child indexes. Replay can therefore change index names, attachment relationships, and validity even when the SQL executes successfully.

The dump generator now creates partition tables and leaf indexes first, creates standalone partitioned indexes with ON ONLY, and emits explicit ALTER INDEX ... ATTACH PARTITION statements from the inspected relationships. For example, event_lookup retains its attached child early_custom_lookup, while an equivalent index named early_unattached_lookup remains unattached.

• The inspector reads the parent index name and schema from pg_inherits.
• The multi-file formatter keeps deferred parent-index statements in numbered indexes/ files.
• The generator postpones FKs that need those partitioned unique indexes until the indexes and attachments exist.
• Migration generation retains its existing recursive index strategy. Only dump uses explicit index attachments.

This builds on merged #642. It contains no FK-fix commits from that PR.

Nested-partition prerequisite

The nested-index regression exposed an existing table-generation defect: an intermediate partition lost its own PARTITION BY clause. PostgreSQL then rejected its child tables with "archive_events_early" is not partitioned.

A separate commit preserves that clause for intermediate partitions in both dump and migration generation. A native plan/apply fixture covers nested table creation, an intermediate column-default override, and an empty follow-up plan.

Ignored partition-index dependencies

If .pgschemaignore excludes an attached child index or its table while retaining
the parent index, dump returns an error naming both indexes before generating
SQL or writing files. Existing output files remain unchanged. This also applies
to nested partitions and indexes without dependent FKs. Ignored unattached indexes
and consistently excluded index families remain supported.

The inspector records the attachment metadata before filtering. That metadata is
excluded from serialization and fingerprints. Migration generation retains its
existing ignore behavior.

Validation

Passed locally with Go 1.24.7 on macOS:

go test -count=1 -v ./...
PGSCHEMA_POSTGRES_VERSION=14 go test -count=1 -v ./cmd/dump -run '^TestDumpCommand_(IgnoredPartitionIndexes|PartitionIndexIdentity)$'
go vet ./...
go build -o /tmp/pgschema-partition-review .

The full suite passed on PostgreSQL 18, including the ignore tests. The partition-index regression passed all eight combinations of single/multi-file output, --no-comments, and --qualify-schema on PostgreSQL 14 and 18.

The regression compares replayed index and FK catalogs with an independently created source and requires an empty plan against that source. It covers custom names, two partition levels, unattached equivalent indexes, incomplete and invalid parents, an empty partitioned table, unique and partial indexes, expressions, operator classes, included columns, quoted names, comments, and deterministic output.

The ignored-index regression passed nine scenarios in both output modes, with and without existing output files, on PostgreSQL 14 and 18. It verifies actionable errors, no partial output, unchanged existing files, and successful replay for allowed configurations. Unit tests also verify schema-qualified dependency matching, unchanged migration generation, and exclusion of inspection metadata from serialization.

Formatting and whitespace checks also pass. The SQL in the query binding matches the source query. Upstream Linux CI remains pending.

Scope

The explicit attachment strategy applies to standalone indexes in the managed schema. Constraint-backed primary, unique, and exclusion indexes remain under existing constraint management. This does not change migration-time index reconciliation or add cross-schema partition reconstruction support.

AI assistance

AI assisted with implementation, test preparation, and source review. The validation results above are from executed tests.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] Adds partition index attachment tracking to schema dumps.

Fix the ignored-child case before merging, because it can produce a dump that cannot be loaded.

Findings

1.  Ignored indexes break dump replay ▶

Summary

The PR gives dumps a separate path that preserves standalone partition-index names and attachments.

• Database dumps keep partition index names and attachments.
• Intermediate partitions keep their own partition clause.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Inspect selected schema] --> B[Create tables and leaf indexes]
  B --> C[Create nested parent indexes with ON ONLY]
  C --> D[Attach recorded child indexes]
  D --> E[Create root indexes and attach children]
  E --> F[Add deferred foreign keys]
  F --> G[Write one SQL file or ordered includes]

Loading

_{Reviews (1) · Last reviewed commit: "fix(diff): preserve partition keys on ne..." · Reviewed by Greptile}

[Copilot]

🟡 Changes recommended

Attachment discovery misses constraint-backed child indexes, potentially producing invalid parents and failing dependent foreign keys.

