peter
473d9f02f2
Deep-review round 2 of the 2026-09 bug hunt, over the round-1 fix diff. - A window was LATE by the distance to the keyframe its input seek landed on, and then doubled for good once the full pass merged in: a cue already running at the window's start came out at a negative time, and the webvtt muxer's default shifted the whole output so it sat at 0. Measured on the real Westworld S01E06 at 25:00: every window cue +5.005 s, 682 cues merged instead of 608. `-avoid_negative_ts disabled` keeps the times exact (the parser reads ffmpeg's signed-field spelling of a negative time); now 608/608 on that file, and a real-ffmpeg test pins it. - A full pass that failed was re-armed every minute while the viewer watched what it did publish — a whole-container read per minute. It now backs off for an hour; windows keep running meanwhile. - A failed WINDOW set the track's failure memo, so any uncovered position read "failed" — a final answer that ends the player's poll, even with the full pass about to deliver, and one any account could plant with an absurd `at`. It now marks only the span it failed to read; "failed" is given once the full pass AND a window there have both failed. - Windows join by what they COVER, not a bucketed identity: a backward seek no longer waits behind a window that cannot cover it, and a playhead crossing a bucket edge no longer queues a second read. The queue is bounded per track and overall. - shutdown() sets a flag a worker checks before it spawns ffmpeg (one still in its Plex lookup used to start a read after the kill sweep); queued-job cancellation and the lifespan call are now tested. - A window starts where ffmpeg is told to seek (rounded to the ms it gets), a rounded-up millisecond carries into the minute (no more `00:00:60.000`), a key's first publish claims nothing before its VTT lands, and `at`/`offset`/`start`/`aoff` refuse non-finite values (`at` also an absurd one).
Backend conventions
The rules below are already followed across backend/app — they lived only in a few module
docstrings (demo.py, userscope.py, routes/_common.py) until this file collected them (E3 S3,
B-core-20). When you add code, match these so the next reader doesn't have to guess "which pattern do
I copy".
Module layering
Three layers, by who runs the code — not by feature:
-
routes/…— request-only. A route module holds HTTP concerns: auth dependencies, request/ response models, status-code mapping, serialization. It owns logic that runs only on a request. Business logic that a request happens to trigger does not live here if anything else runs it too. -
plex/,downloads/,sync/— domain logic shared by more than one caller. Anything the scheduler or the worker runs (not just a request) lives in one of these domain packages, so the route, the worker, and the scheduler share one definition. Examples from this codebase:downloads/service.pyowns enqueue + the job state transitions (routes call it);plex/sync.pyis run by both the admin route and the scheduler;downloads/pipeline.pyis the worker's child pipeline. -
Flat
app/*.pycore modules — cross-cutting primitives.auth,sysconfig,quota,audit,notifications,userscope,realtime,ratelimit,progress, … are single-file primitives many layers import. They depend on nothing above them (a core module must never import aroutes/…module).
Extraction trigger: the SECOND consumer. Keep new logic inline in its first caller. The moment a second caller needs it (a worker as well as a route; a scheduler as well as a route), extract it down a layer — into the relevant domain package, or a flat core module if it's cross-cutting. Don't pre-abstract for a single caller; don't leave shared logic duplicated once there are two.
Transaction ownership
This follows CLAUDE.md's canonical rule verbatim — "the handler/job-wrapper commits; helpers flush
(only quota owns its session)" — so read it that way, not as an exception to it. The commit belongs
to the operation owner; anything called inside another transaction only stages its change
(db.add / flush) and lets the owner commit, so the whole operation is one transaction.
audit.record() is the reference: it adds the audit row and the caller commits it together with the
mutation it records, so an audit entry can never exist without its change (or vice-versa).
- The operation owner is usually the route handler. A
service.*operation the handler delegates a whole transition to (e.g.service.cancel) is that owner and commits its one transition — it isn't a mid-operation helper. A helper that owns its OWN short-lived session (the worker'sjobdbwrites, the same carve-out CLAUDE.md grantsquota) commits that session. Everything else flushes. SessionLocalisexpire_on_commit=False: after a rawUPDATE+ commit, an ORM object you already loaded is stale. Read the fresh value from the row (RETURNING, or a re-SELECT), not the ORM attribute — seeservice.decr_ref_count.
Shared counters and status: do it atomically at the DB
- Never read-modify-write a shared counter in Python (
x.count = x.count + 1). Concurrent sessions lose an update. Use a DB-atomic statement —col = col + 1, orgreatest(col - 1, 0)withRETURNINGso the caller branches on the post-value under the row lock (service.incr_ref_count/decr_ref_count). - Download job status is a
JobStatusStrEnum (downloads/status.py), not bare strings. TheHOLDINGset defines which states hold aMediaAsset(so release accounting stays correct);STOPPED_BY_ROUTEis the route-initiated stop set. A worker child's terminal write goes throughservice.finalize_if_running(an atomicUPDATE … WHERE status='running') so a finishing child can't resurrect a job a route just stopped.
Config: one home per kind
app.config.settings— boot/env config (DB URL, download root, worker concurrency): read once at startup, not user-tunable at runtime.sysconfig.SPECS— the runtime admin knobs (DB-override-or-default). A new admin-tunable value goes inSPECS; don't invent a fourth store. (state.py'smaintenance_batchpredates this and is slated to fold intosysconfig— do not add more like it.)
The gate
Everything below must be green before a change is done (run via siftlode check):
ruff (backend) · tsc ×3 + eslint + prettier + knip (frontend) · vitest · pytest. The
pytest lane runs in backend/docker-compose.test.yml — an ephemeral Postgres + the backend tree
bind-mounted — so the DB-backed integration tests (the db fixture) actually run; SIFTLODE_TEST_ REQUIRE_DB=1 there turns a would-be silent skip into a hard failure. See docs/development.md.