Pivot: fold drizzlebox as utils, HonkerEventTarget, OperationSpecs as repo surface

- Update architecture docs to reflect pivot from @libsql/client to Honker
- Fold @alkdev/drizzlebox Phase 0 into src/sqlite/utils/ (ADR-046)
- Add HonkerEventTarget adapter for pubsub TypedEventTarget (ADR-047)
- Replace hand-written CRUD with OperationSpec generation (ADR-048)
- Resolved OQ-26: Honker replaces Redis for single-node pub/sub (POC validated)
- Updated OQ-17, OQ-18, OQ-19 for OperationSpec repository surface
- Added OQ-30 (composite event target), OQ-31 (consumer naming), OQ-32 (Drizzle Kit)
- POC results: adapter buildable, same-process pub/sub works, transactional
  outbox semantics confirmed, concurrent listeners/streams work
- Research doc at docs/research/pivot-honker-sqlite-adapter.md
This commit is contained in:
2026-06-01 16:31:40 +00:00
parent 6aa2fcc6ff
commit 412ad98f11
10 changed files with 1342 additions and 230 deletions

View File

@@ -0,0 +1,53 @@
---
status: accepted
date: 2026-06-01
supersedes: ADR-005, ADR-018
---
# ADR-046: Fold @alkdev/drizzlebox as src/sqlite/utils/
## Context
`@alkdev/drizzlebox` (a fork of `drizzle-typebox` adapted for
`@alkdev/typebox`) provides `createSelectSchema` and `createInsertSchema`
functions that derive TypeBox validation schemas from Drizzle table
definitions. It is consumed as an external npm dependency by all SQLite
table files in `@alkdev/storage`.
The parent project `@alkdev/dbtype` also planned a Phase 1 (UJSX→HostConfig
→Drizzle pipeline) that was never implemented. With SQLite as the sole
database target (ADR-038), the multi-dialect column mappings in dbtype
(PG, MySQL, SingleStore) are dead weight for storage.
ADR-018 deferred dbtype integration to post-v1. But the fold of just the
Phase 0 subset (column→TypeBox mappings + schema generation) is a
straightforward import path change with no behavioral difference.
## Decision
Fold the SQLite-only subset of `@alkdev/dbtype` Phase 0 into
`src/sqlite/utils/`:
| Source (dbtype) | Target (storage) | Changes |
|-----------------|-------------------|---------|
| `schema.ts` | `utils/schema.ts` | Remove PgEnum handling. Keep createSelectSchema, createInsertSchema, createUpdateSchema. |
| `column.ts` | `utils/column.ts` | Strip PG, MySQL, SingleStore branches. Keep SQLiteInteger, SQLiteReal, SQLiteText + generic dataType dispatch. |
| `schema.types.ts` + `schema.types.internal.ts` + `column.types.ts` | `utils/types.ts` | Merged. Remove PgEnum overloads. |
| `constants.ts` | `utils/constants.ts` | Keep as-is. |
| `utils.ts` | `utils/utils.ts` | Remove PgEnum alias. Keep isColumnType, isWithEnum, JsonSchema, BufferSchema. |
All table files change their import from `@alkdev/drizzlebox` to `../utils/schema.ts`.
## Consequences
- **No external drizzlebox dependency** — one fewer npm package in the
dependency graph.
- **SQLite-only column mappings** — dead multi-dialect code removed. If a
new database host is added later, mappings would need to be re-added.
- **Same API surface** — `createSelectSchema`, `createInsertSchema`,
`createUpdateSchema` produce the same TypeBox schemas.
- **Co-located with tables** — utility code lives next to the tables it
derives schemas from. Easier to maintain.
- **dbtype Phase 1 (UJSX→HostConfig) not affected** — that remains a
separate architectural concern, could live in storage or as its own
package when built.

