Add support for extensions in the IR (#436)#443
Conversation
Extensions installed in the database are now read into the IR via a new `Extension` entity and emitted by the diff/dump path so dumps remain replayable on a fresh cluster. Without this, schemas that rely on extension-provided types or operator classes (e.g., a GIST index on UUID via `btree_gist`) couldn't be reproduced from `pgschema dump` output. Implementation: - `ir.Extension` lives at the IR root (cluster-level, not per-schema) - `Inspector.buildExtensions` queries `pg_extension` via a new sqlc query, excluding the always-present `plpgsql` - `internal/diff/extension.go` emits `CREATE EXTENSION IF NOT EXISTS` at the head of the create-phase output and `DROP EXTENSION IF EXISTS` at the tail of the drop-phase output - `WITH SCHEMA` is intentionally omitted from emission. Preserving the user-declared install schema requires more care than this PR takes on (`pg_extension` shows the *actual* install schema, which becomes pgschema's temporary schema during plan generation). Tracked as a follow-up Note on tests: extensions are cluster-level state and persisted across test cases on the shared embedded postgres instance. Without cleanup between cases, any setup.sql that installs an extension would leak into later tests' inspections. `TestPlanAndApply` now registers a `t.Cleanup` that drops non-default extensions when each case exits. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Greptile SummaryThis PR adds extension support to the schema IR and migration output. The main changes are:
Confidence Score: 2/5These issues should be fixed before merging.
Important Files Changed
Reviews (1): Last reviewed commit: "Add support for extensions in the IR (#4..." | Re-trigger Greptile |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| func generateExtensionSQL(ext *ir.Extension) string { | ||
| return fmt.Sprintf("CREATE EXTENSION IF NOT EXISTS %s;", ext.Name) | ||
| } |
There was a problem hiding this comment.
ext.Name is emitted directly into the CREATE EXTENSION statement. Valid PostgreSQL extension names can require quoting, such as uuid-ossp; this renders as CREATE EXTENSION IF NOT EXISTS uuid-ossp;, which PostgreSQL does not parse as that extension name. A dump or plan containing that common extension will fail to apply.
| func generateExtensionSQL(ext *ir.Extension) string { | |
| return fmt.Sprintf("CREATE EXTENSION IF NOT EXISTS %s;", ext.Name) | |
| } | |
| func generateExtensionSQL(ext *ir.Extension) string { | |
| return fmt.Sprintf("CREATE EXTENSION IF NOT EXISTS %s;", ir.QuoteIdentifier(ext.Name)) | |
| } |
Context Used: CLAUDE.md (source)
| Source: ext, | ||
| CanRunInTransaction: true, | ||
| } | ||
| collector.collect(context, fmt.Sprintf("DROP EXTENSION IF EXISTS %s;", ext.Name)) |
There was a problem hiding this comment.
The drop path has the same raw identifier interpolation as the create path. If an installed extension name needs quoting, such as uuid-ossp or a mixed-case custom extension, the generated DROP EXTENSION statement fails or targets the wrong identifier.
| collector.collect(context, fmt.Sprintf("DROP EXTENSION IF EXISTS %s;", ext.Name)) | |
| collector.collect(context, fmt.Sprintf("DROP EXTENSION IF EXISTS %s;", ir.QuoteIdentifier(ext.Name))) |
Context Used: CLAUDE.md (source)
| func generateExtensionSQL(ext *ir.Extension) string { | ||
| return fmt.Sprintf("CREATE EXTENSION IF NOT EXISTS %s;", ext.Name) |
There was a problem hiding this comment.
The inspector records Extension.Schema, but this renderer always emits CREATE EXTENSION without WITH SCHEMA. For a database with CREATE EXTENSION hstore SCHEMA utils and tables using utils.hstore, the replay installs hstore into the current search path instead of utils, so later DDL that references utils.hstore can fail with a missing type or create a different schema layout.
Context Used: CLAUDE.md (source)
| oldExtNames := sortedKeys(oldIR.Extensions) | ||
| for _, name := range oldExtNames { | ||
| if _, exists := newIR.Extensions[name]; !exists { | ||
| diff.droppedExtensions = append(diff.droppedExtensions, oldIR.Extensions[name]) | ||
| } |
There was a problem hiding this comment.
This compares all database extensions and drops any extension missing from the desired IR, even when the command is scoped to one target schema. If the current database has an extension used by another schema or application and the desired SQL for public does not declare it, a schema-scoped plan can emit DROP EXTENSION for unrelated database state.
