ef6d3a077ab8770910246e8c3999d9ea57512109
95 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
MechaCat02
|
ef6d3a077a |
fix: close what nine adversarial reviews found, most of it mine
Some checks failed
Checks / Backend — cargo test + clippy + fmt (push) Failing after 1m5s
Checks / Frontend — vitest + svelte-check (push) Failing after 5m55s
Checks / E2E — typecheck + lint (push) Failing after 49s
E2E / Playwright E2E (chromium + webkit) (push) Failing after 10m42s
E2E / Cross-UA smoke matrix (push) Failing after 7m57s
Audit / cargo audit (backend) (push) Failing after 11m12s
Audit / npm audit (frontend) (push) Successful in 44s
Nine focused reviews (export state machine, upload path, auth/abuse, client queue, guest UI, database, deploy/ops, regression hunt, test honesty). Every finding below was re-verified against the code before being acted on; several plausible-sounding ones were checked and rejected. ## Data loss and denial of service **One request could OOM-kill the app container.** `client_upload_id` was read with `Field::text()` — axum builds its multipart reader with no SizeLimit, so the only bound was the route's 576 MiB body limit, then decoded into a second full String. `caption` and `hashtags` go through `read_text_field_bounded` for exactly this reason; this field arrived later and missed it. Any guest, one request, and every SSE stream drops and every in-flight temp file is stranded. **Nothing bounded concurrent upload bodies.** The headroom gate can only refuse to COMMIT — the body is already streamed to a temp file by the time it runs, and neither axum, the tower stack nor Caddy limits how many stream at once. ~100 guests tapping "upload all" after the ceremony puts 10-20 GB of .tmp on a 40 GB volume, invisible to the gate, eating the reserve that keeps Postgres able to write WAL. New `UploadAdmission` budgets bytes (not requests, so one video and two hundred photos coexist) via a permit that releases on drop, so every exit path returns it. **The export decode bypassed the memory permit the compression path takes.** Same class of work — decode + resize every image in the gallery — in a bare spawn_blocking. A release fired while the last photos were still compressing put both in the same 1 GiB cgroup; the OOM kill marks the export failed and `recover_exports` re-spawns it into the same conditions on the next boot. The permit is now process-wide in `imaging`, because the constraint it expresses is the container's memory, not one worker's. **`MediaTotalCache` cached its own failure as 0.** For the whole TTL the gate then saw an empty event and collapsed to the flat reserve — the behaviour the two-halves design replaced — with no log line. And the trigger correlates with the danger: with max_connections 10 the query fails exactly during a burst. Now falls back to the last good reading and says so. **V8's heap ceiling sat above the frontend container's entire budget** (measured: 259 MB inside a 256M limit), so GC could never intervene and the only backpressure was SIGKILL under an arrival burst. ## Guest-visible **The feed stopped being newest-first after the first reconcile.** It fetches whole 100-item server pages while `uploads` grows in 20s, so everything in the gap was absent from `present`, classified as new, and prepended — ~80 photos from earlier in the evening above the newest ones. It also stalled infinite scroll, since the cursor still pointed at item 20 and the observer only re-fires on a change. The union is now sorted on the server's own (created_at, id) key, which additionally places an SSE arrival correctly. **A stale `loadMoreError` outlived every refresh and filter change**, leaving a false error above a button that returns immediately on `!nextCursor`. **A failed derivative toasted "Ein Upload konnte nicht verarbeitet werden."** for a photo sitting right there on screen — the handler still assumed 1d9fb11's pre-fix behaviour (row deleted, quota refunded, card evicted), none of which is true any more. It was the last surviving route for the "your photo is gone" signal that fix set out to remove. ## Enforcement that existed only in comments `recover_name_rate_per_15min` is clamped at the point of use: the ordering `3 x ceiling <= PIN_LOCK_THRESHOLD` is the whole control against one source locking any guest whose name is on the feed, it was asserted in a comment, and `patch_config` accepted 1..100_000. The test pinned the default constant rather than the enforced bound; it now pins the bound. ## Tests that could not fail - The gate test asserted only its own premise (`500MB x 100 > 35GB`) and never touched the gate. It now checks both controls against the same state and requires them to disagree in the right direction. - `the_banner_always_fires_before_the_upload_gate_closes` reduced to `G < G + G/4` — true for any margin, including zero, so it could not detect the banner moving to exactly the gate. It now pins the gap. - `disk_is_low`'s `free < LOW_DISK_FLOOR_BYTES` clause was unreachable (warn_at is always >= 12.5 GB against a 10 GB floor). Two tests were named after it and neither could fail if it were deleted. Clause and constant removed. - The suspension test I added last commit hard-coded the credit cap instead of importing it, so changing STALL_TIMEOUT_MS would leave it passing against a system that no longer exists. Now imports MAX_SUSPEND_CREDIT_MS. ## Stale comments corrected The prune doc still argued at length for the pre-build ordering that |
||
|
MechaCat02
|
214f9e3062 |
fix: close four confirmed defects an adversarial review found
Findings from a multi-angle review, most of them in code I wrote in the last
few commits. Each was verified against the code before being acted on.
## Backend
**The export daily limit was bypassable ~60x/minute.** `SseTicketStore` is
untyped, and the export download quietly started reusing it. `POST
/stream/ticket` is free and rate-limited at 60/min per user; `POST
/export/ticket` charges one of three PER-DAY downloads. So a guest could mint
at the cheap endpoint and redeem at the expensive one, each redemption
streaming the whole multi-GB keepsake, `no-store`, off the same filesystem
Postgres writes WAL to. Tickets now carry a `TicketKind` and `consume` requires
it to match, asserted in both directions. The comment claiming "one mint is at
most one download" was simply false.
**`export_ticket` answered 200 `{"ticket": null}` when the store was full** —
after charging a daily slot. `issue` returns `Option`; `sse.rs` handles the
None with a 503 and this call site unwrapped it into the JSON body. The page
toasted success, the iframe navigated to `?ticket=null`, and one of three
downloads was gone. That is the phantom-success failure the pre-validation in
|
||
|
MechaCat02
|
253878e027 |
fix(export): give the keepsake viewer the two-phase preflight it was meant to get
The two-phase preflight from
|
||
|
MechaCat02
|
5b705317ef |
fix: close the last three guest-facing dead ends (items 7-9)
C1 — a failed page-append silently ended infinite scroll `loadMore`'s catch showed a toast and changed no state, unlike every sibling error path in the file. `nextCursor` survived so the feed was still technically paginable, but the IntersectionObserver only fires on a CHANGE: after a failed append nothing scrolls and no rows are added, so it never re-fires. One 429 or wifi blip and the guest concluded the gallery was 20 photos. Now leaves a retry control at the sentinel — a toast that fades in 5s is not an affordance — resuming from the untouched cursor. C2 — the export page reported downloads that never happened `downloadFile` toasted 'Download gestartet' the instant it assigned the iframe's src, before a single byte existed. Since the iframe swallows errors BY DESIGN (a top-level navigation to a 404 would unload the PWA), a failure produced a green success message, a consumed single-use ticket, and one of only three daily slots spent — repeatable until the day's allowance was gone, on the screen that is the whole point of the app. Root cause is two sources of truth: `export_status` reports `done` from `export_job` and enables the button, while the download resolves through `export_current.file_path` plus a `Path::exists()`. They can legitimately disagree. `export_ticket` now takes a `kind` and calls the existing `resolve_export_file` BEFORE charging the rate slot, so a missing archive fails honestly on a plain fetch that `toastError` already renders. Not the HEAD probe ruled out elsewhere: it reads the same indexed row the download will read and touches no ticket, so it cannot consume anything. The parameter is optional, so an older client degrades to today's behaviour rather than breaking. C3 — the WhatsApp journey could dead-end with no error at all The join link travels through guest group chats, and a link tapped inside one opens in that app's browser, where the file picker and getUserMedia both depend on the host app having wired them up. When they aren't, the buttons do nothing — no error, nothing to act on. Two targeted changes rather than a UI rebuild: the camera error panel now offers "Aus Galerie wählen" (its advice to change "Browsereinstellungen" refers to settings that do not exist in a webview, so retrying could never help those guests), and the sheet carries a standing one-line hint to open the link in Safari or Chrome. Deliberately no user-agent sniffing: a sniff list is wrong for every browser it has not heard of, while a quiet standing hint costs one line and is never wrong. The hint lives in UploadSheet rather than the root layout because both layout banners are gated on `$showBottomNav`, which `/upload` turns off — one there would never render on the composer. Verified: 151/151 backend tests against a live Postgres, clippy clean, 58/58 vitest, svelte-check 0 errors, eslint clean, both builds. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
MechaCat02
|
05063694d2 |
fix: close eight regressions the audit pass found, five of them mine
Two adversarial reviews over |
||
|
MechaCat02
|
23e2f485dd |
fix(upload): correct two errors in the keepsake headroom gate
Both found reviewing my own change rather than by a test, which is the point. DOUBLE-SUBTRACTION. The gate computed `free - size`, but the body is streamed to its temp file during multipart parsing — far above the gate — so the free-space reading already excludes those bytes. Subtracting again refused uploads a full file-size early; with max_video_size_mb at 500 that is half a gigabyte of phantom pressure. `media_total` genuinely does need `+ size` (its row is not committed yet), which is what made the asymmetry easy to miss. BLOCKING SCAN ON THE HOTTEST PATH. It called `disk::free_bytes`, whose doc comment says it deliberately bypasses DiskCache — but that rationale is the export preflight's: a rare, high-stakes decision where a sibling worker can move free space by tens of GB inside the TTL. Per upload it means sysinfo re-scanning every mount, synchronously, on the async runtime, on a 2 vCPU box with two worker threads. Now uses the cached snapshot, the same 15s staleness the quota check immediately below already accepts for the same question. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
MechaCat02
|
eb0e405562 |
fix: gate uploads on keepsake headroom, and close five unattended-event gaps
The box is 2 vCPU / 4 GB / 40 GB, not the 4 vCPU / 8 GB / 80 GB that the audit, the committed comments and README's sizing section all assumed. That correction is what the first change is about; the rest are the remaining pre-event items. THE ARCHIVE COULD BECOME UNBUILDABLE WHILE UPLOADS KEPT SUCCEEDING `required_free_bytes` is `media × 1.1 × 2` — the ZIP and the HTML viewer are each gallery-sized — and the export preflight also wants DISK_RESERVE_BYTES on top. The upload gate, though, only refused below a FLAT 10 GB reserve. On 40 GB that let uploads run to ~25 GB of media while a release needed `2.2 × 25 + 10` = 65 GB free. Every upload in that band succeeded and the keepsake could then never be built: the product's entire promise, failing silently at the end of the night with nobody there. The gate now enforces the invariant that actually matters — never accept an upload that would make the keepsake unbuildable — sharing `required_free_bytes` with the preflight so the two cannot drift into disagreeing about the same question. Uploads stop at ~8 GB of media on this disk, with a German message naming the cause. Refusing the 1001st photo beats losing all 1000. `media_total.rs` backs it: SUM(user.total_upload_bytes) over ~100 rows, cached 5s, rather than `estimate_export_bytes`'s join across every upload. It counts hidden and banned users' bytes, which the export excludes — skew in the SAFE direction, so the gate closes marginally early rather than late. Fails open on a query error. A test pins the gate against the preflight across the whole gallery-size range, and a second asserts the per-user floor alone would over-commit the volume — i.e. that the global gate is what must bind. THE WATCHDOG ABORTED HEALTHY UPLOADS EVERY TIME A PHONE WAS POCKETED `Date.now()` advances while a backgrounded phone is frozen but `setInterval` does not, so the first tick after a screen lock read the whole sleep as silence and aborted — re-sending a video from byte zero and burning one of five PERMANENT auto-attempts. The interval is now its own suspension detector: a tick that arrives 125s late for a 5s schedule credits that window back, because a period the watchdog could not observe is not evidence of silence. Chosen over a `visibilitychange` listener, which only covers causes that fire that event — a throttled-but-visible tab, a closed lid and an occluded window all freeze timers without one — and which would have needed module state, an SSR guard and a teardown for strictly less coverage. `performance.now()` was rejected because Safari pauses it across system sleep on some paths and Chrome does not. The credit buys one fresh window, not immunity: a socket iOS reaped while backgrounded still aborts ~90s after resume rather than hanging for `xhr.timeout` (5-60 min) with the queue's `processing` latch held. Two latent leaks found while in there: `xhr.abort()` on a request already in readyState DONE emits no `abort` event, so `settle()` never ran and the interval re-aborted every 5s forever while `activeUploads` kept a stale entry (the ✕ button silently stopped working); and a synchronous throw from `xhr.send` — a blob whose backing store the OS purged — leaked the same way. Both closed. OKLCH MADE THE DELETE BUTTON INVISIBLE ON SAMSUNG'S DEFAULT BROWSER red/amber/green were never in the @theme block and fell through to Tailwind v4's `oklch()` defaults, which Safari <15.4, Chrome <111 and Samsung Internet <22 cannot parse: `var(--color-red-600)` is then invalid at computed-value time, `background-color` falls back to transparent, and `.btn-danger` renders white text on nothing. Pinned to Tailwind's own defaults gamut-mapped to sRGB by Lightning CSS — the converter already in this pipeline — so modern browsers render exactly what they render today. Verified against seven hex fallbacks it had already emitted for the /alpha forms. rose and teal (avatar chips) had the same leak. The app CSS goes from 40 oklch declarations to 0. Also fixes `--color-purple-950`, which was simply missing: `dark:bg-purple-950/50` on the host dashboard was rendering default violet on EVERY browser, off-brand. The keepsake viewer only picks this up on a rebuild, so its committed artefact is rebuilt here too — still single-file, still zero external references. A BRICKED BOOT LOOKED LIKE A SPINNER FOREVER With `ssr = false` the page is empty until the bundle mounts, so a chunk 404 after a redeploy or a dead uplink left the guest on the boot spinner with no message, no reload control, and in a standalone PWA no URL bar. A 15s timeout in the existing nonce'd IIFE (no CSP change) swaps in German copy and a reload button. Deliberately a timeout rather than feature detection: a SyntaxError in the bundle is invisible to any capability check. Plus a <noscript>, since there was nothing at all to see without JS. EVERY 4xx WAS INVISIBLE AT ANY LOG LEVEL tower_http counts 4xx as a success, so it logs at DEBUG while production runs at info. If guests spend the evening hitting 429s or 413s, the post-event logs said nothing. Now one WARN per client error; 5xx excluded because Internal already logs its source chain and the pool-exhaustion 503 logs at construction. A DEAD FRONTEND SERVED A BLANK 502 `handle_errors 5xx` with an inline German page (the caddy service mounts only the Caddyfile, so there is no volume to ship a static file through). Verified empirically against this config, not from documentation: an upstream 404 through `reverse_proxy` still arrives as untouched `application/json`, and only a dial failure renders the page. That mattered — the keepsake download navigates a hidden iframe and DEPENDS on a real 404/429 arriving, and swallowing those would have been worse than the blank 502. CONFIG CORRECTIONS FOR THE REAL HARDWARE DATABASE_MAX_CONNECTIONS 30 → 15: sized to 2 vCPU rather than to the guest count. Since migration 024 a feed page costs well under a millisecond, so connections are no longer spent waiting, and 30 backends crowd the db container's 1 GB on a 4 GB host. COMPRESSION_WORKER_CONCURRENCY stays at 2 — the merged heavy-image permit already serialises anything over 150 MiB, so the "two 48 MP photos" worst case that number was sized against is unreachable; dropping to 1 would halve light-path throughput and push more feed tiles onto full-size originals. README's sizing section rewritten for the actual disk. Verified: 149/149 backend tests against a live Postgres, clippy clean, 57/57 vitest, svelte-check 0 errors, eslint clean, vite build, export-viewer rebuild, caddy validate, compose YAML parse. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
MechaCat02
|
1d9fb11c7b |
fix: close the nine ways an unattended event loses photos or dies
Every one of these was found in the pre-event audit, verified against source, and
survives to production on the current main. Grouped by what actually goes wrong.
