Asset entity — UUID-keyed, clean-room design

Esta página aún no está disponible en tu idioma.

Context

artist-alley’s storage layer (ADR 0008, implemented in Phase 1.4) sits below the user-facing entity. storage_objects deduplicates byte streams by sha256; storage_variants records renditions; storage_pins reference-counts ownership. None of these are the thing a user uploads, titles, tags, and shares. That layer needs its own entity.

Two earlier ADRs frame what we want:

ADR 0009 (collection model) assumes resource_id UUID as the membership target. Collections can’t land before this entity does.
ADR 0007 (federation thinking ahead) locks in UUIDs and origin_server_id on every entity that might cross peer boundaries. A typical DAM’s resource.ref BIGINT auto_increment is the opposite of that.

Per memory rule feedback_pre_mvp_everything_is_volatile.md, the DB is throwaway until MVP — so we design the schema cleanly for federation rather than retrofitting a legacy table.

Decision

Name: `assets`

The entity is called assets (table name) and lives under /api/v1/assets. The legacy resource table stays in the baseline migration for PHP coexistence and is gradually retired as PHP pages move to Go.

This conflicts with the Phase 1.4 endpoint group which currently uses /api/v1/assets for raw byte upload/download. We rename the Phase 1.4 endpoints to /api/v1/storage/objects — they were always the low-level byte plane and the name now reflects that.

Schema

The new assets table carries most of the typical legacy resource columns (per user guidance: keep what we haven’t replaced yet, drop later as porting reveals what’s actually used), with three structural additions: UUID primary key, file-hash link to the storage layer, and federation-ready origin tracking.

CREATE TABLE assets (
    id                 UUID         PRIMARY KEY DEFAULT gen_random_uuid(),
    title              TEXT         NOT NULL DEFAULT '',
    description        TEXT         NOT NULL DEFAULT '',
    asset_type      BIGINT       NOT NULL REFERENCES asset_type(ref),
    owner_user_ref     BIGINT       NULL,                     -- nullable so system uploads can exist
    status             TEXT         NOT NULL DEFAULT 'active'
                                    CHECK (status IN ('draft','active','archived')),
    file_hash          TEXT         NULL REFERENCES storage_objects(hash) ON DELETE SET NULL,
    file_extension     TEXT         NULL,                     -- 'jpg', 'mp4', 'glb', ...
    file_size_bytes    BIGINT       NULL,                     -- denormalized from storage_objects for fast filter
    -- Legacy columns we keep until porting reveals whether they're needed:
    rating             INTEGER      NULL,
    user_rating        REAL         NULL,
    hit_count          BIGINT       NOT NULL DEFAULT 0,
    new_hit_count      BIGINT       NOT NULL DEFAULT 0,
    request_count      BIGINT       NOT NULL DEFAULT 0,
    archive_state      INTEGER      NOT NULL DEFAULT 0,        -- legacy "archive" enum; status above is the new layer
    access             INTEGER      NOT NULL DEFAULT 0,        -- legacy "access" enum (0 open, 1 restricted, 2 confidential)
    thumb_width        INTEGER      NULL,
    thumb_height       INTEGER      NULL,
    image_red          SMALLINT     NULL,                      -- dominant-colour theming
    image_green        SMALLINT     NULL,
    image_blue         SMALLINT     NULL,
    colour_key         TEXT         NULL,
    geo_lat            DOUBLE PRECISION NULL,
    geo_long           DOUBLE PRECISION NULL,
    country            TEXT         NULL,
    has_image          BOOLEAN      NOT NULL DEFAULT FALSE,
    is_transcoding     BOOLEAN      NOT NULL DEFAULT FALSE,
    -- artist-alley additions:
    metadata           JSONB        NOT NULL DEFAULT '{}'::jsonb,  -- EXIF / extracted / custom fields
    origin_server_id   UUID         NULL,                          -- federation prep (ADR 0007)
    created_at         TIMESTAMPTZ  NOT NULL DEFAULT NOW(),
    updated_at         TIMESTAMPTZ  NOT NULL DEFAULT NOW(),
    deleted_at         TIMESTAMPTZ  NULL                            -- soft-delete
);

CREATE INDEX assets_owner_idx       ON assets (owner_user_ref) WHERE deleted_at IS NULL;
CREATE INDEX assets_type_idx        ON assets (asset_type)  WHERE deleted_at IS NULL;
CREATE INDEX assets_status_idx      ON assets (status)         WHERE deleted_at IS NULL;
CREATE INDEX assets_file_hash_idx   ON assets (file_hash)      WHERE file_hash IS NOT NULL;
CREATE INDEX assets_created_at_idx  ON assets (created_at DESC) WHERE deleted_at IS NULL;
CREATE INDEX assets_metadata_gin    ON assets USING gin (metadata);