Context Used: CLAUDE.md (source)
| { | ||
| extNames := sortedKeys(newIR.Extensions) | ||
| for _, name := range extNames { | ||
| newExt := newIR.Extensions[name] | ||
| if _, exists := oldIR.Extensions[name]; !exists { | ||
| diff.addedExtensions = append(diff.addedExtensions, newExt) | ||
| } | ||
| } | ||
| oldExtNames := sortedKeys(oldIR.Extensions) | ||
| for _, name := range oldExtNames { | ||
| if _, exists := newIR.Extensions[name]; !exists { | ||
| diff.droppedExtensions = append(diff.droppedExtensions, oldIR.Extensions[name]) | ||
| } |
There was a problem hiding this comment.
Extension diffing only checks whether the name exists in both IRs. The IR now carries Version, Schema, and Comment, but changes to those fields produce no diff, so pgschema plan can report no changes while the installed extension metadata still differs from the desired state.
Context Used: CLAUDE.md (source)
| // Extensions are cluster-level and persist across tests on the shared | ||
| // embedded postgres instance. Reset any non-default extensions when | ||
| // this test case finishes so the next case (or the next top-level test) | ||
| // inherits a clean cluster. | ||
| t.Cleanup(func() { | ||
| cleanupSharedClusterExtensions(t) | ||
| }) |
There was a problem hiding this comment.
The cleanup is registered only when a fixture has a non-empty setup.sql. The new extension fixture installs btree_gist from new.sql through the shared embedded provider, and ApplySchema drops only the temporary schema, not database-level extensions. Because this fixture has no setup.sql, btree_gist can remain installed in the shared embedded database and contaminate later plan tests.
Context Used: CLAUDE.md (source)
Address Greptile review on pgplex#443: - CREATE/DROP EXTENSION now go through ir.QuoteIdentifier, so names like uuid-ossp render as a properly quoted identifier instead of invalid SQL. - The shared-cluster extension cleanup in runPlanAndApplyTest used to register only when setup.sql was non-empty. Extensions installed via new.sql (the create_extension fixture does exactly this) could leak across cases on the shared embedded postgres. Register the cleanup unconditionally. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1 participant
Summary
ir.Extensionas a cluster-level IR entity (lives at the IR root, not per-schema). Mirrors the per-entity pattern (struct +GetObjectName+ thread-safe accessors).Inspector.buildExtensionsreadspg_extensionvia a new sqlc query and skips the always-presentplpgsqlbuilt-in.CREATE EXTENSION IF NOT EXISTS <name>;at the head of the create-phase output andDROP EXTENSION IF EXISTS <name>;at the tail of the drop-phase output. Dumps now include extensions, sopgschema dump's output is replayable on a fresh cluster (the headline ask of Include Required Extensions in Dump Output #436).DiffTypeExtensionwithMarshalJSON/UnmarshalJSONsymmetry, golden fixture undertestdata/diff/create_extension/add_extension/.Fixes #436
Decisions worth flagging for review
WITH SCHEMAis intentionally omitted from emission.pg_extensionshows the actual installed schema, but during plan generation that's pgschema's temporarypgschema_tmp_*schema, which would leak into golden plans and be non-deterministic. Preserving the user-declared install schema requires either pinning it from the SQL source (currently unavailable to the inspector) or filtering out transient schemas. Tracked as a follow-up.pg_dependfiltering for "only-used" extensions deferred. Include Required Extensions in Dump Output #436 mentions detecting extensions actually referenced by schema objects (viapg_depend deptype='e'). This PR emits all installed extensions exceptplpgsql. The common cases (btree_gist, hstore, pgvector, citext, etc. — explicitly installed by users) are addressed cleanly; selective emission is the natural follow-up PR.ALTER EXTENSION ... UPDATE) not implemented. CurrentlyaddedExtensions/droppedExtensionsare tracked;modifiedExtensionsis out of scope to keep this PR small. The fingerprinting would still detect a version change on the source side; emitting the upgrade DDL is a follow-up.Test infrastructure note
Extensions are cluster-level state and persist across test cases on the shared embedded postgres instance. Without cleanup, a
setup.sqlthat installs (say)btree_gistwould leak into every subsequent test's inspector view. This PR adds at.CleanupinrunPlanAndApplyTestthat drops non-default extensions when each case exits. Touchescmd/migrate_integration_test.go— happy to split that out if preferred.Test plan
testdata/diff/create_extension/add_extension/— declaresCREATE EXTENSION IF NOT EXISTS btree_gist;innew.sql, expects matching DDL emission and round-trip viaTestPlanAndApply.go test ./..., ~8 min including cleanup).Run just the new fixture:
🤖 Generated with Claude Code
Note (2026-05-23): pgproj (a downstream DACFX-shaped fork at https://github.com/guillaume86/pgproj) has adopted a hard-fork posture and will continue independently. This contribution remains open without expectation of merge; happy to address review feedback if a maintainer engages.