PHOTOS DISAPPEAR
* compression.rs no longer soft-deletes on a failed derivative. The guest got a
201, watched the card appear, then watched it vanish — the row left v_feed,
find_visible_media and BOTH keepsakes, while its bytes sat on disk for 14 days
waiting for a cleanup nothing announced. No screen anywhere lists compression
failures, so recovery meant hand-written SQL that also had to re-add the
refunded quota. Now it does exactly what the ENOSPC arm beside it already did
and documented as correct: keep the row, serve the original, retry on the next
boot (bounded by derivative_attempts). `upload-deleted` is no longer emitted;
`upload-processed` is, so the card re-renders instead of sitting on a
placeholder.
* A 413 is now a reversible lock, so the blob survives. The quota moves — free
disk falls, uploader count rises — so a guest goes over it having done nothing,
and treating that as permanent meant a 400 MB video was pushed across cellular
in full and THEN deleted from IndexedDB. Gone on both sides, and unrecoverable
for an in-app camera capture that exists nowhere else.
* quota_limit_bytes gained a floor and a stable divisor. The ceiling used to
decrease monotonically all evening; it now settles at max(uploaders,
estimated_guest_count) — a config key that was seeded, validated in the admin
whitelist, and read by no code at all. The floor is clamped to what the disk
can actually back, so a full volume still yields zero rather than handing out
an allowance it cannot honour.
* Because that floor gives up the aggregate guarantee the formula used to imply,
uploads now check a hard 10 GB reserve first, independent of every quota
toggle. postgres_data, media_data and exports_data share one filesystem: the
end state was not a degraded feature, it was Postgres unable to write WAL.
THE ARCHIVE DISAPPEARS
* prune_superseded_archives runs only after the new generation lands. It ran
before the preflight, reasoning the old archive was already unreachable — true
of reachability, false of recoverability. An epoch is a value that can be
rolled back; deleted bytes cannot. Any failed rebuild left the event with NO
keepsake at all.
* The export preflight reserves the same 10 GB. `free < needed` authorised an
export sized at exactly free, which ran for half an hour and landed the box at
zero with the keepsake still unfinished.
THE APP DIES
* The feed reconcile re-reads the id set after its awaits instead of reusing one
captured up to three round-trips earlier. The new-upload SSE handler prepends
during exactly that window, so the row was both already present and absent from
the stale set — prepended twice, and a duplicate key in a keyed {#each} throws
in production, not just dev. The SSE handler and loadMore now dedupe too.
* Added routes/+error.svelte. Without it any uncaught error fell through to
SvelteKit's unstyled English 500 with no reload control — inside a chromeless
standalone PWA with no URL bar, for the rest of the evening.
THE OPERATOR IS LOCKED OUT
* admin_login verifies the password BEFORE charging the rate bucket, and a
correct password is never throttled. The old order made this a denial of
service against its own operator: every guest shares one NAT IP, the check ran
first, so five requests a minute from any phone in the room kept the bucket
full — and the escape hatch needed the admin session being blocked. A generous
separate ceiling still bounds bcrypt CPU.
THE PROJECTOR DIES
* The preload budget is now strictly inside the dwell. At the 3s option the 4s
budget could never land a commit on a slow uplink, so the wall froze on one
photo while the queue drained silently behind it.
* The wake lock retries every 30s while visible, and the page says so on screen
when the browser has no wake lock API. visibilitychange was the only retry
trigger and a kiosk never changes visibility, so one refusal — iOS in Low Power
Mode, say — was permanent.
* Caddy: /api/v1/upload/*/display joins the cacheable carve-out. The backend set
max-age=300 on it and the blanket no-store silently replaced it, so a projector
re-fetched a full-size JPEG per slide, ~2-4 GB over an evening on the uplink
the guests are uploading over.
Also removes Upload::soft_delete, now unreferenced and an unscoped footgun next
to soft_delete_in_event.
Verified: 146/146 backend tests against a live Postgres, clippy clean, 51/51
vitest, svelte-check 0 errors, eslint clean, vite build, caddy validate, compose
YAML parse.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
MechaCat02
|
61119be817 |
Merge branch 'fix/deploy-unattended-blockers' into main
Two independent lines of production hardening diverged at |
||
|
MechaCat02
|
e1c689d1a7 |
chore(migrations): renumber 021-023 to 023-025 to clear the collision with main
main independently shipped 021_hashtag_counts_respect_bans and 022_client_upload_idempotency. sqlx::migrate! embeds ./migrations and refuses two files per version, so these three had to move before the branch could merge at all. Contents are untouched — only the version prefixes change. Renumbering (rather than renumbering main's) is the safe direction: main is already deployed, so its 021 and 022 are applied in production and their versions are now immutable. |
||
|
Fabian Hamm (Privat)
|
edc5f1f62c |
chore(export-viewer): rebuild the embedded bundle, and fix the lint ignores
The offline keepsake viewer had the same two filter defects as the app: typed suggestions were capped so a matching tag could be unselectable, and the dropdown used `onmousedown` with a backdrop that swallowed the selection. Rebuilt into `backend/static/export-viewer/index.html`, which `include_dir!` embeds in the binary. The eslint ignores were unanchored, so once the export-viewer's dependencies were installed its nested `.svelte-kit` output was linted as source. Anchored with `**/`. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Fabian Hamm (Privat)
|
87d01a8a26 |
fix(export): charge the download limit where the client can see the answer
The keepsake download is an iframe navigation, so its response is invisible to the page. The rate limit was enforced inside the zip/html handler — i.e. inside that navigation — while the ticket POST in front of it always returned 200. A guest over the limit therefore tapped "Herunterladen" and absolutely nothing happened, forever, with no explanation, on the one screen that is the emotional payoff of the whole app. With the default of 3/day, ZIP + HTML costs 2 and one retry locks them out until tomorrow. Minting is a normal `fetch`, so the limit moves there and the 429 reaches the user. The limit is not weakened: tickets are single-use with a 30s TTL and can only be obtained from that authenticated endpoint, so one mint is at most one download — and charging it in both places would have cost every download two slots. The message named the wrong timescale too. It shared the generic "warte kurz" wording with the per-minute limiters, but this bucket is a DAY, so a guest was told to wait a moment for something that could not work again until tomorrow. Verified live: three mints succeed, the fourth returns 429 in German; raising `export_rate_per_day` through the admin API takes effect on the next request with no restart, and the HTML keepsake then downloads. Also here, from the same pass: - `looks_bcrypt` checks the SHAPE of ADMIN_PASSWORD_HASH, not just placeholder-ness. A hash corrupted by shell or Compose escaping is not a placeholder, so the app booted green, `/health` said ok, and every admin login 401'd — unrecoverable mid-event, because the Admin row is only created BY a successful admin login and promoting a host requires one. - The rate limiter indexed `timestamps[0]` while holding its mutex, so a `max == 0` configuration panicked and poisoned the lock process-wide. Uses `first()`. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Fabian Hamm (Privat)
|
496dba5a1f |
fix(feed): filter on the server, exactly, and keep banned uploads out of the chips
Filtering was split across two independent client-side states and applied to whatever page 1 happened to hold, by caption SUBSTRING. So a tag chip selected in the list view was silently still applied in the grid without being shown; a filter matched photos whose caption merely contained the text; and anything past the first page was invisible to it. Verified against the seeded data: `hashtag=tanz` returned 6 photos by substring, 1 by tag. `FeedQuery` now carries `hashtag` (single, list view), `hashtags` (CSV, OR'd, grid chips) and `uploader` (exact, AND'd), normalised through one function that trims, strips `#`, lowercases and dedupes, and yields None when empty — so an empty filter means "no filter", never "match nothing". The two SQL branches collapse into one with `h.tag = ANY($4)`. Tag-OR plus tag+user-AND is a specified feature, not an accident: `e2e/specs/03-feed/filter-search.spec.ts` and USER_JOURNEYS §8 pin it, which is why the semantics moved to the server rather than being simplified away. Tags travel as CSV safely because the backend restricts them to ASCII alphanumerics and `_`; `uploader` stays a single exact parameter because a display name can contain a comma. New `GET /api/v1/uploaders` reads `v_feed`, so banned and hidden uploaders are excluded for free. Migration 021 gives `v_hashtag_counts` the same treatment. It counted every upload regardless of the uploader's ban state, so banning a guest left their tags in the chip list as ghost filters that lead to an empty feed. Verified: after banning the guest who owned all six `tanz*` photos, the chips went 6 -> 0. `?limit=-5` returned a 500 — only the upper bound was clamped, so Postgres was asked for `LIMIT -4`. Clamped at both ends. `is_banned` is added to `/me/context` so the client can show a read-only notice instead of letting a banned guest discover the ban one 403 toast at a time. `add_comment` sorts and dedupes hashtags on the normalised key, matching the upload path — the two disagreed, which is a lock-ordering deadlock between concurrent upserts. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Fabian Hamm (Privat)
|
1d0df3ebf6 |
feat(upload): make uploads idempotent so a lost response cannot duplicate a photo
The ordinary mobile failure, not an exotic one: the server receives the body, validates it, commits the row — and the response is lost on the way back because the guest walked out of range or the AP dropped the connection. The client sees a network error with the blob still in hand and re-sends it, both when the guest taps "Erneut" and automatically when the queue requeues on reconnect. Every attempt minted a fresh `Uuid::new_v4()` server-side, so the same photo landed in the gallery two or three times and was charged against the guest's storage quota each time. The client already has a stable per-queue-item UUID, so it costs nothing to send. Migration 022 adds `client_upload_id` with a partial unique index — partial so the NULLs of every pre-022 upload, and of any caller that doesn't send one, keep working untouched. Two paths, because there are two races: - Sequential retry: a lookup before the transaction finds the stored row, deletes the re-sent bytes and replays the original response as 200. The body has necessarily already been streamed, since the key arrives as a multipart field — re-sending is the client's cost and is already paid by the time we see it. What must be prevented is a second ROW. - Concurrent retry: two attempts in flight at once. `ON CONFLICT DO NOTHING` returns no row to the loser, which abandons its transaction (quota increment included) and replays the winner. Letting the unique index raise instead would only surface after the transaction had aborted, as an opaque error the caller would have to string-match. The replay reads live state rather than assuming a fresh row: a reconnect can be minutes later, by which time the derivatives may exist and the photo may have been liked. Every read there fails soft — the upload is already safely stored, so a sparser response is fine and failing the request is not. Verified live: the same photo sent three times returns 201, 200, 200 with one id, one row, and the quota charged exactly once. Also in this file: the two image-header probes at admission now run on `spawn_blocking`. Both open the file and run the codec's header parse synchronously, and `#[tokio::main]` gives two worker threads on a 2-vCPU box — so every upload stalled half the runtime's request-serving capacity. Everything else that blocks here (image encode, bcrypt) was already offloaded; this was the one that wasn't. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Fabian Hamm (Privat)
|
faea555967 |
fix(ops): make /health a readiness probe, and bound how long the pool waits
`/health` returned the literal string "ok" and touched nothing. Every request in this app needs the database, so that answered a question nobody asked: the container reported healthy while every request 500'd, and with no operator watching during the event there was no signal at all. It now runs `SELECT 1` with a 2s timeout. Verified live against the production stack: 200, stop Postgres, 503 "database timeout", start Postgres, 200 again — with no app restart, because sqlx revalidates on acquire. Deliberately NOT wired to automatic recovery. Compose's `restart: unless-stopped` does not react to healthcheck state anyway, and an autoheal sidecar would be actively wrong here: it would truncate every in-flight upload to "fix" an outage that, as the test above shows, clears on its own. This is a diagnostic — including for the runbook's event-day `curl`. The pool had only `max_connections` set. Three additions: - `acquire_timeout(5s)`. sqlx defaults to 30, so a DB blip parked every request AND all ~100 SSE session revalidations for half a minute before erroring — the app looked hung rather than degraded, and the backlog outlived the blip. - `min_connections(2)`, so the first request after the setup-to-guests-arriving gap doesn't pay TCP + auth. - `statement_timeout=15s` / `lock_timeout=5s` per connection. Without them a single pathological query holds a pool slot indefinitely and no client-side timeout can take it back, because the slot is only released when Postgres finishes. Those two SETs are sent as two statements. `sqlx::query` uses the extended query protocol, which permits exactly one per call — as `SET a; SET b` every new connection failed, which surfaced as the pool never opening one and `create_pool` reporting a connect timeout. Caught by booting against an empty database rather than a warm one. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Fabian Hamm (Privat)
|
157499d493 |
fix(ops): reclaim abandoned upload temp files, and supervise the task that does
`stream_field_to_file` removes its `.tmp` on every error return, which covers everything the handler can see. It cannot cover what actually happens at a party: the client goes away — a phone sleeps, a guest walks out of range, the PWA is evicted mid-video — and axum DROPS the handler future rather than returning an error, so no cleanup runs at all. The shutdown backstop force-exits in-flight handlers for the same net effect. Nothing else reclaimed them. `cleanup_deleted_media` only visits rows with `deleted_at`, and an abandoned upload never got a row; `export::sweep_orphan_temps` is only ever pointed at the exports volume. `grep -rn read_dir src/` had three hits, all in export.rs — the media tree was never read by anything. So every abandonment stranded up to `max_video_size_mb` of unowned bytes permanently, on the same 40 GB filesystem as `postgres_data`. Worse than a leak: the per-user quota is computed from live free disk, so those bytes were also subtracted from what everyone else was allowed to upload. An evening of flaky venue wifi could take the event down. The threshold is on modification time, not creation time, which is what makes an hour safe: a live upload is written to continuously so its mtime keeps advancing and it can never age into the sweep no matter how slow the connection. The clock only starts once the writer stops. The periodic task is now supervised. It carries every piece of recurring hygiene in the app — session pruning, media reclamation, this sweep, and the rate-limiter and SSE-ticket maps — as a bare `tokio::spawn` with no retained handle, so a single panic anywhere inside it stopped all five permanently and silently. No log line, no symptom until the disk or a HashMap grew into one. Tested: an hour-old temp is reclaimed, a temp still being written to is not (deleting that one destroys a live upload), a committed `.jpg` is never touched, and a media tree that does not exist yet is a silent no-op rather than an error logged 24 times a day. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Fabian Hamm (Privat)
|
2f952494c2 |
fix(media): stop a poster-frame failure from deleting the guest's video
Reproduced live, by accident, while smoke-testing on a machine with no ffmpeg: the clip uploaded fine, returned 201, and roughly six seconds later had `deleted_at` set and was gone from the feed. The `Ok(None)` "this clip yields no frame" case was already handled — that fix landed when sub-second clips were being destroyed. But the `?` on the call itself still routed every OTHER failure into the same give-up path, which soft-deletes: ffmpeg missing from the image, ffmpeg hanging on a truncated `.mov` and tripping the timeout, an ENOSPC on `thumbnails/`, or a DB blip in `set_thumbnail_path`. None of those says anything about the video, and `get_original` serves the file byte-for-byte, so a post that merely lacks a poster is fully watchable. No failure in the video branch may fail the upload. iPhone `.mov` is exactly the input most likely to trip it, and a wedding clip is not retakeable. ENOSPC gets its own classifier. It was the one failure the retry loop actively made worse: a disk does not drain during six seconds of backoff, so all three attempts failed identically while holding a compression permit that photos were queued behind — and the give-up path then refunded the quota and soft-deleted the row while deliberately KEEPING the original. That freed nothing, removed the photo seconds after a 201, and handed the guest the allowance to upload it again into the same full disk. Now: no retry, no refund, no delete. The row stays live and the photo is served from its original, and `backfill_stale_derivatives` regenerates the derivatives on the next start once there is room. `is_storage_full_error` has to look inside `ImageError::IoError` as well as at bare io errors, because `image` wraps rather than sources it and a plain chain walk would miss every derivative-write failure. FFMPEG_TIMEOUT drops 120s -> 45s. It was never a budget for honest work — a poster from a phone clip takes well under a second, and `-ss` before `-i` means even a 500 MB file seeks rather than scans. It is the ceiling on how long a pathological input holds a permit that guests' photos are waiting behind, so it should be as tight as it can be without cutting off real work. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Fabian Hamm (Privat)
|
43d37269b6 |
deploy: pull prebuilt images instead of building on the event server
The production compose still carried `build:` keys and no `image:` keys, so a
`git clone` onto the CX22 followed by `docker compose up -d` would have started a
fat-LTO release build of 427 crates on a 2-vCPU/4 GB box — the outcome the whole
build-on-the-Mac decision exists to avoid, reached silently because `pull` skips a
service it is told to build rather than failing.
