Files
wiggleverse-ecomm/docs/products-domain.md
T
ben.stull a851d3587c
ci / check (push) Has been cancelled
ci / check (pull_request) Has been cancelled
fix(products): conservative dialect detection — canonical-distinctive columns veto Shopify (SLICE-8 review)
Addresses adversarial-review findings: a canonical file carrying a stray
Shopify-signature name (e.g. SEO Title) no longer misdetects as Shopify and
silently drops canonical Type / corrupts the weight unit (kg->g). A
canonical-distinctive column (Description/Variant Cost/Variant Weight/Google
Product Category/Variant Volume/Tax ID/Position) now vetoes detection, so
detection leans conservative (under-detection warns; over-detection corrupts).
Also closes the dual-named-column shadow (Body (HTML)+Description) and the
stale-weight-unit-on-clear case. Tests cover all three.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-06-12 02:36:09 -07:00

276 lines
15 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# products domain — developer notes
DOC-4 (SD-0002 §11): the import/export spine as built in SLICE-5, extended by
SLICE-68. Operator-facing material is in [`OPERATIONS.md`](./OPERATIONS.md);
the spec is SD-0002 in the content repo.
## Layout and pipeline
`backend/app/domains/products/` is layered like the rest of the app
(main → domains → platform, enforced by import-linter):
- `models.py` — canonical row model + the column registry.
- `codec.py` — bytes → `ParsedFile`; file-level gates only.
- `validate.py` — rows → `CanonicalProduct` blocks + per-row errors.
- `diff.py` — catalog × canonical products → apply plan + preview records.
- `serialize.py` — the export half: `CatalogProduct` snapshot → canonical CSV
(the inverse of `codec`/`validate`). DB-free, like `diff.py`.
- `repo.py` — SQL only: catalog snapshot, draft/run CRUD, apply primitives.
Never commits or rolls back.
- `service.py` — use-case orchestration; owns every transaction boundary and
emits the TEL events via `backend/app/platform/telemetry.py`.
```mermaid
flowchart LR
subgraph validate_path [import_validate]
A[parse_csv] --> B[build_products] --> D[compute_diff] --> E[(import_draft)]
C[load_catalog] --> D
end
subgraph confirm_path [confirm_draft]
E --> F[re-derive: parse → build → diff] --> G{fingerprint match?}
G -- yes --> H[one-transaction apply<br/>run + products + delete draft]
G -- no --> I[PreviewStale<br/>draft kept]
end
```
Confirm re-derives everything from the draft's stored `file_bytes` against the
live catalog, then checks the fingerprint — so what lands is exactly what the
preview showed, or the confirm refuses (`preview_stale`). A confirm with no
adds and no updates refuses with `nothing_to_apply`.
## Canonical model and blank-vs-absent
`models.py` is the one model every dialect maps to (INV-17). `KNOWN_COLUMNS`
is the registry header detection, unknown-column warnings, and validation all
read; `CLEAR_DEFAULTS` holds the reset values for clearable fields
(`status`, `published`, `product_type`, `tags`).
The §6.5.1 cell semantics, as implemented:
- **Absent column** → the field never enters `fields{}` → untouched by diff
and apply (never a change).
- **Present-but-empty cell** → `fields[name] = None` (an explicit clear) →
resolved at diff time to its `CLEAR_DEFAULTS` entry, or NULL where none
exists.
- **The position exception:** a cleared `Variant Position` has no
`CLEAR_DEFAULTS` entry — `diff.resolved_variant_fields` resolves it to the
variant's 1-based file order within its product, never to NULL.
Option names live both in `CanonicalProduct.option_names` (the values) and in
`fields{}` (the file-presence marker the diff needs for the absent-vs-clear
distinction).
## Dialects (INV-17, SLICE-8)
A file is normalized to canonical names **at the codec boundary**, so every stage
downstream of `parse_csv` (validate → diff → apply) is dialect-agnostic — it only
ever sees canonical cells. Two dialects exist: `canonical` and `shopify`.