View File

@@ -0,0 +1,58 @@
---
status: accepted
date: 2026-06-01
resolves: OQ-26
---
# ADR-047: HonkerEventTarget Adapter for pubsub
## Context
`@alkdev/pubsub` defines a `TypedEventTarget` interface that all transport
adapters implement: in-process `EventTarget`, Redis, WebSocket client/server,
and Worker. This provides transport-agnostic pub/sub — consumers call
`createPubSub({ eventTarget })` without knowing the underlying transport.
Honker provides SQLite with built-in pub/sub primitives:
- `db.notify(channel, payload)` / `db.listen(channel)` — ephemeral, fire-and-forget
- `db.stream(name).publish(payload)` / `db.stream(name).subscribe(consumer)` — durable, offset-tracked
POC 2-4 (2026-06-01) validated:
- Same-process notify→listen works. ~17ms median latency.
- Multiple concurrent listeners on different channels work.
- `tx.notify()` only fires on `tx.commit()`. Rollback suppresses notification.
- `queue.enqueueTx(tx, payload)` only visible after commit. Rollback suppresses.
- Stream publish/subscribe works with consumer offset tracking.
These results confirm Honker can back the pubsub `TypedEventTarget`
interface for single-node deployments.
## Decision
Implement `HonkerEventTarget` in `src/sqlite/event-target.ts`. It adapts
`@alkdev/pubsub`'s `TypedEventTarget` to Honker primitives with two modes:
1. **Ephemeral mode**: `addEventListener``db.listen()`, `dispatchEvent`
`db.notify()`. Fire-and-forget semantics, no delivery guarantee.
2. **Durable mode**: `addEventListener``db.stream().subscribe()`,
`dispatchEvent``db.stream().publish()`. Per-consumer offset tracking,
crash recovery replays from last saved offset.
`@alkdev/pubsub` is a peer dependency (needed only when using
HonkerEventTarget). The `graphs/` module remains zero-dep.
## Consequences
- **Single-node hub can use Honker instead of Redis** — no separate Redis
deployment needed for pub/sub.
- **Transactional outbox semantics** — `dispatchEvent` inside a Drizzle
transaction (via `tx.notify()` or `stream.publishTx()`) commits atomically
with data writes. No dual-write problem.
- **Hub-spoke symmetry** — both hub and spoke use the same `createPubSub()`
call. Different event target instances determine routing.
- **Multi-node still needs Redis or WebSocket** — Honker events don't cross
process boundaries. For multi-node, the `WebSocketServerEventTarget` (hub)
and `WebSocketClientEventTarget` (spoke) handle cross-process routing.
- **Latency trade-off** — ~17ms Honker round-trip vs sub-ms in-process. For
hot-path call protocol, pair with in-process `EventTarget`. Design of a
composite event target is an open question (OQ-30).

View File

@@ -0,0 +1,74 @@
---
status: accepted
date: 2026-06-01
supersedes: ADR-033 (partially — OQ-17, OQ-18, OQ-19 updated)
---
# ADR-048: OperationSpecs as Repository Surface
## Context
ADR-033 established JSON path queries with hand-written CRUD for v1. The
question was whether the repository layer would be hand-written query
functions, auto-generated from table definitions, or something else.
The `@alkdev/operations` package provides `OperationSpec` — a serializable
descriptor for a query, mutation, or subscription operation. Hubs and spokes
register specs in an `OperationRegistry` along with handlers. The operations
runtime handles execution, call protocol routing, subscriptions, access
control, and validation.
This presents a natural alternative to hand-written CRUD: storage outputs
`OperationSpec[]` describing CRUD operations per table, and the consumer
wires them into the operations registry.
## Decision
Storage does not ship a "repository layer" of hand-written query functions.
Instead, it outputs `OperationSpec[]` per table — flat arrays describing
CRUD operations. The consumer (hub/spoke) imports these specs, registers
handlers, and the operations runtime handles execution.
```ts
// Storage defines table + operation contracts
export const graphTypes = sqliteTable("graph_types", { ... });
export const graphTypeSpecs: OperationSpec[] = [
{ name: "create", namespace: "graph_types", type: "mutation", ... },
{ name: "find", namespace: "graph_types", type: "query", ... },
{ name: "list", namespace: "graph_types", type: "query", ... },
{ name: "update", namespace: "graph_types", type: "mutation", ... },
{ name: "delete", namespace: "graph_types", type: "mutation", ... },
];
// Hub registers specs + handlers
for (const spec of graphTypeSpecs) {
registry.registerSpec(spec);
registry.registerHandler(`${spec.namespace}.${spec.name}`, handler);
}
```
The handler is consumer-provided. Storage defines the contract (input/output
schemas, operation type); the hub provides the execution (Drizzle queries,
Honker transactions, notifications).
`@alkdev/operations` is a type-only peer dependency of storage. No circular
dependency.
## Consequences
- **No hand-written repository functions** — storage avoids the maintenance
burden of typed CRUD code. The contract is the OperationSpec.
- **Consumer owns execution** — the hub decides how to handle each operation
(raw Drizzle query, with Honker notification, with access control, etc.).
Storage doesn't execute queries.
- **Subscriptions for free** — OperationSpec supports `type: "subscription"`.
Table-level change streams become operation subscriptions rather than
custom event handling.
- **Clean dependency graph** — storage depends on operations only for the
`OperationSpec` type (peer dep). No runtime dependency.
- **Attribute queries remain JSON path** — for metagraph tables where
attributes are dynamic JSON, `json_extract()` is still the query mechanism.
Domain-specific tables with native columns produce simpler specs.
- **Supersedes ADR-033 partially** — the "hand-written CRUD" part is replaced.
The "JSON path queries for attributes" part still applies to metagraph
tables.