Both services now pull `registry.mc02.dev/eventsnap/*:${EVENTSNAP_VERSION}` with the
`:?` form, so a missing tag fails the command instead of resolving to an empty one.
`docker-compose.build.yml` restores the `build:` keys for the workstation that
produces the images, from the same context paths.
Also here:
- `DOMAIN` gets the same `:?` guard. Blank did not fail — it produced `https://` for
the frontend's ORIGIN and collapsed the Caddyfile's site block into a malformed
global block, so the stack came up with no TLS and no site.
- `stop_grace_period: 20s` on the app. Docker's default stop timeout is 10s, exactly
the app's own drain budget, so a redeploy could SIGKILL the process at the moment it
was finishing — truncating the in-flight upload the graceful shutdown protects.
- `COMMENTS_ENABLED` is pinned "false" alongside MEDIA_PATH. It is a product decision
for this event, and `.env.example` ships the generic `true`; pinning it means an
operator who copies the example and edits only the secrets cannot ship comments on.
- The frontend runtime stage now copies the lockfile and uses `npm ci`. Without it the
three `^`-ranged deps re-resolved at build time, so an image rebuilt days later could
differ from the one that was tested. Image also drops 120 MB -> 65 MB.
- `docker-compose.dev.yml` told the operator that production had the same `$`-eating
bug and to escape the hash as `$$` in `.env`. That is wrong and it breaks a working
deployment: Compose uses single-quoted env_file values literally, and doubling
produces a 74-character string `looks_bcrypt` rejects. Verified with `printenv`.
The runbook's rollback pointed at `v0.12.0`, which has 6 migrations against HEAD's 22
and was never built or pushed — running the emergency card's rollback line would have
crash-looped the app with `VersionMissing` during the event. §9 now has you tag one
build twice so the rollback target is bit-identical, and says plainly what that can and
cannot fix. Every `$DOMAIN` command gained the `set -a; . ./.env` it needs, the
down-migration psql commands are wrapped in `sh -c` so the container expands the
credentials rather than sending `-U ""`, and the advice to lower `max_video_size_mb`
is withdrawn: the client guard it was premised on does exist, but is pinned to a
compile-time constant, so lowering the DB value only moves failures later.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
MechaCat02
|
117e67fa80 |
fix(auth,upload): close the admin lockout and four unbounded-input paths
ADMIN LOCKOUT. admin_login looked its user up BY NAME. Migration 007 makes
display_name unique per event case-insensitively and join had no reserved-name
guard, so any guest joining as "admin"/"Admin"/"ADMIN" before the operator's first
login made find(role == Admin) miss, the fallback create("Admin") violate that
index, and `?` return a 500 — permanently, with no in-app recovery. Moderation,
config and gallery release all gone; the fix was hand-editing the database.
The root cause is the lookup key, not the creation. The name was never the
identity. User::find_admin_for_event resolves by role, which makes the whole class
of name collisions irrelevant — including the homoglyph bypasses of the new
reserved-name list, which is now defence in depth rather than the control.
Promoting the squatting row would be the obvious fix and is a serious mistake: it
carries a recovery_pin_hash the guest knows, so it would hand them the admin
dashboard via /recover, permanently, through a path needing no password. A
separate row under a fallback name is worse UX and much better security. Verified
against the real schema — the guest keeps their uploads, PIN and session under a
freed name, and the role lookup then finds exactly one admin.
Second, independent bug in that block: create() followed by a SEPARATE UPDATE ...
SET role = 'admin' manufactures the same poisoned state if anything fails between
them. Collapsed into create_with_role.
UNBOUNDED INPUTS — one root cause, four places: validation ran after the
allocation.
- upload caption/hashtags used Field::text(), which buffers the whole field, on
the one route whose DefaultBodyLimit is 576 MiB — so 576 MiB of heap per
concurrent request in a 1 GiB container, with the length check running
afterwards on a string already built. Now refused mid-read.
- the hashtag CSV was never length-checked at all and was upserted tag by tag
INSIDE the commit transaction, which holds FOR SHARE on the event row — one
request could stall every other upload behind tens of thousands of round trips.
Capped at 30 tags of <=50 chars.
- /recover and /recover/request built rate-limiter keys by format!() from an
unvalidated, unbounded display name, retained up to 24h in a map pruned hourly:
the limiter itself became the memory-exhaustion primitive it exists to prevent.
join validated first; that check is now shared by all three. /recover/request
also had no per-IP ceiling at all — /join got one in 017, /recover in 019, and
019's own comment describes exactly this attack. It returns 204 rather than 400
on a bad name, because a 400 would be a new signal on an endpoint whose contract
is that it cannot enumerate guests.
- the SSE ticket store had no size cap, no per-session cap and no rate limit on
its endpoint, while prune ran hourly against a 30s TTL. Now pruned on issue,
capped, and rate-limited. At capacity it REFUSES rather than evicting a
stranger's ticket — evicting would let one client deny SSE to the venue. Not
one-ticket-per-session either: two tabs open their EventSources concurrently.
PATCH /upload/{id} had no rate limit, no validation, and called
invalidate_and_arm unconditionally — outside both `if let Some` guards. So
PATCH {} bumped export_epoch and armed a fresh pair of full-gallery export workers
every call; REGEN_DEBOUNCE bounds the rate of that, not the total work, so a guest
could keep the keepsake permanently un-downloadable. All three fixed. The
validation also resolves a divergence: upload normalised tags while edit stored
them raw, so #Party via edit and party via upload became two hashtag rows.
PIN LOCKOUT was an ordering bug before a policy one: the account-lock threshold
(3) sat BELOW the per-(IP, name) ceiling (5), so three requests from one IP locked
any guest whose name is on the feed, every 15 minutes, forever. The tier meant to
protect a guest was the cheapest way to attack them. Ceiling drops to 4, threshold
rises to 12, so locking a victim now needs at least three distinct sources.
Brute-force cost is unchanged — 48 attempts/hour means 10k PINs still take ~208h
regardless of IP count — and increment_failed_pin now decays the streak after 15
minutes, since the counter previously only cleared on success and honest typos
accumulated across days. Both invariants are pinned by tests rather than comments.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
MechaCat02
|
f275de5c8f |
perf(feed): stop every feed page from aggregating the whole event
v_feed computed like_count/comment_count with LEFT JOINs and a GROUP BY. Postgres CAN push `event_id = $1` and the keyset predicate through the view — verified with EXPLAIN, it uses idx_upload_event_created_id — but it CANNOT push ORDER BY ... LIMIT across a GroupAggregate. So every request aggregated every upload in the event, times its likes and comments, and only then sorted and took 21 rows. Cost grew with the event, not with the page. Measured on a throwaway database seeded to a real reception (1000 uploads, 100 guests, 27k likes, 10k comments), same query, same data: before GroupAggregate (actual rows=1001) -> Sort -> Limit 449 ms after Index Scan (actual rows=21) -> Limit -> SubPlans 0.58 ms migration 022 replaces the joins with correlated scalar subqueries, which puts the counts ABOVE the Limit so they run 21 times instead of 1001. Exactly equivalent, not merely close: "like" is keyed (upload_id, user_id) so COUNT(DISTINCT user_id) == count(*), comment.id is the PK so COUNT(DISTINCT c.id) == count(*), and the GROUP BY was on u.id so it was already one row per upload. Column names, order and types are unchanged, so no Rust changes. No new index needed — idx_like_upload and idx_comment_upload already match the subqueries. Note the existing load harness cannot see any of this: e2e/loadtest/driver.mjs creates no likes and no comments, so the expensive path had never been exercised. The amplifier, feed/+page.svelte: every open feed subscribed to `upload-processed` and refetched page 1 — the most expensive page — so ~100 open feeds each fired one per completed upload. Now gated on whether this client actually shows the card that changed, and the debounce is jittered, because a fixed delay just moves a simultaneous herd 800 ms later. Nothing is lost by skipping: a client without the card also missed its `new-upload`, and the reconnect `feed-delta` already schedules a refresh. Load shedding, because the above reduces the risk rather than removing it: db.rs set no acquire_timeout, so sqlx's 30 s default applied — longer than the frontend's own 20 s fetch timeout, meaning the browser gave up while the server kept holding the slot and the work was done for nobody. Now 5 s, and PoolTimedOut maps to 503 + Retry-After instead of a generic 500. That mattered because the upload queue classifies 5xx as transient and retries: a 500 sent the retries straight back into the saturated pool with nothing to pace them. PoolClosed stays Internal — it only occurs during shutdown, where a 503 would invite a retry against a server that is going away. The Retry-After extraction in into_response matches on variants, so unlike message() a missing arm is not a compile error — it would silently drop the header. Pinned by a test covering both retry-carrying variants. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
MechaCat02
|
e52b2f1cd1 |
fix(upload): reclaim the bytes of uploads that never finish
Reclaim was a dozen explicit remove_file calls on the handler's return paths. That covers every way the handler can FINISH and none of the ways it can simply STOP. When a client disconnects mid-body — a phone leaving wifi, iOS killing a backgrounded PWA, the user hitting back — axum drops the handler future at an .await inside field.chunk() and no return path runs at all. The partial file then survives forever: it has no upload row, so cleanup_deleted_media (row-driven) can never see it, and no sweeper covered the originals directory. The triggers are routine rather than adversarial, and the client keeps the blob and auto- retries on every `online` event, so one large video over bad wifi leaves several copies. Those bytes are also invisible to the quota while still consuming the free disk that compute_storage_quota divides among guests — so orphans silently shrink every guest's ceiling while the admin widget under-reports. All three volumes share one filesystem; the end state is Postgres unable to write WAL. TempFileGuard is an RAII guard, because dropping the future is exactly what runs Drop — it is the only construct that survives cancellation. Armed before the file can exist, disarmed only after tx.commit() succeeds. The twelve explicit cleanups are deleted so one owner holds the rule. The subtler half is the rename. It happens BEFORE the commit, so between them the file exists under its final name with no row pointing at it — an orphan that looks legitimate. The guard is RETARGETED there rather than disarmed, and the retarget sits on the same poll as the rename with no .await between, which is what makes that window uncancellable. sweep_orphan_originals is the backstop for the process that was killed, where no Drop can run at all. Hourly, alongside the existing sweeps: .tmp files past the window go unconditionally (a .tmp never has a row by construction), other files are batched 500 at a time through a single NOT EXISTS query. Two things that look like oversights and are not, both commented in place: - the 6h window is what makes the sweep safe against the rename-before-commit ordering, since a committing upload is briefly indistinguishable from an orphan. It must not be shortened to speed up a test. - the NOT EXISTS deliberately does NOT filter deleted_at IS NULL. A soft-deleted row still points at its file during its retention window, and reclaiming that is cleanup_deleted_media's job; filtering here would race the two sweeps and destroy the files the recovery window exists to preserve. Verified against the real schema: given a live original, a soft-deleted one and a true orphan, the query returns only the orphan. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
MechaCat02
|
5969ec74ea |
fix(compression): bound derivative retries so one bad upload can't loop forever
The OOM in the previous commit was survivable; what made it an outage was that it repeated. The upload row is committed before compression starts, derivatives_rev defaults to 0, and set_derivatives_rev only runs on success — so a row whose processing killed the container survived at rev 0, and backfill_stale_derivatives (called unconditionally at every boot) re-selected it and re-ran the identical workload. With restart: unless-stopped that is an infinite kill loop, and every cycle also drops every SSE stream and truncates every in-flight upload. Verified end to end against the real schema in a scratch database: with the new guard the backfill selects the row on boots 1-3 and zero rows from boot 4 on, and a later success resets the counter. migration 021 adds derivative_attempts and derivative_last_error. The counter is incremented WRITE-AHEAD, before the work is attempted. This is the whole design: the failure being bounded is a cgroup SIGKILL, so no Err is returned, no error handler runs and no Drop fires. A counter bumped in a failure path increments zero times per crash and the loop would be unchanged. set_derivatives_rev clears it, so success is the only reset and both the live path and the backfill get it without a new call site to forget. Also in the backfill: - one task walking the rows sequentially instead of one task per row. A large backlog used to spawn thousands of tasks, each holding a pool handle and queueing on the same two permits, competing with live uploads for a whole boot. - LIMIT 200 per boot, and original_path <> '' replacing an IS NOT NULL that was dead (the column is NOT NULL; cleanup_deleted_media blanks it instead). - a once-per-boot error log naming how many uploads have given up. Without it the give-up is invisible — the loop stops, which is the point, but the photos keep a stale derivative forever with nothing to notice. Adds backfill_video_posters for the mirror-image gap: a video interrupted by a restart has its compression_status flipped processing -> failed by startup_recovery and is never re-enqueued, so thumbnail_path stays NULL for the rest of the event while the clip itself plays fine. It shares the same attempt budget, which means a genuinely posterless sub-second clip (Live Photo, mis-tap) stops being re-ffmpeg'd after three boots. That is intended, not a bug to fix later — Ok(false) is a normal permanent outcome there. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
MechaCat02
|
a4bac03628 |
fix(compression): stop a single large PNG from OOM-killing the container
An 8000x8000 RGBA PNG passes admission — 256,000,000 bytes is just under the 256 MiB max_alloc, and smooth content is under 3 MB on disk, far below any size cap. Processing it peaked at ~1250 MiB inside a 1 GiB cgroup. Measured, not argued: the new (ignored) test builds exactly that image and reads VmHWM around the pipeline, resetting the watermark via /proc/self/clear_refs so the number covers only the code under test. Three independent causes, all of which had to go: 1. The decode outlived everything. `resize` takes &self, and the no-downscale arm bound `img` into `display`, so the ~244 MiB buffer was still alive when oxipng ran — and oxipng decodes the PNG *again*, holding a full-size buffer per filter trial. The decode now lives in a block that yields the display derivative; the else arm moves `img` out, which is what makes "the block's value is the only survivor" true in both arms. 2. oxipng was unbounded in every dimension: preset 2 with timeout: None, and the default features pull in rayon, which evaluates filter trials concurrently with a full-size buffer each and has no Options knob to cap it. Now gated at 8 MP, given a 20 s timeout, and built with default-features = false so oxipng's own sequential shim is used. "filetime" is kept — without it preserve_attrs silently no-ops. Dropping "binary" also stops compiling clap/glob/env_logger (a CLI's deps) into the server image. 3. Even with those fixed it still measured 516 MiB, and compression_concurrency defaults to 2 — so two guests uploading big photos at once was another OOM, 1032 MiB against a 1 GiB limit. The cost is dominated by image's Lanczos3 resize, which accumulates in f32: the intermediate is new_width * old_height * 16 bytes, i.e. 262 MiB for this image — larger than the decode itself, and invisible to max_alloc. Two changes: the preview now derives from the 2048px display instead of re-resizing the original (one full-size pass, not two), and a job whose header-estimated peak exceeds 150 MiB takes an exclusive permit so two giants can never overlap. Ordinary photos (a 12 MP JPEG estimates ~50 MiB) never touch that permit, so throughput is unchanged for everything except the case that must not run in parallel. Chaining 8000 -> 2048 -> 800 for the preview is not a quality trade: a staged Lanczos3 downscale is standard for large ratios and is visually indistinguishable at 800px. The blocking half is now a free function so its memory behaviour is testable — the fix is a scoping property a future edit could silently undo. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