`dialect_shopify.py` is the Shopify adapter and the source of truth for the
mapping (mirrored by `tests/test_products_dialect_shopify.py` and
`tests/fixtures/shopify-export.csv`):
- **Detection** — `is_shopify_header()` returns `shopify` iff the header carries a
Shopify-only *signature* column (`Body (HTML)`, `Variant Grams`, `Cost per item`,
`Gift Card`, a `Google Shopping / …` column, …) **and no canonical-distinctive
column**. A canonical-distinctive name — one Shopify renames away (`Description`,
`Variant Cost`, `Variant Weight`, `Google Product Category`) or a canonical-only
column Shopify lacks (`Variant Volume`, `Variant Tax ID 1/2`, `Variant Position`) —
**vetoes** Shopify detection. Detection is by signature, not an exact header set, so
extra market/region columns don't defeat it; the veto biases it conservative, because
under-detection is safe (Shopify-named columns are warned as not-imported) while
over-detection corrupts (canonical `Type` / `Variant Weight Unit` would be dropped).
So a hybrid or ambiguous header falls through to `canonical`, which maps shared
columns identically and warns the rest — it never misparses (§7.4).
- **Mapping** — `map_shopify_header(header)` returns `(mapped, not_imported)`:
`mapped[i]` is the canonical name for column `i` (or `None` when it has no
canonical home), and `not_imported` is the warning list surfaced as the draft's
`unknown_columns`. The rule per column, in order: in `SHOPIFY_RENAME`
canonical name; in `SHOPIFY_DROP` (`Type`, `Variant Weight Unit`) → not imported;
in `KNOWN_COLUMNS` → pass through; else → not imported (this last branch absorbs
the unbounded `… / <Market>` price columns without enumerating them).
- **The two name-collision overrides** are why `SHOPIFY_DROP` exists: Shopify's
`Type` is a free-text category, but canonical `Type` is *structural*
(`standalone`/kits), so it is warned rather than folded in; and Shopify's
`Variant Weight Unit` is dropped because the `Variant Grams``Variant Weight`
rename is the weight source.
- **The one value transform** lives in `parse_csv`: when a Shopify row has a mapped
`Variant Weight` (from `Variant Grams`), the codec synthesizes
`Variant Weight Unit = "g"`, since Shopify grams are unitless integers.
`detect_dialect()` in `codec.py` is the INV-17 seam; a new dialect is a new adapter
module plus a branch there, with no change to validate/diff/apply.
## Error granularity
`validate.py` never raises on a row problem: every violation is recorded as a
merchant-language `RowError` and **poisons its whole product block** — the
product previews as `kind="error"` and is excluded from apply, while parsing
continues so one pass yields a complete accounting (BUC-1a). On apply, error
rows are recorded per line in `import_run_error`; `rows_errored` on the run is
the **error-row count**, while the preview's errors tile counts error
**products** — the two numbers legitimately differ.
File-level problems (`not_csv`, `missing_required_column`, `too_many_rows`,
`file_too_large`) raise `FileRejected` in `codec.py` instead: no draft is
created. The BFF adds an early 413 for oversized uploads (`main.py`).
## INV-11 mechanics
`compute_diff` makes **one walk** that produces two views of the same
computation: the typed apply `plan` (resolved natives — `Decimal`, `bool`,
lists) that `confirm_draft` executes, and the JSON-safe preview `records`
stored as draft JSONB and served verbatim to the SPA. Because both derive from
the same walk they cannot diverge. The `fingerprint` is
`sha256(json.dumps(records, sort_keys=True, separators=(",", ":")))`; a
mismatch at confirm means the catalog drifted since preview → `PreviewStale`
(409 `preview_stale`, draft kept for re-validation). The whole apply — run row,
product/variant/image writes, error rows, draft delete — is one transaction;
any exception rolls it back and emits TEL-6.
