feat: add module migration contracts and validation
CI / verify (pull_request) Successful in 21s
CI / infrastructure-smoke (pull_request) Successful in 20s

This commit is contained in:
ekklesia
2026-07-21 08:09:30 -05:00
parent 37f54edd89
commit 88c28d75c9
30 changed files with 2947 additions and 48 deletions
@@ -0,0 +1,151 @@
# 0008: Module-owned migration contracts and offline validation
- Status: Accepted
- Date: 2026-07-21
## Context
ADR 0007 gives each module exclusive ownership of its namespaced Drizzle tables, but it does not
define how those tables evolve. Ekklesia needs a cloneable module to carry its migration history
without a central registry, and Core needs to inspect every installed history deterministically
before a future runner is allowed to touch PostgreSQL.
Drizzle Kit already generates PostgreSQL SQL files, a journal, and schema snapshots. Replacing that
format would separate module authors from supported Drizzle tooling. At the same time, Drizzle's
journal does not express Ekklesia's destructive-change policy or module ownership boundary.
## Decision
### Ownership and packaging
Each module owns its migration history and may create, alter, or remove only database objects that
it owns. A dependency on another module does not grant permission to migrate that module's tables,
indexes, constraints, sequences, types, or other private database objects. Cross-module foreign
keys and imports of another module's private schema or migrations remain prohibited by ADR 0007.
A module with migrations declares the optional manifest field
`entrypoints.migrations`, normally as `./drizzle`. It points to a packaged directory inside the
module, not executable code. A module with no migrations omits the field. Static module discovery
must validate the directory for every installed module before migration files are inspected. The
directory is included in the module package so cloning a module into `modules/`, installing, and
building remains sufficient; no central migration registry is introduced.
### Supported Drizzle format and workflow
This slice pins Drizzle Kit `0.31.10`, matching the stable pre-v1 format compatible with the
repository's Drizzle ORM `0.45.2`. A module configures PostgreSQL, its source schema, an output
directory, index-prefixed migrations, and statement breakpoints. Authors generate migrations with
the module's `migration:generate` script, review the generated SQL and snapshots, add the required
Ekklesia policy entry, and run `pnpm migrations:validate` from the repository root.
The accepted directory format is:
```text
drizzle/
├── 0000_initial.sql
└── meta/
├── 0000_snapshot.json
├── _ekklesia.json
└── _journal.json
```
`_journal.json` and indexed snapshots are the native Drizzle Kit PostgreSQL version-7 metadata.
Every journal entry has a corresponding `<tag>.sql` file and `<index>_snapshot.json`. Extra SQL or
snapshot files and unsupported file types are invalid. Files and the `meta` directory must be real
entries within the declared directory; symbolic links are rejected rather than followed.
Ekklesia adds strict `meta/_ekklesia.json` policy metadata:
```json
{
"version": 1,
"moduleId": "example",
"entries": [
{
"tag": "0000_initial",
"destructive": { "mode": "none" }
}
]
}
```
The sidecar has exactly one entry for each journal entry in the same order. A destructive migration
uses `{"mode":"approval-required","reason":"..."}`. The source tree never records an
`approved` boolean: a future approval record must be external to the module and bound to the
module ID, migration identity, and checksum so editing the migration invalidates approval.
### Identity, checksums, and ordering
Within a module, `_journal.json` is authoritative for ordering. Indices start at zero, are
contiguous, and match each indexed tag; timestamps strictly increase; snapshot IDs form the native
Drizzle predecessor chain. The validator does not repair or re-sort an invalid journal.
Between modules, Core uses the dependency-first order produced by complete static module
discovery, with its deterministic identifier tie-breaking. It then preserves journal order within
each module.
A migration's stable identity is `<module ID>:<Drizzle tag>`, for example
`example:0000_initial`. Core calculates a lowercase SHA-256 checksum from the exact SQL file bytes.
Line endings, whitespace, breakpoint comments, and final newlines therefore affect the checksum.
SQL must be non-empty valid UTF-8, but it is not normalized. The source metadata does not duplicate
the calculated checksum. When execution exists, Core's applied database history is the authority
for comparing the stored applied checksum with current packaged bytes and rejecting edits to an
already-applied migration.
### Destructive changes and SQL limits
Migrations are forward-only for now. There are no downgrade or uninstall migrations. A migration
declared `approval-required` must never be silently applied; offline validation reports that state,
while the future runner will require an explicit external approval contract before execution.
This slice neither grants approval nor executes SQL.
The ownership and destructive declarations are enforceable author contracts and review inputs,
not claims that the offline validator can prove arbitrary SQL safe. SQL comments, quoted and
schema-qualified identifiers, procedural bodies, dynamic SQL, extension behavior, and
`search_path` changes make simplistic regular-expression classification incomplete and
misleading. The validator deliberately treats SQL as opaque bytes and does not claim to detect
cross-module access or destructive statements. Stronger enforcement, if added, requires a real
PostgreSQL-aware model and must state its limits.
### Atomic validation and future execution
Offline migration discovery begins only after complete static module discovery succeeds. It opens
no PostgreSQL connection, imports no schema or server entrypoint, and executes no SQL. It validates
the full inventory, metadata, ordering, identities, paths, snapshots, policy, and checksums for all
modules, then publishes one frozen aggregate. Any issue prevents publication of every module's
result. Errors contain deterministic codes, module IDs, and logical metadata paths only; caught
exception messages, SQL text, URLs, secrets, machine paths, causes, stacks, and partial results are
not exposed.
Migration execution remains exclusively controlled by Core. A later ADR and slice will define the
database history schema, pending/applied comparison, advisory locking, transaction boundaries,
approval lookup, failure recovery, and execution semantics.
The selected format follows Drizzle Kit's stable
[`_journal.json` and snapshot handling](https://github.com/drizzle-team/drizzle-orm/blob/drizzle-kit%400.31.10/drizzle-kit/src/utils.ts),
its [migration generation flow](https://github.com/drizzle-team/drizzle-orm/blob/drizzle-kit%400.31.10/drizzle-kit/src/cli/commands/migrate.ts),
and Drizzle ORM's [SQL reading and SHA-256 behavior](https://github.com/drizzle-team/drizzle-orm/blob/0.45.2/drizzle-orm/src/migrator.ts).
## Alternatives
- A custom migration language was rejected because it would abandon supported Drizzle generation
and metadata without improving runtime safety.
- Expected checksums in packaged migration policy metadata were rejected because they duplicate the
artifact being checked and do not prove whether an edited migration was already applied. Applied
database history will provide that authority.
- Timestamp-only or filename-only identity was rejected because independently installed modules
can legitimately use the same generated tag and timestamp.
- Lexicographically sorting migration files was rejected because Drizzle's journal is the native
ordered history and must be internally consistent.
- Automatically classifying SQL with regular expressions was rejected because it would imply an
ownership and destructive-change guarantee the implementation cannot provide.
- Module-controlled execution was rejected because ordering, history, locking, approval, and
failure behavior are platform-wide invariants owned by Core.
## Consequences
Module authors commit generated SQL, native Drizzle metadata, and the Ekklesia policy sidecar. They
must review ownership and destructive impact explicitly. Modules with no database history remain
valid and need no placeholder directory. Offline validation gives CI a stable inventory and exact
checksums without requiring PostgreSQL, but intentionally does not prove SQL semantics or predict
whether a migration will succeed against a particular live database.