MechaCat02
|
6fd75adb27 |
fix(ops): bound logs, drain ffmpeg's stderr, and stop three silent hangs
Six independent operational defects, none of which needed a new feature to fix. Log rotation. Docker's json-file driver is unbounded by default, and those files land on the HOST filesystem — outside every deploy.resources.limits in the compose file, and on the same disk as postgres_data and media_data. A full disk stops Postgres writing WAL, which takes the event down. Capped at 10m x 3 per service. Log level. RUST_LOG was set in neither .env.example nor docker-compose.yml, so the code fallback WAS the production level — and it was `debug`, with tower_http=debug emitting a line per request and per response into that unrotated file. Now info, with tower_http=warn to state that those spans are diagnostics, not an access log. ffmpeg pipe deadlock. run_ffmpeg piped stdout and stderr and then called wait(), which drains neither. Once the ~64 KiB pipe buffer filled, ffmpeg blocked writing and wait() never returned — burning the full 120s timeout, twice per seek position, three times per compression attempt. And the timeout is an Err, so the end state was a soft-deleted upload: a guest's playable video destroyed by a poster-frame failure. Now stdout is null (nothing ever read it) and stderr is drained by wait_with_output, whose tail is logged on a non-zero exit. Note wait_with_output consumes the child, so the old kill-on-timeout is gone; kill_on_drop(true) already covers it. Readiness probe. /health never touched the pool, so the disk-full endgame above stayed green all the way down. Adds /health/ready (SELECT 1 under 2s) as a SECOND route — the compose healthcheck deliberately keeps pointing at /health, because caddy gates its startup on it and a DB-dependent probe would turn a Postgres blip into the reverse proxy refusing to start. api.ts request timeout. The abort timer was cleared in a finally around fetch(), which resolves on the response HEAD — leaving res.text() uncovered and no longer abortable. An upstream that sends headers then stalls the body hung the call forever. The timer now lives until the body is read, including the 204 path (which otherwise leaked a live 20s timer per no-content request). Upload XHR watchdog. The XHR had no timeout while processQueue held the isProcessing latch across it; on a half-open socket neither error nor abort ever fires, so the latch pinned and the queue wedged. Bounds SILENCE rather than total duration — a 500 MB video over a venue uplink legitimately runs 30+ minutes while making steady progress. Rejects as NetworkError, which is already the retryable branch, so a stalled upload now recovers like any network blip. IndexedDB failures. addToQueue called getDb() unguarded and handleSubmit had no catch, so a private-mode refusal or a QuotaExceededError on a large blob left a permanent "Wird hochgeladen…" spinner, no toast, and — for an in-app camera capture — the only copy of the photo gone. Now reported as 'failed', which keeps the staged files on screen and stays on the page. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
fabi
|
402215d405 |
fix(video): extract a real poster frame, and stop claiming one that isn't there
Both the compression worker and the HTML export ran the same invocation:
ffmpeg -i <src> -vframes 1 -ss 00:00:01 -vf scale=… -y <out>
`-ss` AFTER `-i` is an output-side seek. Against a clip of a second or less ffmpeg
exits 0 and writes nothing, and both call sites gated on the exit status:
- the worker wrote `thumbnail_path` and logged "thumbnail generated" for a file that
was never created, so GET /upload/{id}/thumbnail 404s in the live feed;
- the export listed media/<id>_thumb.jpg in data.json while the ZIP writer skipped the
unopenable file, so the keepsake drew a broken image tile.
Any clip at or under a second, which phones produce constantly — mis-taps, Live
Photos, boomerangs. Not data loss; the .mp4 is in both archives and plays. Every
server-side signal stayed green.
New `services/video.rs` owns the extraction for both callers, mirroring the imaging.rs
precedent (created for the same duplication, and it paid off when the max_alloc fix
landed in both workers at once). Three changes in it:
- `-ss` before `-i`, an input-side seek. NOT sufficient alone: verified against the
production image, seeking to 1s in a 1.000s clip is still past the last frame and
still exits 0 with no file. The 0s fallback is what actually fixes this, and 1s is
tried first only because an opening frame makes a poor poster.
- Verify the artifact, not the exit status. This is the check both sites were missing.
- Carry compression.rs's 120s timeout. export.rs had NONE — a hung ffmpeg there would
strand the job at `running` and the keepsake would never complete.
The worker's call used `?`. Tightening the check without also making a missing poster
non-fatal would have been far worse than the bug: every sub-second clip would fail
compression, exhaust its retries and be soft-deleted. It now logs a warning and leaves
`thumbnail_path` NULL, which FeedListCard, VirtualFeed and LightboxModal already
handle.
The export now sets `thumb: ""` and skips the manifest entry when there is no poster —
and does the same for the IMAGE branch, whose decode failure left the identical
dangling reference. No viewer change was needed: +page.svelte already guards
`{#if post.media.thumb}` and falls back to a video tile with a play glyph. The comment
claiming "viewer handles missing thumbs gracefully" was true about the viewer and false
about what the backend sent — the guard never fired because the string was never empty.
e2e/specs/06-export/export-video.spec.ts had DOCUMENTED this as intended behaviour
("the fixture clip is <1s, so ffmpeg extracts no thumbnail frame — but exits 0 …
that's the intended shape here"). sample.mp4 is exactly 1.000s, so every video test in
the suite ran at that boundary and none ever fetched the poster. That comment is now
corrected to say what it actually was.
Tests: 2 unit; a new sample-5s.mp4 fixture so the ordinary first-seek path is covered
at all; video-playback now FETCHES the poster rather than asserting the attribute (the
one extra request that nine rounds of green never made); a new spec covering both
fixtures plus the mirror that a posterless video still uploads and plays; and a
keepsake spec asserting every <img> in the opened viewer resolves — naturalWidth === 0
is exactly the broken-tile case, whatever produced it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
fabi
|
3bcb7c6a76 |
fix(deploy): close the POSTGRES_PASSWORD trap that broke a fresh deploy either way
README step 2 said to set "DOMAIN, JWT_SECRET, ADMIN_PASSWORD_HASH, EVENT_NAME, etc." -- POSTGRES_PASSWORD was not in that list. .env.example said to set it and keep it in sync with DATABASE_URL. The two documents disagreed, and both readings ended badly. Branch A, README verbatim: the stack came up GREEN and healthy on CHANGE_ME_use_a_strong_password -- a database credential published in the public repo. The production secret guard covered JWT_SECRET and ADMIN_PASSWORD_HASH, and nothing anywhere looked at the Postgres password. Branch B, .env.example verbatim: a permanent restart loop, "password authentication failed for user eventsnap". The guard's own design made Branch B near-certain. It stops the APP on the first `docker compose up -d` -- but not the `db` service in that same command, which initialises its data directory and bakes in whatever password was in .env at that moment. POSTGRES_PASSWORD is honoured ONLY at initdb. So the intended recovery -- see the refusal, fix your secrets, boot again -- was exactly the sequence that broke it. Nothing in the error named the cause, and the remedy (`down -v`) is both unguessable and the one command you must never run once real data exists. Three changes, which have to ship together: the guard alone would just move operators out of Branch A and into Branch B. - The guard now rejects a placeholder DATABASE_URL in production (the password rides in that URL, which is what the app actually reads). Branch A can no longer boot. - It reports EVERY unset secret in one message instead of returning on the first. Fixing two secrets used to cost two boot cycles, on a stack where Caddy waits on the unhealthy app throughout, and each avoidable cycle is another chance to reach for -v. - A 28P01 handler in db.rs turns the unguessable failure into a self-explaining one: it names the initdb semantics, gives `down -v` with an explicit "deletes db + media + exports, no undo", and gives the ALTER ROLE alternative for when data already exists. Docs: step 2 now names POSTGRES_PASSWORD and says every secret must be set BEFORE the first up; the troubleshooting block covers the auth-failure loop and both remedies. Also replaces htpasswd (apache2-utils -- not on a stock VPS) with `docker run --rm caddy:2-alpine caddy hash-password`, an image the stack already pulls, in README, .env.example and the guard's own message. Verified the output ($2a$14) against the shipped $2y$12 example. Verified on a real Postgres, not just in tests: Branch A refuses and names DATABASE_URL; all three placeholders report in one boot; a volume initialised with one password and connected to with another prints the diagnostic; an unrelated connect failure (dead port) stays silent. 8 unit tests, including that non-prod ignores all of it so the e2e stack is unaffected. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
fabi
|
08d92b7531 |
fix(audit-7): keepsake viewer integrity, readable archives, quota_tolerance floor
Seventh round, both defects in the keepsake — the one artifact that leaves the
system entirely, and so the one where a green server-side signal carries no
information at all.
Squashed from 2 commits, original messages preserved below.
──────── fix(export): stop a caption bricking the viewer, and ship readable archives
Two defects in the keepsake, both silent server-side and both only visible by
extracting the real artifact and trying to use it.
1. A CAPTION COULD BRICK THE VIEWER.
The viewer's data is inlined as `<script>window.__EXPORT_DATA__={…}</script>` -- it has
to be, since guests open index.html over file:// where fetching a sibling data.json is
blocked. The escape was `</` -> `<\/`. Against XSS that holds; I fired
`</script><img src=x onerror=…>` through a real Chromium parser and it round-trips
inert.
It does not stop the caption steering the HTML TOKENIZER. `<!--<script` with no later
`-->` drives the parser into script-data-double-escaped state, where the template's own
`</script>` only steps back to script-data-escaped instead of closing the element.
Everything after it -- including the viewer bundle -- is swallowed as script data.
Nothing executes and nothing leaks: `__EXPORT_DATA__` is simply never assigned and the
keepsake opens blank. A denial of the deliverable, not an XSS.
Reproduced in Chromium before changing anything, and the near-miss is worth recording:
`<!--<script>alert(1)</script>-->` comes back CLEAN, because the trailing `-->` returns
the parser to script-data state. A probe using the terminated form quietly repairs the
thing it is testing for.
Fix: escape every `<` as `<`, not just `</`. `<` never appears in JSON structural
syntax -- only inside string values -- so a global replace is sound, and one rule covers
`</script`, `<!--` and `<script` together. That is the point: the old escape was named
for the single case it handled. Only the INLINED copy is escaped; data.json is written
separately, in no HTML context, and stays literal.
2. EVERY ENTRY IN BOTH ARCHIVES WAS STORED MODE 0000.
`ZipEntryBuilder::new` leaves the external file attribute at zero and async_zip's host
compatibility defaults to Unix, so `unzip -Z` showed `?---------` on every line of both
Gallery.zip and Memories.zip. Windows Explorer ignores Unix modes, which is why this
survived; on Linux and macOS `unzip` faithfully applies what the archive asks for and
the guest gets a folder of photos none of which they can open.
Unconditional -- every keepsake ever produced, no hostile input required -- and
invisible server-side: the export succeeds, the ZIP is well-formed, the job writes
`done`, /export/status is green.
Found by accident. The browser test for defect 1 failed with ERR_ACCESS_DENIED on
file://, which looked exactly like a Playwright sandbox quirk; I twice "worked around"
it (fresh context, then a separately launched browser) before checking the extracted
files and finding mode 000. The workaround was suppressing a real bug. Both workarounds
are gone -- the ordinary `page` fixture loads the archive fine now.
Fix: all six ZipEntryBuilder sites route through one `keepsake_entry` helper stamping
`S_IFREG | 0644`, so the mode cannot be forgotten at a call site.
Tests: 3 unit (no `<` survives; the payload still decodes to the original value, because
this is a transport encoding and not a sanitiser; a clean payload is untouched) and 2
e2e that release for real, download the real archives, and check them from outside the
app -- one opening index.html over file:// in Chromium and asserting the viewer booted,
the captions came back verbatim and nothing executed; one asserting every stored mode
and every extracted file is readable. Both assertions verified to FAIL against the
pre-fix artifacts.
──────── fix(admin): reject quota_tolerance = 0 instead of silently blocking every upload
Zero is inside the documented 0–1 range and catastrophic. The per-user limit is
`free_disk * tolerance / active_uploaders`, so a tolerance of 0 makes every limit 0 and
refuses EVERY upload -- mid-event, with "Du hast dein Upload-Limit für dieses Event
erreicht", an error naming the wrong cause entirely. An admin reaching for an off-switch
wants `storage_quota_enabled`; the rejection now says so.
Rejecting the value rather than raising the floor. A floor of 0.01 was the obvious fix
and it is wrong: very small tolerances are legitimate -- they are how a large disk is
throttled down to a sensible per-guest ceiling, and how the quota specs steer it
(tolerance = target * active / free lands around 1e-5 on the 174 GB volume this suite
runs on). A floor would forbid real configurations, and would have broken the entire
storage-quota describe block, to prevent one typo. Verified: those four tests still pass.
Tests: the rejection, that the stored value is untouched (validation fully precedes any
write), and the mirror -- 0.00001 still round-trips -- so the guard can't quietly become
a floor later.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
fabi
|
06bc9ddcb3 |
refactor(export): share the visibility filter between the row query and the estimate
`query_uploads` selects the rows the archives are built from; `estimate_export_bytes` sizes them for the disk preflight. They stated the same WHERE clause separately, and the direction of drift matters: an estimate that MISSES rows the archive writes under-reserves, which is precisely the ENOSPC the preflight exists to prevent. The integration test claimed to guard this and cannot. Both sides of `the_estimate_sums_exactly_the_rows_the_archive_will_contain` are `SRC:`-marked hand-copies in tests/common/mod.rs -- neither is production code -- so drift means production moved while both copies sat still, and the test goes on passing. The convention is sound for pinning behaviour; it is structurally incapable of detecting divergence from the thing it copies. So fix it where it can be fixed. One `export_visibility_where!()` fragment, `concat!`-ed into both queries at compile time (still `&'static str`, no allocation), with the `u`/`usr` alias contract stated. Divergence is now impossible by construction rather than watched for. The tests keep their value and lose the overclaim: the docstrings now say they pin WHICH uploads may be counted -- each excluded row in the fixture is excluded by a different predicate, so weakening any one of them still fails here -- and say plainly that they do not detect drift, with a pointer to what does. No behaviour change. The filters were verified identical before the hoist (`u.event_id = $1 AND u.deleted_at IS NULL AND usr.uploads_hidden = FALSE AND usr.is_banned = FALSE`); 99 backend tests still pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
fabi
|
5f702f2b40 |
fix(audit-5): export disk preflight, media reclaim, low-disk warning, restore docs
Fifth round, all storage. The keepsake could not fit on the documented hardware
and failed halfway through a multi-GB write, leaving the deliverable stuck; the
quota had stopped bounding the disk; and the backup had no restore procedure.