## Export & the round-trip (SLICE-6)
`serialize.py` is the export half of "one codec, two directions": it turns the
`CatalogProduct` snapshot (the same one `repo.load_catalog` builds for the diff
engine) back into canonical CSV, writing exactly the columns `codec.py` /
`validate.py` parse, in the §6.5.1 row grammar — `HEADER` is the full canonical
column set; product fields + option names sit on the first row; one variant per
row; images interleave; a product with more images than variants emits
image-only rows.
`repo.export_catalog` returns the status-filtered snapshot list (sorted by
handle, deterministic); the `service.export_catalog` generator streams
`serialize.catalog_to_csv` over it and emits TEL-3 (`catalog_exported`) once the
stream is exhausted. The BFF wraps it in a `StreamingResponse`; an empty
(filtered) catalog raises `EmptyCatalog` **eagerly**`409 empty_catalog`
before any bytes stream.
**INV-12** (`diff(catalog, import(export(catalog))) = ∅`) is locked two ways: a
property test (`test_products_serialize.py`) runs the real
export→parse→diff loop over 200 generated **text-field** catalogs and asserts
every product is `unchanged`; the `e2e_roundtrip_noop` browser scenario does the
same through the UI (export download → re-upload → all-unchanged preview, import
disabled). Numeric/`Decimal` fields — the property test's deliberate blind spot
(string-form vs value-identity) — get explicit round-trip unit tests.
## Image pipeline (SLICE-7)
### `platform/objectstore`
A two-adapter port (`backend/app/platform/objectstore/`):
- `local.py` — writes blobs under a configurable local directory; used in
tests and local dev (`ECOMM_OBJECTSTORE_KIND=local`).
- `gcs.py` — wraps `google-cloud-storage`; bucket name comes from
`ECOMM_OBJECTSTORE_BUCKET` (`ECOMM_OBJECTSTORE_KIND=gcs`). Auth is the
VM's service-account ADC — no credential bytes in the overlay or secrets.
Object keys for product images follow the prefix `product-images/` (the
`provision-bucket` gesture also sets a lifecycle rule on `import-drafts/` for
forward-compat, even though the draft blob remains BYTEA this slice — see
seams below).
### `platform/images`
`backend/app/platform/images.py` decodes an uploaded or fetched image byte
string and produces up to four renditions stored in the objectstore:
- **`original`** — stored verbatim after the resolution bar passes.
- **`thumb`** — 150 px on the shorter side, WebP.
- **`card`** — 400 px on the shorter side, WebP.
- **`detail`** — 800 px on the shorter side, WebP.
**Resolution bar (`MIN_IMAGE_SHORT_SIDE = 500`):** images whose shorter side
is below 500 px are rejected (Q-3). This is the only hard gate; oversized
images are scaled down without rejection. All renditions share the same
objectstore key prefix (`product-images/<image_id>/`).
### `domains/products/imagefetch.py` — the fetch phase
After `confirm_draft` commits, the run enters `fetching_images` (if it has
any pending image URLs) and a `ThreadPoolExecutor(max_workers=4)` processes
them concurrently:
- **SSRF guard (INV-18, extended):** only `http`/`https` schemes are
permitted; the resolved IP must be public (RFC-1918, loopback, link-local,
and multicast ranges are blocked); redirects are followed only within the
same guard — a redirect to a private IP is rejected mid-chain.
- **Bounds:** ≤ 20 MB response body, ≤ 30 s per fetch.
- **Per-image presence/idempotency check (`claim_image_for_fetch`):** before
processing, each `CatalogImage` row is checked so an already-handled row is
skipped. This is a presence guard adequate for the **single-worker,
in-process** model — the deployed app runs one uvicorn worker, recovery runs
at startup over prior-process runs, and a fresh confirm always targets a new
run, so two phase invocations never process the same run's images
concurrently. It is **not** a cross-process lock. A true multi-worker claim
would need a distinct in-flight status transition (`pending → fetching` with
`RETURNING`, or `SELECT … FOR UPDATE SKIP LOCKED`) — that is the future seam
if the fetch phase ever moves multi-process or onto a queue (§6.9).