-- Tags as a join table (indexable, queryable; jsonb tags can't be
-- filtered as cheaply at our expected scale).
CREATE TABLE asset_tag (
    asset_id   UUID  NOT NULL REFERENCES assets(id) ON DELETE CASCADE,
    tag        TEXT  NOT NULL,
    added_at   TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    PRIMARY KEY (asset_id, tag)
);

CREATE INDEX asset_tag_tag_idx ON asset_tag (tag);

Notes on the column choices:

status is artist-alley’s new vocabulary (draft / active / archived). The legacy archive_state enum (0..5 with site-specific semantics) lives alongside as an integer; mapping happens at the PHP/Go boundary until legacy pages retire.
file_hash ties the asset to the byte plane (storage_objects). Multiple assets pointing at the same hash is fine and intentional — that’s the whole reason content-addressed storage exists.
metadata jsonb is the extensibility safety valve. EXIF blobs, extracted technical metadata, custom-field overrides land here rather than as new columns. Indexed via GIN.
tags are a join table, not jsonb. Indexable; query patterns like “every asset tagged X” stay fast.

API surface

# raw-byte storage (renamed from Phase 1.4's /assets)
POST   /storage/objects                    — upload raw bytes, return hash
GET    /storage/objects/{hash}             — download bytes
GET    /storage/objects/{hash}/variants/{variant}  — download variant

# asset entity (Phase 1.8, this ADR)
POST   /assets                             — create asset (body: title, type, optional file_hash)
GET    /assets                             — list (paginated, filter by owner/type/status/tag)
GET    /assets/{id}                        — fetch one
PATCH  /assets/{id}                        — partial update
DELETE /assets/{id}                        — soft-delete

GET    /assets/{id}/file                   — proxy to /storage/objects/{file_hash}
GET    /assets/{id}/variants/{variant}     — proxy to /storage/objects/{file_hash}/variants/{variant}

POST   /assets/{id}/tags                   — add tags {tags: ["x","y"]}
DELETE /assets/{id}/tags/{tag}             — remove

The proxy endpoints (/assets/{id}/file, /assets/{id}/variants/*) mean the front-end never has to know an asset’s underlying hash — better encapsulation, and we can swap storage backends without breaking links.

Storage pinning

Creating an asset auto-pins the underlying byte blob with {subject_type: "asset", subject_id: <asset.id>}. Deleting the asset removes the pin; the storage GC sweeper later removes the bytes if no other pin remains. The existing user: pin from upload gets removed when the asset is created (its work is done).

Migration from legacy data

Explicit non-goal for Phase 1.8. The DB is volatile until MVP; fresh installs start with the new model. If we ever need to import legacy data, it lands as a separate Phase later under scripts/import-rs-resources.go.

Consequences

Positive:

Clean UUID-keyed entity ready for federation, collections, search.
Storage layer / entity layer naming finally distinguishes the byte plane from the user-facing record.
Soft-delete + JSONB metadata + tag join table cover the extensibility cases without re-architecting later.
Legacy coexistence preserved — resource table untouched; PHP keeps rendering through the translator.

Negative:

Breaking rename of Phase 1.4 endpoints (/assets → /storage/objects). Acceptable per volatility rule; will be a single-commit change with test-shim updates.
Two tables (assets + the legacy resource) describe overlapping things during the transition. Operators with no legacy data installed can ignore resource entirely; we drop it from the baseline when no PHP page on the served route table needs it.

Deferred:

Search / filter DSL (ADR 0010) is what powers GET /assets?... beyond simple equality filters. Until 0010 lands, list endpoint supports only owner / type / status / tag exact-match filters and cursor pagination.
Variant generator (was Phase 1.4.D) finally has somewhere to attach — it lives as a background job that produces thumbnails and previews for each new asset, writing storage_variants rows. Scope: ships with Phase 1.8 if simple (image only), else immediately after.
Resource_type as UUID-keyed entity (currently BIGINT FK). Phase 1.8 keeps the BIGINT FK to avoid a coordinated change; if/when asset_type ports to UUIDs, assets.asset_type swaps.

Open questions

Whether asset versioning belongs in this model (replace the file but keep title/tags/comments). Useful for “v2 of this concept piece” — but adds complexity. Lean: defer to a follow-up, model versions as separate assets linked via metadata.parent_asset_id until demand is clear.
Whether owner_user_ref should be UUID. The user table itself still uses ref BIGINT; porting that is its own large phase. Keep BIGINT for now, swap when users move to UUIDs.