Squashed from 6 commits, original messages preserved below.
──────── fix(export): refuse an export that cannot fit, and stop peaking at two generations
Nothing in export.rs ever asked whether the keepsake would fit. Both archives write
their media `Compression::Stored`, so each is essentially a byte-for-byte second copy
of the originals -- Gallery.zip always, and Memories.zip for every video and every
image at or under 5 MB. On the documented CX33 (80 GB, all three volumes on one
filesystem) the upload quota's fixed point leaves ~40 GB free, and a release spawns
BOTH halves concurrently against it.
The failure is not "the export failed", it is "the deliverable is stuck":
1. ENOSPC lands partway through a multi-GB write.
2. The epoch has already moved, so the job row is `failed` at the CURRENT
generation and readiness (epoch = event.export_epoch AND status = 'done') is
false -- GET /export/zip 404s.
3. The last good archive sits on disk, unreferenced and unreachable.
4. POST /host/export/rebuild, the only escape, re-arms the same doomed write.
Three changes.
Reclaim before building. `prune_stale_export_files` ran only after the new archive
was written, renamed and finalised. That reads as durability but buys nothing: the
moment `invalidate_and_arm` bumps the epoch the old archive is ALREADY unreachable,
so keeping it reserves gigabytes for a download nobody can perform -- and for a
takedown it is content someone explicitly asked to have removed. Peak usage is now
one generation. Narrower than the post-finalize prune on purpose: final archives
only, never a `.tmp` or a `viewer_tmp_` dir, since a superseded worker can still be
streaming into those and at build START is far more likely to be alive.
Preflight the space. SUM(original_size_bytes) over exactly `query_uploads`'
visibility filter, +10% for ZIP overhead, multiplied by the number of armed jobs --
without that multiplier each of the two concurrent halves independently sees "it
fits" and together they don't. Runs AFTER claim_job, not before as reported: bailing
before the claim leaves the row `pending` with no worker and no error, the
spinner-forever state `mark_failed`'s status guard exists to prevent. Fails open when
the mount can't be read, exactly as the upload quota does.
Show the host the reason. /export/status returned {status, progress_pct} and nothing
else, so the host dashboard could only render "fehlgeschlagen" next to the retry
button. The message was written to the row and surfaced solely in the ADMIN job list
-- a different screen, possibly a different person. It now travels with the status,
and only on a failure, so a message left on a since-succeeded row can't appear beside
a green "ist bereit".
Tests: 10 unit (the u128 clamp caught a real bug in the first draft -- saturating_mul
then /100 turns an overflow into a number ~100x too small, the one direction that
authorises the write being guarded against; the carried-forward archive must survive
its own older epoch in the filename), 4 DB-backed (the estimate is asserted against
the row set the archive actually contains, not against a restatement of the WHERE
clause, so the two queries cannot drift), 3 e2e over the four-hop plumbing.
──────── fix(maintenance): reclaim the media of deliberately deleted uploads
The quota stopped bounding the disk. `soft_delete_in_event` stamps `deleted_at` and
refunds `total_upload_bytes`, but nothing ever removed the bytes, and the hourly
sweep reached only `compression_status = 'failed'`. Upload 500 MB, delete, quota back
to zero, upload another 500 MB. Not an attack -- a guest curating their camera roll,
which is what people do. The host then sees guests hitting "Du hast dein Upload-Limit
erreicht" while the admin widget shows a disk full of files no upload row points at,
and the quota message is actively misleading because the space really is gone, just
not to anyone the accounting can name.
Two retention windows, because the two deletes mean different things. A compression
failure keeps its 14 days: the guest didn't ask for it and may not be able to retake
the photo. A deliberate removal gets 24 hours -- 14 days outlives the whole event, so
a deliberate delete would never reclaim anything while it mattered, and a day still
covers a mis-tap.
Wider than reported: ALL FOUR paths are reclaimed, not just the original. Preview,
display and thumbnail are each a separate file, none counted in
`original_size_bytes`, and nothing ever removed them either. That was invisible while
the sweep only saw failed compressions (which produce no derivatives) and becomes
three leaked files per upload the moment it reaches a successful one. A row is
re-selected until every path is cleared, and the columns are cleared only once every
file for that upload is gone -- clearing after a partial success would strand the
survivors in exactly the unowned state this drains.
`backfill_stale_derivatives` selects on `display_path IS NULL AND preview_path IS NOT
NULL`, which is close enough to the post-sweep state to be worth pinning: it is
guarded on `deleted_at IS NULL`, so it cannot re-decode an original that is no longer
on disk. Covered.
Residual, deliberately: within the 24h window the bytes are still spent and still
unaccounted, so delete-and-re-upload through an eight-hour event can outrun the
sweep. Bounding that means holding the quota until the file is reclaimed rather than
refunding at `deleted_at`. The low-disk warning is the net under it.
Tests: 6 DB-backed, replacing 3. The one asserting an owner-deleted upload IS
reclaimed is the exact inverse of what this file used to assert.
──────── feat(host): warn about low disk before it becomes unrecoverable
Storage visibility existed in exactly one place: a passive Speicherauslastung widget
on the ADMIN dashboard. A host who isn't the admin had no view of it, and nothing
warned anyone. README carried "Low-disk alert (< 10 GB free)" under Planned since v1.
Two things make this a safety net rather than a nice-to-have. postgres_data,
media_data and exports_data are all Docker named volumes on ONE filesystem, so
running out doesn't degrade a subsystem -- Postgres stops being able to write and the
whole event goes down. And the keepsake needs room for two gallery-sized archives,
which the export preflight can only ever refuse AFTER the release, when the event is
over and every remedy is harder.
So the threshold is not a fixed number alone. It fires on the 10 GB floor the README
always named, OR on "you could not build the keepsake right now" -- the trigger a
host can still act on, computed with the same arithmetic the preflight uses. Unknown
free space is NOT low: it fails open like the upload quota and the preflight do,
because a banner that cries wolf on an unreadable mount is a banner nobody reads.
Carried on GET /host/event, which the dashboard already fetches on load and on every
reload -- no new endpoint, no new poll. Rendered above everything else including the
PIN-reset queue, and it names the consequence (the event, not just the download)
rather than only the number.
Also fixes the host page's formatBytes, which topped out at MB: 30 GB free would have
rendered as "30720.0 MB", and a guest with 2 GB of uploads was already being shown
that way in the user list.
Tests: 5 unit on the threshold (including that plenty of free space is still low when
the keepsake wouldn't fit -- the case a fixed threshold misses entirely), 3 e2e.
The e2e drives it through `original_size_bytes` rather than a genuinely full disk:
the estimate is pure SQL over that column, so overstating one row moves the
accounting without touching a byte on disk.
──────── docs: add a restore procedure, fix the backup cadence, and correct quota_tolerance
Four things, all found by the same question: what does an operator standing at the
venue actually need?
A RESTORE PROCEDURE. There was none anywhere, and a backup you have never restored
isn't a backup. Two hazards worth writing down: media must be extracted preserving
ownership (the app runs as uid 100 / gid 101, and a root-owned restore makes every
upload fail with EACCES surfacing as a generic 500), and the app must be STOPPED
first, because migrations run on boot and a live pool will fight the restore.
Both the backup and the restore commands were run against the real stack before being
written down, which caught two that would have failed:
- The plain `pg_dump` did not restore: `psql` aborted on `ERROR: schema
"_sqlx_test" already exists`. pg_dump emits no DROPs without --clean --if-exists,
so the documented dump could only ever be restored into an empty database. Fixed
at the source (the dump is now self-cleaning) and verified end to end: 16 tables
back, exit 0.
- `--same-owner` does not exist in BusyBox tar, which is what `alpine` ships, so
the extract aborted before unpacking anything. `--numeric-owner` plus the
explicit chown, verified to land 100:101.
BACKUP CADENCE. "Weekly offsite" is the wrong shape when every irreplaceable byte is
created in one eight-hour window and nobody can retake a wedding. The backup that
matters runs that night, and again after the release so the keepsake is captured.
Also: take the DB dump and the media tarball back to back, or you get rows pointing
at files the dump doesn't know about.
quota_tolerance WAS DOCUMENTED AS SOMETHING IT ISN'T. .env.example called it "fraction
of disk that triggers the low-storage warning". It is the multiplier in
`floor(free_disk * tolerance / active_uploaders)` -- so an operator who wants "warn me
later" and sets 0.95 is actually authorising guests to fill 95% of the disk, moving
the fixed point from 43% to ~49% and eating the export headroom. The admin UI labelled
it "Toleranz (0-1)" with no explanation at all, which invites exactly that reading;
it is now "Speicher-Anteil für Gäste" with the formula in the hint. Wrong docs on a
tuning knob are worse than no docs.
SIZING. New section with the arithmetic: three volumes on one filesystem, the quota
fixed point at tolerance/(1+tolerance), and the fact the 80 GB baseline does not cover
the keepsake -- both archives are built concurrently and each is roughly a second copy
of every original. Provision ~3x expected media, or give exports its own volume.
Also ticks the low-disk alert off the roadmap, since it now exists.
──────── chore: raise the db memory limit and rate-limit social writes
Two smaller operational items.
POSTGRES 512M -> 1G. DATABASE_MAX_CONNECTIONS is 30 for a ~100-guest event (feed
polling + SSE + uploads at once), and 30 backends plus Postgres 16's default
shared_buffers leaves very little headroom at 512M. An OOM here doesn't degrade one
feature -- every request path touches the database, so it takes the event down.
Memory is the cheaper knob than shrinking the pool back and reintroducing the
queueing it was raised to fix. .env.example now names the pairing explicitly, the way
it already does for COMPRESSION_WORKER_CONCURRENCY.
SOCIAL WRITES WERE UNTHROTTLED. toggle_like, add_comment and delete_comment were the
only mutating endpoints in the app with no limit at all -- upload, join, recover,
export and admin login all carry one. Asymmetric coverage rather than a deliberate
decision.
Low severity, and honestly so: a like fans an SSE broadcast to every client, but the
export regeneration a comment deletion triggers is contained (REGEN_DEBOUNCE 20s,
workers born with their epoch, superseded ones inert). So the ceiling is 120/min --
far above anything a real guest produces. This bounds a script, not an enthusiastic
double-tapper.
ONE bucket across all three actions: separate buckets would let a caller triple the
aggregate write rate by alternating between them. Keyed per USER, matching the feed
and upload limits -- at a venue every guest is behind one NAT, and an IP key is what
made the /join and /feed limits turn guests away in the first place.
Migration 020 seeds both keys, and both are wired into the admin allowlist, the
config UI and the e2e reseed -- the step two earlier per-area toggles missed, which
left switches that existed in code and could never be flipped.
Tests: 4 e2e, including that the shared bucket really is shared (the part most likely
to be lost in a refactor) and that one guest hitting the ceiling doesn't block
another behind the same IP.
──────── fix(e2e): stop the video poster assertion racing the ffmpeg thumbnail
Pre-existing, and it fired for real during the full-suite run on a cold stack.
The lightbox binds `poster={upload.thumbnail_url ?? undefined}`, so the attribute is
absent until compression produces the thumbnail. This test asserted on it immediately
after seeding, never waiting for the worker -- unlike the Range test further down the
same file, which does poll. Against a warm stack the worker usually wins; against a
freshly rebuilt one (`stack:down -v`, cold ffmpeg) it doesn't.
That is the worst possible time for a false failure: the first run after a rebuild is
exactly when you are trying to establish whether a change broke something. Poll for
`compression_status = 'done'` before the poster assertion. The `src` assertion needs
no wait and keeps none.
Verified with --repeat-each=3.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
fabi
|
31faccfdf8 |
fix(audit-4): refuse undecodable images at admission, stop retrying permanent failures
Fourth round: reject at the door what the worker can only fail on, and stop retrying errors that cannot succeed — while keeping IoError retryable so ENOSPC still gets another attempt. Squashed from 2 commits, original messages preserved below. ──────── fix(upload): refuse undecodable images at the door, and stop retrying them Two halves of the same complaint: an oversized photo was accepted with a 201 and then silently soft-deleted minutes later, after the worker had burned six seconds of backoff re-reaching a conclusion it could not change. Admission. The compression budget now runs at upload time, against the header only, so a guest is told immediately and told why: "Bild hat zu viele Bildpunkte (ca. 99 Megapixel) und kann nicht verarbeitet werden. Bitte verkleinere es und lade es erneut hoch." instead of watching the photo vanish behind a vague "could not be processed" — which arrived only if they happened to still be on the feed with that card loaded. Nothing is stored, so there is no row to soft-delete and no orphan for the sweep to reclaim. Admission and the worker share ONE function (`decoder_within_budget`), so they cannot drift apart and start disagreeing about what is acceptable — a photo accepted at the door and rejected by the worker would be worse than either behaviour alone. The worker keeps its own check: the backfill decodes files that predate this check, and defence in depth is the whole reason the budget exists. Retries. The loop retried every failure, including ones that are a property of the input. An image over the budget, a corrupt file, an unsupported format: each fails identically on all three attempts, so the only effect was 2s + 4s of sleep and three near-identical warnings before the same outcome. `is_permanent_image_error` classifies the `ImageError` variants that cannot change between attempts — Limits, Unsupported, Decoding — and the loop gives up on those at once. `IoError` is deliberately excluded: an ENOSPC while writing a derivative is exactly the transient case the retry exists for, and misclassifying it would turn a blip back into the data loss round 1 fixed. Measured: retry log lines went from 3 per oversized upload to 0. Tests: unit tests for both sides of the classifier (a Limits error is permanent, a missing file is not) and for admission agreeing with the decoder on accept AND reject. The e2e spec is rewritten for the new contract — 400 with an actionable message, nothing stored, backend alive after a burst of four — plus a mirror asserting an ordinary photo still uploads and processes, since a budget that rejected everything would satisfy the other two. ──────── fix(upload): narrow the admission check to the memory budget only The admission check I just added rejected ANY image the decoder couldn't build — corrupt, truncated, or unsupported, not only over-budget. That broke two adversarial tests, and they were right to break. 07-adversarial/file-upload-attacks pins, deliberately, that acceptance follows the MAGIC BYTES: a payload whose first three bytes are a JPEG header is accepted regardless of what follows, because the security property under test is that the client-declared Content-Type has no influence. Both failing cases upload 1024 bytes of JPEG magic followed by zeros. Rejecting those at admission is a different, broader contract than the one asked for, and rewriting an adversarial test to match new behaviour is precisely the thing that needs justifying rather than doing quietly. So admission now checks only what it was meant to: `exceeds_decode_budget` returns true solely for `ImageError::Limits`. A corrupt file goes to the compression worker exactly as before — which handles it gracefully and, since the retry classifier in the previous commit, no longer burns backoff on it. The resource guard is the part that had to move earlier; nothing else did. Tests: the size agreement between admission and the worker is still asserted in both directions, plus a new one writing a magic-bytes-only stub and asserting admission accepts it WHILE the worker still rejects it — pinning the boundary between the two checks so a future widening fails here rather than in the adversarial suite. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
fabi
|
06ade4e158 |
fix(audit-3): restore the decode allocation guard, route /health in production
Third round. The decode allocation guard is a regression from round 1: swapping
reader.decode() for into_decoder() silently dropped the max_alloc enforcement
while keeping the comment that claimed it held.