1 open finding

• Omit partitioned unique index attachments when child backs a constraint · New

What changed in this PR

Preserves partition-index names and attachment relationships in dumps and complete new-hierarchy migrations.

Changes:

• Inspects index attachment metadata and generates explicit ON ONLY indexes and attachments.
• Preserves nested partition clauses and defers dependent foreign keys.
• Adds dump, migration, ignore, formatting, and documentation coverage.

File	Description
testdata/​diff/​create_table/​nested_partitions/​plan.txt	Adds expected text plan.
testdata/​diff/​create_table/​nested_partitions/​plan.sql	Adds expected plan SQL.
testdata/​diff/​create_table/​nested_partitions/​plan.json	Adds expected JSON plan.
testdata/​diff/​create_table/​nested_partitions/​old.sql	Defines empty initial schema.
testdata/​diff/​create_table/​nested_partitions/​new.sql	Defines nested partitions.
testdata/​diff/​create_table/​nested_partitions/​diff.sql	Verifies nested partition DDL.
ir/​queries/​queries.sql.go	Updates generated index-query binding.
ir/​queries/​queries.sql	Reads index-parent relationships.
ir/​ir.go	Adds partition attachment metadata.
ir/​inspector.go	Populates attachment and hierarchy metadata.
ir/​index_test.go	Tests transient metadata serialization.
internal/​dump/​formatter.go	Adds numbered partition-index files.
internal/​diff/​table.go	Preserves nested partitioning and defers FKs.
internal/​diff/​qualify_schema_test.go	Updates generator calls.
internal/​diff/​partition_index.go	Plans explicit index creation and attachment.
internal/​diff/​partition_index_test.go	Tests ordering, filtering, and eligibility.
internal/​diff/​index.go	Supports CREATE INDEX ... ON ONLY.
internal/​diff/​diff.go	Adds dump-specific generation flow.
docs/​cli/​plan.mdx	Documents new-hierarchy migration behavior.
docs/​cli/​ignore.mdx	Documents attachment dependency errors.
docs/​cli/​dump.mdx	Documents explicit attachment output.
cmd/​plan/​plan.go	Normalizes attachment schema metadata.
cmd/​plan/​plan_test.go	Tests attachment normalization.
cmd/​partition_index_migration_test.go	Adds plan/apply regressions.
cmd/​dump/​partition_index_integration_test.go	Tests dump identity and replay.
cmd/​dump/​partition_index_ignore_test.go	Tests ignored dependency handling.
cmd/​dump/​dump.go	Uses partition-aware dump generation.

Files not reviewed (1)

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

🧠 Review effort: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[zikolach]

@tianzhou I addressed the feedback and tested changes on my use case for which this fix was made. I would appreciate if you can have another look and verify the change. Please let me know if I can help?

[Copilot]

🟡 Changes recommended

Attachment steps are rendered as concurrent index rebuilds in plan output.

1 open finding

• Index attachments incorrectly reported as concurrent rebuilds · New

1 resolved since last review

• Omit partitioned unique index attachments when child backs a constraint

Files not reviewed (1)

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

🧠 Review effort: Balanced

[zikolach]

The attachment-rendering fix is in 83da9e1, and CI passed. I replied to and resolved the corresponding review thread. GitHub denied my request for another Copilot review because my account lacks permission to request reviewers. Could a maintainer request a fresh Copilot review of the latest commit?

[Copilot]

🔵 Needs a closer look

It changes catalog inspection and dependency-sensitive DDL ordering across dump and migration paths, warranting final human review despite extensive coverage.

0 open findings

1 resolved since last review

• Index attachments incorrectly reported as concurrent rebuilds

Files not reviewed (1)

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

🧠 Review effort: Balanced

[tianzhou]

LGTM. Thanks for the contribution

[zikolach]

Following up on this review, I opened #645 because this PR is already merged. Further testing found that PostgreSQL can give an inherited index a name needed by another index, causing dump replay to fail. The follow-up restores the original names before creating the other indexes.