- **Startup recovery scan:** on process start the app scans for runs stuck in
`fetching_images` and re-enqueues their pending images (emits TEL-5). This
covers crash/restart scenarios without a separate scheduler. On a mid-fetch
process restart (e.g. a deploy) the phase is interrupted and the run is left
resumable: pending images stay `pending`, fetched ones stay `fetched`, and
the startup scan completes the run. Any pool-closed error in an in-flight
worker at shutdown is benign — the uncommitted image stays `pending` and is
re-fetched on resume. Mid-flight interruption is tolerated by construction
(§6.9).
- **Progress tracking:** the run row carries `image_progress`,
`image_counts`, and `image_outcomes` (updated per image); these fields feed
the run-detail UI panel.
On completion the phase sets the run status to `complete` and emits TEL-4.
### Image-serving route
`GET /api/products/images/{id}/{rendition}` — storefront-authorized,
streams the objectstore blob. The response sets
`Cache-Control: public, immutable, max-age=31536000` (the object key encodes
the image id, so the URL is stable for the lifetime of the image). INV-16:
an image belongs to exactly one storefront; the route enforces storefront
scope before reading from the objectstore.
### Hosted-URL recognition and INV-12 over images
`domains/products/hosted.py` provides a predicate that recognises whether an
image URL is already served by this app (i.e., a `GET /api/products/images/…`
URL at the current `APP_URL`). The diff pre-pass `_resolve_hosted_images`
uses it to convert hosted URLs back to their `CatalogImage` FK before hashing
the diff fingerprint — this closes the INV-12 round-trip guarantee over
image columns (a re-imported export previews as all-unchanged even when
images are hosted here).
## Named seams (what later slices replace)
- **`import_draft.file_bytes` remains BYTEA (deferred past SLICE-7).** Moving
the draft blob to object storage was scoped out of SLICE-7. The media
bucket holds only `product-images/` objects this slice; the `provision-bucket`
script sets the `import-drafts/` lifecycle rule for forward-compat so the
migration will be clean when it lands.
- **`codec.detect_dialect` → Shopify (SLICE-8).** Today it always returns
`"canonical"`; SLICE-8 recognizes Shopify's exact header set here (INV-17).
- **Run-status complete shortcut → shipped in SLICE-7.**
`confirm_draft` now inserts the run as `applying`, then transitions to
`fetching_images` (if the run has pending image URLs) or directly to
`complete`. The fetch phase fills `image_progress` / `image_counts` /
`image_outcomes` and sets `complete` when done.
## Test map
| File | Covers |
| --- | --- |
| `backend/tests/test_products_codec.py` | file-level gates: parse, caps (INV-18), required columns, dialect |
| `backend/tests/test_products_validate.py` | every §6.5.1 row-error rule, one fixture each |
| `backend/tests/test_products_diff.py` | classification, blank-vs-absent, option matching, fingerprint |
| `backend/tests/test_products_serialize.py` | serializer grammar + INV-12 property test (round-trip no-op over generated catalogs) + decimal round-trip |
| `backend/tests/test_products_export.py` | status-filtered snapshot, streamed export, EmptyCatalog, TEL-3 |
| `backend/tests/test_products_service.py` | draft lifecycle: validate/preview/discard, expiry, TEL-1 |
| `backend/tests/test_products_invariants.py` | INV-10 (never deletes), INV-14 (storefront isolation), apply transactionality, TEL-6 |
| `backend/tests/test_products_endpoints.py` | §6.4 API scenarios + auth/storefront gates |
| `e2e/tests/import-preview-confirm.spec.ts` | happy path: upload → preview → confirm → history |
| `e2e/tests/import-errors.spec.ts` | actionable row errors at preview and on the run report |
| `e2e/tests/import-file-rejected.spec.ts` | file-level rejection, picker stays live, no trace |
| `e2e/tests/import-cancel.spec.ts` | cancel at preview leaves no trace |
| `e2e/tests/export-download.spec.ts` | export downloads canonical CSV, status filter respected |
| `e2e/tests/roundtrip-noop.spec.ts` | export → re-import → all-unchanged, import disabled (PUC-10) |