Squashed from 3 commits, original messages preserved below.
──────── fix(imaging): restore the decode allocation guard I removed in round 1
This is a regression I introduced, not a pre-existing gap. Before 05948d8 the
compression worker used `ImageReader::decode()`, which does:
let mut decoder = Self::make_decoder(format, self.inner, limits.clone())?;
limits.reserve(decoder.total_bytes())?; // enforces max_alloc
decoder.set_limits(limits)?;
Reading the EXIF orientation tag needs `into_decoder()` instead, and that skips
the reserve entirely — the crate's own FIXME concedes `from_decoder` doesn't
compensate. Nothing else enforces `max_alloc`: the JPEG decoder's `set_limits`
only checks support and dimensions. So the 256 MiB budget has been inert since
that commit, and round 2 then propagated the weakened path into export.rs through
the shared helper, in a commit whose message claimed the helper "carries" the
decompression-bomb cap. It didn't, and the comment saying max_alloc "hard-caps
the decode allocation" was simply false.
What was left was only the per-axis cap, which permits 12000x12000 — 412 MiB
decoded, 824 MiB for the two concurrent decodes the worker runs by default,
against a 1 GiB container. Deploy-blocking right now because bumping
DERIVATIVES_REV makes the first boot after a deploy re-decode the entire gallery
two at a time: an OOM kill there restarts the container, which re-runs the
backfill. A boot loop, on the first deploy of these fixes.
Re-add the reserve exactly as `decode()` does it. Per the budget decision it stays
at 256 MiB (~89 MP for RGB8, above any mainstream phone's real output); two
concurrent decodes now peak at 512 MiB. Oversized images take the graceful path
from round 1 — original retained, quota refunded, upload-error toast — and fail
after the header parse but BEFORE any pixels are read, so they cost a header read
rather than an allocation. Measured peak during a concurrent oversized burst: 3.0
MiB.
Test parity is the other half, and the reason this was invisible: the e2e app
container had NO memory limit while production is capped at 1 GiB, so a decode
that would OOM-kill production simply succeeded in CI. Mirror the 1 GiB cap in
docker-compose.test.yml. That is the third divergence of this shape, after WebKit
missing from CI and /health existing only in Caddyfile.test.
Tests: a fixture that is 568 KiB on disk and 283 MiB decoded (11000x9000 = 99 MP,
deliberately UNDER the per-axis cap so the axis check cannot be what rejects it).
A unit test asserts the refusal — it fails against the old code, which decoded it
into an 11000x9000 buffer — with a companion asserting an ordinary photo still
decodes AND still gets its orientation applied, so the guard didn't become a
blanket refusal. An e2e test uploads it singly and as a concurrent pair, asserting
compression lands in 'failed' and the backend is still serving and still
processing afterwards.
──────── fix(deploy): route /health in production, and actually apply Caddyfile changes
Two defects in the update procedure I wrote last round, both of which make a
successful-looking deploy a lie.
1. The documented health check could never pass.
`curl -fsS https://DOMAIN/health` 404s against a perfectly healthy production
stack. The backend registers /health on its ROOT router, not under /api/v1, and
the production Caddyfile proxies only /api/* and /media/* — so /health fell
through to the SvelteKit catch-all, which has no such route and returns its 404
page. With -f, curl exits 22 and the `&& echo` never runs. My own gloss
("Anything other than ok means check the logs") then sent the operator chasing a
phantom outage.
e2e/Caddyfile.test has carried `reverse_proxy /health app:3000` since it was
written — precisely because the catch-all would otherwise swallow it. Production
never did. Per the fix-the-gap-not-the-doc call, production gets the same line,
and /health joins the no-store matcher so a cached response can't report the last
known state instead of the current one. Verified by running the production
Caddyfile against the real backend: /health -> 200 "ok", Cache-Control: no-store,
with /api/v1/event and / unaffected.
2. The sequence never reloaded Caddy, so a Caddyfile-only change was dropped.
`--build` only rebuilds services with a `build:` section, and caddy is a pinned
upstream image. Compose decides whether to recreate a container from its config
hash, which covers the mount SPECIFICATION but not the mounted file's CONTENTS —
so a git pull that changes ./Caddyfile produces no delta, Compose reports
`Running`, and Caddy serves its old config indefinitely. Exit code 0 throughout.
Round 1's iOS download fix (137c4ee) is exactly this shape: Caddyfile plus four
e2e files, so 100% of its production effect is in that one file. Following the
README to the letter deployed it, showed both image IDs changing, and left iOS
downloads broken.
Demonstrated rather than assumed — added a probe header to a Caddyfile, ran the
old sequence (`up -d --build`): header absent, change silently dropped. Ran the
new step 4 (`up -d --force-recreate caddy`): header served.
`--force-recreate` rather than `restart` or `caddy reload` because the bind mount
is resolved to an inode at container-create time and git pull replaces the file
rather than editing in place, so a restart can re-read the stale content — the
exact failure I hit in round 1 when `caddy reload` didn't pick up an edit.
Also rewrites the "db and caddy are untouched … so data volumes survive" sentence.
I wrote it as reassurance; "caddy is untouched" was the bug.
──────── chore: take the Bash(*) permission change back out of the shared settings
`.claude/settings.json` is committed and applies to anyone who clones. Fabi's
local `allow: ["Bash(*)"]` plus deny list ended up in it, inside f0d69f1 — a
commit about the image decode guard, which has nothing to do with permissions.
That was my mistake, twice. The file was already modified when I started the
round: my `git status --short` check printed "(clean)" from an unconditional
`echo` rather than from the status output, so I read a dirty tree as clean. Then
`git add -A` swept it into an unrelated commit, and I reported afterwards that I
had left it untouched. Neither the check nor the claim was true.
Restores the shared file to its previous three narrow entries. The permission
setup itself is preserved, moved to `.claude/settings.local.json`, which
`.gitignore:34` covers precisely so per-user permissions stay per-user — the
existing 442 entries there are kept alongside it.
Not rewriting f0d69f1 to erase this: main is unpushed so it would be safe, but a
visible correction is worth more than a tidy history, and a rebase across the
merge commits carries more risk than the mistake does.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
fabi
|
b601c062bd |
fix(audit-2): video playback, role identity, keepsake EXIF, recover ceiling, WebKit CI
Second round: the two regressions the first round introduced, the remaining
gaps it left, and the CI change that makes the iOS guarantees actually gate a
change rather than being asserted.
Squashed from 9 commits, original messages preserved below.
──────── docs(deploy): document the update path — `up -d` alone ships nothing
The README only ever described a fresh install. There was no update section
anywhere, and `--build` appeared nowhere in the docs.
That matters because `app` and `frontend` are `build:` services with no published
image tag, and Compose has no source-change detection: if an image by that name
exists it is reused. So the natural `git pull && docker compose up -d` reports
"Container app-1 Running", rebuilds nothing, and exits 0. A deploy that shipped
none of the new code is indistinguishable from a successful one — which is how
eleven merged fixes can sit in the repo and never reach the box.
Verified both halves against a real stack rather than asserting them: with a
source change staged, `up -d` left the image ID untouched; `up -d --build`
produced a new image ID and a healthy /health.
Adds an "Updating an existing deployment" section covering backup-before-migrate,
pull, rebuild, health check, and an image-ID comparison to prove a build actually
happened. Also spells out the rollback trap: migrations run on boot and are not
undone by checking out an older commit, so rolling back code without restoring
the snapshot leaves the schema ahead of the binary and the app refusing to start.
──────── fix(video): play the actual video, and answer Range requests
Every video in the app was unplayable. Two independent defects, either one
sufficient on its own, and nothing in the suite covered either — no test
anywhere played media or asserted a `<video>` src.
1. The lightbox handed `<video>` a JPEG.
`pickMediaUrl` is mime-agnostic, and compression only ever produces a THUMBNAIL
for a video (one `ffmpeg -vframes 1` frame) — no preview, no display. So in the
DEFAULT saver mode the element's src resolved to `/api/v1/upload/{id}/thumbnail`,
served as `image/jpeg` with `nosniff` so the browser can't even sniff its way
out. Chromium reports DEMUXER_ERROR_COULD_NOT_OPEN.
Fixed in the lightbox rather than in `pickMediaUrl`: FeedListCard shares that
helper and legitimately wants the thumbnail for its `<img>` poster, so a central
mime branch would break the feed. This mirrors the rule the diashow already
applies ("videos play the original file directly"). Added `preload="none"` so
saver-mode guests on cellular still fetch nothing until they press play — there
is no smaller video derivative to offer them — plus `playsinline`, without which
iOS hijacks playback into fullscreen.
2. `stream_media_file` ignored Range entirely.
It took no request headers, so it could not see `Range`; it always returned 200
with the whole body and never sent Accept-Ranges or Content-Range. iOS Safari
opens every `<video>` with a `Range: bytes=0-1` probe and abandons the load
without a 206 — so video failed on the app's primary platform even in `original`
mode, where the src was already correct.
Adds single-range support (`bytes=N-`, `bytes=N-M`, `bytes=-S`) with 206 +
Content-Range, 416 + `bytes */len` past EOF, and Accept-Ranges advertised on
every response. Anything it won't handle — multi-range, non-bytes units, garbage
— falls back to a full 200, which RFC 9110 explicitly permits and which is safer
than guessing. All four media routes share the helper, so seeking works
uniformly.
`get_original` now serves `inline` instead of `attachment`. An attachment
disposition is hostile to a `<video>` element, and this route is the only source
of playable video bytes; it also matches what the UI promises, since the action
is labelled "Original anzeigen" — view, not download. `no-store` is deliberately
kept so a takedown still revokes access promptly; ranges work fine under it, the
client just re-fetches.
Tests: 11 unit tests pin the parser (the iOS `bytes=0-1` probe, inclusive ends,
suffix ranges, clamping past EOF, 416 vs 200, malformed fallbacks). A new
03-feed/video-playback spec asserts the src is the original and not the
thumbnail, that the browser accepts the bytes as media (readyState > 0, no
MediaError), that no video bytes are delivered before play, and that Range
returns the correct 206 slices and a 416 past EOF — verified on both Chromium
and WebKit.
The "not downloaded before play" test asserts no *delivered body* rather than no
request: WebKit opens a connection for a preload="none" video and immediately
aborts it (GET, no Range, status 0, nothing transferred) while Chromium issues
nothing at all. The portable guarantee is that no response carrying bytes
completes.
──────── fix(auth): bind the role store to the identity, not to the tab
The role store I added in the moderation work is a module-level singleton seeded
ONCE at import. `goto()` is a client-side navigation, so leaving and re-joining in
the same tab re-imports no module and re-runs no onMount — the previous user's
role simply stayed resident. Nothing reset it: not join, recover, admin login,
"Event verlassen", `clearAuth`, nor the api.ts 401 auto-clear.
So a host who left, followed by a guest joining on the same phone, left that guest
with `isStaff === true` and a "🚫 Beitrag entfernen" action on other people's
photos. The backend 403s the delete, so this was a false affordance rather than a
privilege escalation — but `/feed` never fetched `/me/context`, so unlike every
other route it never self-corrected either. It survived until a hard reload.
The mirror case was equally broken and easier to overlook: a guest who recovered
into a host account got NO host affordances.
`clearAuth` already had a hook registry for exactly this shape of problem, with a
comment explaining it exists to avoid circular imports. Add the missing mirror,
`onSetAuth`, fired by both `setAuth` and `setAdminAuth` after the new token is
resident, and have the role store register on both sides: clear to null on
logout, re-seed from the new token on login. That also gives
`syncRoleFromToken` — dead code with zero callers since I introduced it — its
intended purpose.
Seeding from the claim fixes the reported bug, but the claim is frozen for the
token's 30-day life, so a promotion or demotion still wouldn't reach the feed.
`/feed` now calls the existing `refreshEventState()` on mount, which fetches
`/me/context` and applies both the authoritative role and the lock/release state
in one request. The feed is the one route gating a destructive action on the role,
so it should not be the only route running on a stale claim.
Tests: 04-host/role-identity-reset drives the real flows. The first asserts the
host DOES see the action before asserting the newcomer does not — a negative
assertion alone would pass against a build that shipped no moderation at all. The
second covers the mirror, promoting a guest server-side while their resident token
still claims `role: guest`, so a fix that only cleared the role would fail it.
──────── fix(compression): reclaim failed originals instead of leaking them
Round 1 stopped the compression worker deleting an upload's original on failure —
a transient ENOSPC or a codec panic must never destroy the only copy of a photo a
guest cannot retake. But it left `Upload::soft_delete`'s quota refund in place, so
the bytes stayed on disk while the uploader was charged nothing for them.
That is worse than it first looks. The row is soft-deleted, so the file is
invisible and unowned; a guest hitting a reproducible codec failure can accumulate
orphans indefinitely at zero personal cost. And `active_uploaders` counts only
users with non-deleted uploads, so dropping out of that count RAISES everyone's
per-user ceiling — the leak loosens the very quota meant to contain it.
Keep the refund: the uploader didn't cause the failure and shouldn't silently lose
quota to it. Bound the leak instead, with an hourly sweep alongside the existing
session cleanup in `spawn_periodic_tasks`, reclaiming failed originals older than
14 days — comfortably longer than any single event, so an operator investigating a
failed upload still has the file.
The selection predicate is the entire safety argument, so it is deliberately
narrow: `compression_status = 'failed'` AND soft-deleted AND past the window AND
`original_path <> ''`. That is exactly the state the give-up path leaves behind,
and it cannot reach a live upload, an owner-deleted one, or a failure still inside
its recovery window. `original_path` is cleared after a successful reclaim, which
makes the sweep idempotent — otherwise a row whose file is already gone is
re-selected on every tick forever. The row itself is kept as the audit trail.
Tests reproduce the selection verbatim (same pattern as upload_concurrency) and
assert it against five near-misses that must survive, both sides of the retention
boundary, and the idempotence property.
Also fixes two comments in export.rs still claiming "the compression worker
hard-deletes an original when its transcode fails" — no longer true, and the
defensive handling they justify is now justified by this sweep and by ordinary
deletes instead.
──────── fix(export): apply EXIF orientation in the keepsake too
Round 1 fixed EXIF orientation in the compression worker, which corrected the live
app — feed preview and diashow display. The export worker was missed, and it does
not reuse those derivatives: it re-decodes the originals itself with `image::open`,
which ignores the orientation tag, then re-encodes to JPEG, which drops the tag —
so the viewer has no way to recover it.
The damage was oddly shaped, which is exactly why it reads as a viewer bug:
Gallery.zip originals correct (byte-copied, EXIF intact)
Memories viewer grid thumbnails SIDEWAYS (always)
Memories viewer full image >5 MB SIDEWAYS (re-encoded at 2000px)
Memories viewer full image ≤5 MB correct (streamed byte-for-byte)
So in the keepsake people actually keep, every portrait photo in the grid was on
its side, and clicking through silently "fixed" small photos but not large ones.
Rather than paste the decoder dance a third time, extract `services::imaging::
decode_oriented` and route both workers through it, so there is exactly one way to
turn a file on disk into a DynamicImage. It carries a second invariant that had
also drifted: `image::open` applies NO decode limits, so the export path was
decoding arbitrary user-supplied images unbounded — the decompression-bomb cap
existed only in the compression worker. Both now come as a pair, which is the
point of having one function.
Not done: switching export to consume the existing `display` derivative. It would
fix orientation and drop a redundant full-resolution decode per photo, but it
would also replace the pristine ≤5 MB originals in the keepsake with 2048px
re-encodes — a real quality regression in the one artefact people keep forever.
Test uploads the round-1 fixture (40x20 landscape tagged Orientation=6), runs a
real export, pulls the thumbnail out of Memories.zip and asserts it came back
portrait — with a sanity check that the source really is stored landscape, so the
test can't pass against a pipeline that does nothing.
──────── fix(recover): cap name cycling, and stop bcrypt blocking the runtime
Round 1 gave /join a per-IP ceiling and left /recover with only its
`recover:{ip}:{name}` bucket. That key is right for the job it was written for —
stopping someone who knows a display name (they're listed on the feed) from
burning the victim's 3-strike PIN counter and locking them out on repeat. But the
name is ATTACKER-CHOSEN, so cycling names mints a fresh 5-attempt bucket every
time and the per-IP cost is unbounded.
What sits behind that limiter makes it worse than a normal flood: every call runs
a cost-12 bcrypt verify, including an UNCONDITIONAL throwaway verify for names
that don't exist — added deliberately to close a timing oracle. So an unknown name
is the single cheapest way to make the server do ~200ms of hashing.
Adds `recover_ip_rate_per_min` (default 30, migration 019), checked BEFORE the
per-name bucket so a name generator can't walk past it. 30/min is far above any
real recovery attempt while capping a flood. The per-name bucket is untouched and
remains the anti-guessing control.
The second half matters as much as the first: bcrypt was running inline on the
async runtime everywhere. At cost 12 that pins a tokio worker thread for ~200ms,
and there is only one per core — so a login flood stalled every other request on
the box, including the feed. There was no spawn_blocking anywhere in the auth
module, despite SECURITY-BACKLOG claiming bcrypt had been offloaded.
Route all of it through `verify_password` / `hash_password` on the blocking pool.
That covers /recover, /admin/login, the host PIN reset, and — the one most likely
to bite at a real event — the PIN hash minted on every single /join. Saturating
the blocking pool degrades logins; saturating the worker threads degrades
everything.
Tests: cycling distinct names from one IP now hits the ceiling with a Retry-After,
and — the assertion that keeps the fix honest — repeated wrong PINs against ONE
name are still throttled with the ceiling set generously high, so the ceiling
added protection rather than replacing it.
──────── ci(e2e): run WebKit, so the iOS guarantees actually gate a PR
The workflow installed only Chromium and ran chromium-desktop + chromium-mobile.
iOS Safari is the app's stated primary user — a wedding guest opening a QR link —
and WebKit is the only engine in the matrix that reproduces two of its behaviours:
- it enforces X-Frame-Options on the hidden download iframe, so a site-wide DENY
makes the keepsake download silently do nothing. Blink hands attachments to
the download manager before the frame check and never notices.
- it abandons a <video> load unless its Range probe gets a 206.
Both of those shipped. Adding 06-export to the webkit project in the round-1 fix
bought nothing on a PR, because CI never ran that project at all — the regression
test written specifically to catch the blocker only ever executed locally.
Runs 71 tests (67 pass, 4 skip on the documented IndexedDB-blob harness
limitation) in ~1.5 minutes locally, using the exact command added here.
──────── chore(backend): satisfy cargo fmt
`checks.yml` runs `cargo fmt --check`, and it has been failing since the round-1
audit fixes: I gated those on `cargo build` and `cargo clippy` but never ran fmt,
so three files drifted then and eight more this round. Pure formatting — no
behaviour change; clippy stays at zero and all 70 backend tests still pass.
Worth noting for next time: clippy passing is not evidence fmt does.
──────── chore: satisfy prettier in frontend and e2e
`checks.yml` runs `npm run format:check` for both projects and both were failing.
- frontend/src/lib/ui-store.ts is mine, unformatted since the round-1 upload-queue
badge fix — the same miss as the rustfmt one: I gated on svelte-check and eslint
but never on format:check.
- e2e/loadtest/* and e2e/shots.mjs have been unformatted since
|
||
|
fabi
|
0c0eed885a |
fix(audit-1): media gating, per-user limits, moderation UI, upload integrity
First audit round: a blocker and eight high-severity findings across the media
access gate, the guest-facing rate limiters, host moderation, and the upload
pipeline — plus the backup commands and the e2e/production divergences that hid
several of them.
Squashed from 7 commits, original messages preserved below.
──────── fix(export): let the keepsake download through X-Frame-Options on iOS
The keepsake download navigates a hidden, same-origin iframe (deliberately:
a top-level navigation to a 404/429 would unload the PWA). Caddy stamped a
site-wide `X-Frame-Options: DENY` that also covered the proxied `/api/*`.
Blink hands a `Content-Disposition: attachment` response to the download
manager at the network layer, so Chromium never noticed. WebKit enforces XFO
on the frame navigation first and aborts the load — so on iOS Safari, the
app's primary platform, tapping Download did nothing at all, silently.
Carve the two export endpoints out to SAMEORIGIN, which still blocks
cross-origin framing. Implemented as two disjoint matchers rather than an
override: Caddy applies the FIRST `header` directive outermost, so it wins on
write and a later, more specific `header` is silently ignored (verified
against the running test stack).
Also close the test gap that let this ship:
- `06-export` ran on chromium-desktop only; add it to `webkit-iphone`, the
only engine that enforces XFO on the download frame.
- No test in the suite ever clicked a download button — every archive
assertion used Node `fetch`, which has no frame and no XFO enforcement.
Add a spec that clicks it and awaits a real `download` event. Verified
falsifiable: with the blanket DENY reinstated it fails and reports the
WebKit refusal as the cause.
- Fix `ExportPage`'s card-scoped locators, which matched nothing: the cards
carry `class="card p-5"` (a Tailwind `@apply` component class), never the
`rounded-xl` the page object looked for. This had left the "shows enabled
download buttons" test red on main.
──────── fix(media): close the percent-escape bypass of the media gate
`/media/%70reviews/{id}.jpg` served a taken-down photo to anyone,
unauthenticated. Verified against the running stack: the literal path 404s,
the escaped one returned 200 with the full image. Same for displays,
thumbnails and originals, and any escaped byte in any position works.
Cause: the block was four `nest_service("/media/previews", 404)` route
matches sitting above a `ServeDir` on `/media`. axum matches on the RAW path
(matchit does no percent-decoding), while `ServeDir` percent-decodes when it
resolves the file. So `%70reviews` missed every blocker, fell through to the
ServeDir, and was decoded back to `previews/` on disk — reaching the bytes
with no soft-delete and no ban-hide check. That defeats a host takedown,
which is the entire point of the gate.
Remove the `/media` route tree outright instead of racing the decoder.
Nothing needs it: every media URL the backend emits is already a gated
`/api/v1/upload/{id}/{original,preview,display,thumbnail}` alias
(handlers::feed), the frontend contains zero `/media/` references, and the
`/media` in config.rs/disk.rs is the filesystem path while `media/` in
export.rs is a path inside the zip. `/media/**` now 404s regardless of
encoding. The route's own comment already said it "serves nothing" — it
wasn't a backstop, it was the vector.
Caddy keeps proxying /media/* deliberately: the app 404s it, and forwarding
means the e2e gating specs exercise the app's refusal exactly as production
would rather than being masked by the SvelteKit 404 page.
Extend the gating spec with the encoded variants — asserting only the literal
spelling is what let this sit undetected.
──────── fix(rate-limit): key the guest-facing limiters per user, not per IP
At a venue every guest is behind one NAT, so an IP-keyed limiter hands the
whole party a single bucket. On a fresh deploy 12 guests arriving together
meant 5 joined and 7 were turned away, with no Retry-After telling them when
to retry. `/feed` (60/min) and `/export` (3 per DAY — the fourth guest to
fetch their keepsake locked out until tomorrow) had the same defect.
`feed_delta` was already keyed per-user and its comment states the exact
rationale ("so one client can't starve others behind a shared NAT"); this
makes its siblings match.
- feed:{ip} -> feed:{user_id} (auth was already in scope)
- export:{ip} -> export:{user_id} (resolved from the download ticket's
session, which was previously looked up and discarded)
- join:{ip}: pre-auth, so there is no user to key on. Split in two — a loose
per-IP ceiling that only bounds raw volume (new `join_ip_rate_per_min`,
default 60, migration 017), plus the real 5/60s anti-spam bucket keyed
per (ip, name), mirroring the existing `recover:{ip}:{name}`.
admin_login / recover / pin_reset_req stay IP-keyed on purpose and are now
commented as such: they guard credential guessing, where a per-user or
per-name key would just hand an attacker a fresh bucket per guess.
Retry-After: the machinery existed but 7 of 8 sites called `check()` and
hard-coded `None`, so a throttled client was told to back off but never for
how long. Delete the bool `check()` wrapper entirely so `check_with_retry`
is the only entry point and the delay cannot be discarded by accident. Also
surface it for the PIN lockout, where the deadline was already known.
Fix the "unknown" fallback while here: every client_ip() caller passed that
literal, so any request without X-Forwarded-For — anything reaching the app
directly rather than through Caddy — shared ONE global bucket. Serve with
connect-info and use the peer address.
Tests: the reseed forces every limiter toggle off before each test, which is
why this whole class was invisible. Add 01-auth/rate-limit-shared-nat, which
enables them and asserts 12 guests share an IP without collision, that one
guest hammering their own name IS still throttled (so the fix re-keys rather
than removes the limit), and that feed/export buckets are per-user. Retarget
the ddos join test at the new per-IP ceiling — it asserted the defect.
Also seed `admin_login_rate_enabled` (read by the handler, seeded by no
migration and no reseed) and register `join_ip_rate_per_min` in the admin
config allowlist. Unrelated pre-existing red test fixed: 01-auth/join
asserted a "Willkommen!" heading the wedding redesign removed.
──────── feat(moderation): let a host remove a guest's photo or comment from the UI
`DELETE /host/upload/{id}` and `DELETE /host/comment/{id}` were complete on the
backend — transactional, SSE-broadcasting, audit-logged — and had zero frontend
callers. The feed context sheet offered "Löschen" only when
`target.user_id === myUserId`, so the only lever a host actually had against an
unwanted photo was banning the uploader.
That is both disproportionate and ineffective. A ban doesn't retract what was
already posted, and it makes things strictly worse for comments: the ban check
runs BEFORE the ownership check on the guest delete route, so banning an abusive
author leaves their comment on screen and permanently undeletable by them. With
no host affordance, nobody could remove it at all.
- feed: hosts/admins get "Beitrag entfernen" on other people's posts, routed to
the host endpoint (the guest route 403s anything the caller doesn't own) with
moderation-specific confirm copy. Own-post "Löschen" is unchanged.
- lightbox: same for comments, via /host/comment/{id}.
- Ban semantics are deliberately untouched (USER_JOURNEYS §10 — banned users keep
read access and cannot write). The deadlock is broken by giving the host a way
in, not by loosening the ban.
Live role (this had to come first). `getRole()` decodes the JWT claim, but the
token is never reissued — the backend slides the session row forward and treats
the DB row as authoritative. The claim is therefore frozen for the token's
lifetime: up to 30 days. A guest promoted at the party saw no Host-Dashboard and
no moderation actions until they signed out and back in, even though
`/me/context` had been returning their real role on every page load and 4 of its
6 call sites dropped the field on the floor.
Add `role-store.ts`: seeded from the claim so there's no flash of the wrong nav,
then corrected by every `/me/context` response. Point the ad-hoc `getRole()`
callers at it (account, upload, host, admin, and the new feed gate). The host and
admin dashboards now derive `myRole` reactively, so a demotion disables their
controls immediately instead of at next login.
Tests: 04-host/moderation-ui drives the real UI — host removes a guest photo and
it's gone from /feed server-side; a plain guest is offered nothing on someone
else's post (the mirror that keeps the first test honest); a promoted guest gains
the dashboard on reload while their token still carries `role: guest`; and a host
removes the comment of an already-banned guest, asserting first that the author's
own delete 403s so the deadlock is real.
──────── fix(upload): stop destroying originals, apply EXIF orientation, surface rejections
Three defects in the same pipeline, each of which loses a photo or misrepresents
one.
1. A transient error destroyed the guest's only copy.
`process`'s error arm unconditionally `remove_file`d the original. Every failure
routed there: `create_dir_all`, both derivative `save_with_format` calls (disk
full is the canonical case, and it arrives exactly when many guests upload at
once), a panic inside the image codec, or a momentary DB-pool exhaustion. The
row is only SOFT-deleted, so the bytes were the sole unrecoverable part — and
they were the part we deleted. The author already knew this was wrong next door:
`backfill_missing_display` says it "must NEVER soft-delete an upload that already
has a working preview".
Retry up to 3 times with backoff (re-checking the e2e generation guard after each
sleep), and on final failure keep the refund + soft-delete but leave the original
on disk, logging its path. A failed upload is now recoverable instead of gone.
2. Every portrait photo was stored sideways.
Phones don't rotate sensor data — they record the camera orientation in EXIF and
store the pixels as shot. `decode()` returns those raw pixels and the JPEG
re-encode writes no EXIF, so the 800px preview, the 2048px diashow display and
the keepsake were all rotated 90°, while "Original anzeigen" rendered upright
because the original keeps its tag. That asymmetry is why it reads as a viewer
bug. There was no EXIF handling anywhere in the repo and no exif crate.
Read the tag via `into_decoder()` (which carries the decode Limits through, so
the decompression-bomb cap is untouched) and apply it. Missing/malformed tags
fall back to NoTransforms — most images have none.
Existing derivatives are already baked wrong, so migration 018 adds
`derivatives_rev` and `backfill_missing_display` becomes
`backfill_stale_derivatives`: it now also picks up anything below the current rev
and regenerates it once from the original, which still carries its EXIF. Videos
are marked current in the migration — ffmpeg already honours the rotation matrix.
Bump DERIVATIVES_REV for any future change that invalidates derivatives.
3. A rejected upload vanished without a word.
`UploadQueue.svelte` — 162 lines holding the ONLY renderer of an item's error
text, the only "Erneut" retry button and the only rate-limit countdown — was
never imported anywhere, so `retryItem`, `removeItem` and `clearCompleted` were
unreachable at runtime. On a terminal rejection the store purged the blob and
wrote a clear German reason into `entry.error` "so the UI shows a clear reason".
There was no such UI. And `uploadBadgeCount` counted only pending/uploading, so
the badge decremented exactly as if the upload had succeeded.
Mount the queue on /upload, toast the reason immediately (the flow sends the user
to /feed straight after staging, so the list alone would still miss them), and
count blocked/error in the badge so a failure can't read as success.
Tests: 02-upload/exif-orientation uploads a 40x20 fixture tagged Orientation=6
and asserts both derivatives come back PORTRAIT, with a sanity check that the
source really is stored landscape. 02-upload/rejection-visible bans the uploader
between staging and sending, then asserts the toast, the queue row with the
server's reason, and that the item is still counted.
Note: 02-upload/quota's 4 failures are pre-existing and unrelated — see the next
commit.
──────── docs(backup): make the backup commands work; fix the e2e/prod divergences
Backup. Both documented commands failed on the shipped stack, and the sentence
explaining them was wrong too:
- `pg_dump $DATABASE_URL` — `DATABASE_URL` is only ever in the compose
environment, never an operator's shell, and it points at `db:5432`, which is
compose-internal DNS. The app image has no postgres client either.
- `> /media/backups/…` — `/media` is a named volume mounted inside the app
container, not a host path, and nothing ever creates a `backups` subdirectory.
- `rsync /opt/eventsnap/media/` — that path does not exist anywhere.
- "a single path to back up" — false, and dangerously so: exports were moved to
their own `exports_data` volume precisely so a keepsake (which contains every
photo in the event) can't be served off the media tree. Backing up only
`media_data` silently loses every generated keepsake.
Rewritten as three commands — db via `docker compose exec -T db pg_dump`, and one
`docker run … tar` per volume — all verified against the running stack. The
volume mounts use `/src`, not `/media`: I hit the footgun while testing this.
Docker pre-populates an EMPTY volume from the image's own directory and chowns it
to match, so `-v media_data:/media alpine` tars alpine's cdrom/floppy/usb, writes
them into the volume, and leaves it root-owned so the non-root app can no longer
write. Mounting where the image has nothing avoids all of it. Documented inline
so the next person doesn't rediscover it.
Also correct the architecture notes: `/media/*` no longer routes to the backend
(that static tree was removed as a gating bypass), and `exports_data` was missing
from the volume list — the one volume an operator most needs to know about.
e2e stack: add the `EXPORT_PATH` + `/exports` volume it was missing. The file
says "mirrors production layout"; without these, exports landed on the container's
writable layer at the default path, so export-leak and export-video wrote real
archives into ephemeral storage and the "exports live outside media" invariant
was never actually exercised.
Pre-existing red test, unrelated to the audit: all four 02-upload/quota tests
have been failing since
|
||
|
MechaCat02
|
3fb1b5d80d |
feat(comments): hide every comment mention when COMMENTS_ENABLED=false
The kill-switch previously left comment UI/text visible in several places. Sweep the whole surface so a comments-off instance shows no trace: Frontend (main app): - VirtualFeed grid tile: gate the comment button/count on $commentsEnabled (was ungated — the only feed surface still showing it). - admin stats: hide the "Kommentare" count card. - UploadSheet "Uploads geschlossen" notice, host ban-modal description, and host/admin unban confirmations: drop the "…und kommentieren" wording. - export page: drop "Kommentaren" from the keepsake description. Keepsake export (had no concept of the flag): - export.rs: thread comments_enabled into the exported data (ViewerEvent), wired through spawn_export_jobs/recover_exports and their call sites (host.rs, main.rs). - export-viewer: gate comment counts (list + grid) and the lightbox comments section; older exports without the field default to enabled (?? true). Backend still 403s comment writes when disabled (unchanged) — this is the UI half so stale clients and archives match. Verified on the running stack (COMMENTS_ENABLED=false): /event reports comments_enabled=false, a regenerated keepsake embeds "comments_enabled": false with no comment UI, and the uploads-closed notice renders "…ansehen und liken." Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
MechaCat02
|
5009590882 |
feat(diashow): guarantee all eligible photos shown + 2048px display derivative
Diashow completeness rewrite so every eligible upload is shown regardless of
bursts, disconnects, or library size:
- queue.ts: SlideQueue with live/shuffle queues, allKnown map, recentlyShown
ring; merge(dedup, live-first), remove/removeByUser (prunes recentlyShown),
knownIds for reconcile-eviction. Adds queue.test.ts (burst/completeness/race).
- diashow/+page.svelte: reconcile (full paginate + evict, pre-scan snapshot to
spare concurrent uploads) on mount/reconnect/periodic; catchUpNew paginate-
until-known for bursts with debounced maxWait; hard-cut removals; decode
timeout + candidate fallback + bounded skip so a broken image never stalls.
New ~2048px "display" derivative for big-screen sharpness, decoupled from the
data-saver preview (800px) used on phones:
- migration 016: upload.display_path + v_feed rebuilt (DROP+CREATE, not REPLACE,
to slot the column beside preview/thumbnail).
- compression: generate_image_derivatives emits preview+display (downscale-only
guard, no upscaling); backfill_missing_display regenerates on startup (safe:
logs on error, never soft-deletes).
- upload.rs/main.rs: GET /upload/{id}/display (mirrors preview auth/cache),
/media/displays direct-serve blocked.
- feed.rs + types.ts: display_url in feed/delta DTOs.
- diashow candidate chain: display -> original -> preview.
Verified on the running stack: migration applied, 10/10 existing images
backfilled (2048px cap honoured, small images not upscaled), /display serves
200, /feed returns display_url, diashow cycles.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
||
|
MechaCat02
|
44641473ea |
fix(quota): stop /me/quota leaking raw disk to guests
The per-user quota widget was shown to everyone and the /me/quota payload returned free_disk_bytes (raw server free space) and active_uploaders to any authenticated guest. Gate the widget to staff (host/admin) on the upload and account pages, and zero the server-wide telemetry fields for non-staff in the handler. Guests still get their own used/limit so enforcement stays transparent. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
MechaCat02
|
57a907eca5 |
feat(theme+comments): runtime colour theme and COMMENTS_ENABLED switch
Colour theme is configurable at runtime from two seed colours (brand + accent); neutrals stay fixed for contrast safety. Tailwind v4 var()-based tokens let a :root:root override recolour everything with no rebuild; the 50->950 ramps are derived via a color-mix ladder. Config lives in the DB config table (admin UI: Config > Farbschema, presets + custom pickers + live preview), served on the public /event endpoint with env defaults (THEME_PRESET/PRIMARY/ACCENT), propagated live via event-updated SSE, and cached in localStorage for a no-flash boot. The keepsake export mirrors the same ladder in Rust so offline archives match the event theme. COMMENTS_ENABLED (env, default true) is a boot-time kill-switch: the backend rejects new comments with 403 and the frontend hides the comment button (feed card) and panel/composer (lightbox). Existing comments stay in the DB, hidden, and return when re-enabled. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
MechaCat02
|
3c3a7d0082 |
fix(diashow): open SSE on mount; raise upload rate 10→100/hour
diashow only subscribed to SSE events but never called connectSse(), so a kiosk/projector opening /diashow directly (not via /feed) never opened the EventSource — the showcase display got a one-time /feed snapshot and no live updates, showing "Noch keine Beiträge" forever when turned on before any photos. Open the stream in onMount (idempotent) and close it in onDestroy, mirroring the feed page. Raise the default upload_rate_per_hour from 10 to 100 (migration 015, scoped to installs still on the old default so admin overrides are preserved). Guests routinely upload bursts of 10-20 photos; the old default throttled the first burst. Also update the code fallback and the test-mode reseed. Both verified end-to-end against the docker test stack. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
MechaCat02
|
3654aca18b |
feat(export): single-file offline keepsake viewer + guest download
Rebuild the guest "keepsake" (offline HTML gallery) so it renders when the extracted index.html is opened over file://. Browsers block external ES-module scripts and fetch() at origin null, so a normal multi-file SvelteKit build shows a blank window. Build the viewer as one self-contained index.html (inlined JS/CSS via vite-plugin-singlefile, standalone non-SvelteKit entry) and inject the data as window.__EXPORT_DATA__; the backend embeds the built viewer via include_dir! and writes the data global into index.html when zipping. Also surface the guest-facing download: a feed banner and the BottomNav Export tab, both shown once the host releases the export. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
fabi
|
ee554e7f38 |
style(backend): rustfmt the whole tree; gate cargo fmt --check in CI
The backend had never been run through rustfmt. Doing it in one mechanical pass (134 files) so no future functional diff is buried under formatting churn, then gating `cargo fmt --check` in checks.yml so it stays clean. Formatting only — no logic, SQL, or behaviour changed. Verified after the reformat: cargo test 56 passed, clippy --all-targets -D warnings clean, cargo fmt --check clean. This is the deferred cleanup noted when CI's Format step was first left out. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
fabi
|
db7c4459d7 |
fix(upload): editing a caption after release now regenerates the keepsake
edit_upload updated the caption/hashtags and nothing else. A caption lives in the HTML viewer keepsake (the ZIP holds media only — export.rs), so after release the downloadable viewer kept showing the OLD caption forever while the live feed showed the new one. Editing stays allowed while locked/released — like comments and likes, the lock freezes new uploads only (USER_JOURNEYS §9.3) — so the fix is to regenerate, not forbid: the edit and an invalidate_and_arm(ViewerOnly) share one transaction (same atomicity as delete_upload), then start_regen after commit. ViewerOnly carries the finished ZIP forward untouched since the media didn't change; when the gallery isn't released, invalidate_and_arm returns None and this is a no-op. Mutation-verified: without it, an edit doesn't bump the epoch and the caption test fails. Also (test isolation): the compression worker now carries a generation counter that TRUNCATE bumps. A worker queued on the concurrency semaphore when a truncate wiped media would otherwise wake in the NEXT test, fail to find its file, and broadcast upload-error/upload-deleted into that test's SSE stream — corrupting any test asserting on toasts or feed. It now abandons itself when the generation moved. No-op in production (TRUNCATE is the only caller). Export workers were already inert across a truncate (epoch guard on a fresh-UUID event). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
fabi
|
d2ad560df2 |
test(security): prove admin-login rate limit; make its toggle settable
The admin-login IP rate limit already existed (auth/handlers.rs: 5/min/IP, gated by
`rate_limits_enabled` + `admin_login_rate_enabled`, both default true in prod) — the earlier
"no rate-limit or lockout" note was wrong. What was missing was any TEST of it: the e2e
reseed sets `rate_limits_enabled=false`, so the old test observed all-401 under a disabled
limiter and "documented" the opposite of reality.
Replace that test with three that enable the limiter and prove it, mutation-verified
(deleting the check in the handler fails the first two):
- a burst of wrong passwords from one IP starts returning 429 (first is a normal 401, so
the password path ran and the limiter had budget; none is ever 200)
- once throttled, even the CORRECT password is refused — isolates the RATE LIMIT from the
password check (this is the assertion that dies if the limiter is removed)
- with `admin_login_rate_enabled=false` the same burst is NOT limited (the toggle gates it)
Two fixes this surfaced:
- `admin_login_rate_enabled` and `recover_rate_enabled` are read by their handlers but
were missing from patch_config's BOOL_KEYS allowlist — the switches existed in code and
could never be flipped. Added both.
- Test isolation: exhausting the admin-login limiter locks out the NEXT test's truncate
fixture, which must itself call admin_login before it can reset anything. An afterEach
disables the toggle via patchConfig (JWT-based, not rate-limited) so the next login is
clear; truncate then wipes the counter map. Runs on failure too.
Full desktop suite 201 passed / 1 skipped; backend 56 tests; clippy clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
||
|
fabi
|
e5201a9889 |
test(backend): DB-backed tests for the risky SQL; test isolation; clippy cleanup
All 40 backend tests were pure-function tests — not one line of SQL ran under `cargo test`,
even though the riskiest code in the repo is SQL. Add 12 DB-backed `#[sqlx::test]` tests
(fresh throwaway DB per test, real migrations) pinning the invariants that code is
load-bearing for, each mutation-verified to fail on broken code:
export_epoch.rs (8): epoch is post-increment on release (a worker born pre-increment is
inert); epoch strictly monotonic across reopen; a retired-epoch worker's writes are
no-ops; export_current is EXACTLY the invariant (8-case table); the ViewerOnly
carry-forward carries a done ZIP forward AND matches nothing when the ZIP is unfinished
(the
|
||
|
fabi
|
c229b560d8 |
fix(feed,host): two mis-tap bugs where a live re-render destroys the click target
Both are the same class as the autocomplete bug fixed in
|
||
|
fabi
|
2bef6e19ef |
fix(export): deleting a comment mid-build no longer strands the ZIP forever
Found while auditing the test suite: the ViewerOnly carry-forward (added by me in
|
||
|
fabi
|
3c683247c0 |
chore(deploy): expose the export rebuild hatch in the UI; rehearse migration 014
Three deploy-readiness gaps, none of them in the export invariant itself. 1. The escape hatch was unreachable. `POST /host/export/rebuild` was added as the fix for "a failed export is terminal", but nothing in the frontend called it — so recovery required a terminal, the API docs and a valid JWT. The host page instead told the host to reopen uploads and re-release, which is the destructive act the endpoint exists to avoid (it unlocks the gallery to every guest and retracts the release). The failed state now offers "Erneut versuchen"; a ready keepsake can be rebuilt behind a confirm, since a rebuild makes it briefly undownloadable. 2. Migration 014 had never been run against anything. It is the repo's first destructive migration and its backfill has to carry the old notion of "downloadable" across exactly — get it wrong and a released event's keepsake silently 404s, on data that cannot be re-created. backend/scripts/rehearse-014.sh proves it on a throwaway postgres, against a real pg_dump or a synthetic DB covering every pre-014 state, asserting up, down and up-again all preserve downloadability. Verified non-vacuous: retiring nothing (ELSE -1 -> ELSE e.export_epoch) fails it, naming the three keepsakes it would resurrect. It also surfaced that the down migration is an inverse UP TO DRIFT: a ready flag that disagreed with its job row comes back FALSE. That heals corruption rather than restoring it, and costs nothing (no file_path to serve), so the assertion checks what actually matters — no genuinely downloadable keepsake loses its flag — and reports the normalisation instead of failing on it. 3. No advisory scanning. cargo audit + npm audit, in .github/workflows/ because Gitea Actions scans it too and e2e.yml already resolves actions/* from GitHub. The runner is not assumed to have cargo. RUSTSEC-2023-0071 (rsa/Marvin) is ignored SPECIFICALLY, not globally: it is unreachable here (HS256 + bcrypt, Postgres only) and unpatched, so failing on it would mean a permanently red pipeline everyone learns to ignore. Tests: e2e pins that a failed keepsake recovers via rebuild while the event stays released and uploads stay locked — so a future refactor implementing rebuild as an internal reopen+re-release fails — and that rebuild cannot publish an unreleased gallery. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
fabi
|
9666d74a46 |
fix(export): close all 13 findings from the multi-agent review
A 6-lens multi-agent review (103 agents; every finding adversarially verified by three
refute-by-default skeptics) confirmed the epoch redesign is sound — the core invariant
holds — and found 13 real defects around it. All are fixed here.
The redesign's invariant was re-verified and stands: an accepted upload is always in the
keepsake, and a downloadable keepsake always reflects the most recent release.
But a WEAKER adjacent invariant did NOT hold — "a downloadable keepsake reflects the
current content" — and that is the theme of the biggest fixes.
CONTENT REMOVAL NOW ALWAYS REACHES THE KEEPSAKE
Regeneration was wired only to host deletes. Ban, unban, guest self-delete and guest
comment-delete did not trigger it — yet the export query filters `is_banned = FALSE`,
so the pipeline already agreed that content must not be there. Ban someone for abusive
content after release and every guest kept downloading an archive containing it.
All five removal paths now invalidate and rebuild. `query_comments` also gained the
banned/hidden filter it was missing entirely (the ZIP had it; the viewer did not).
Each removal is now ATOMIC with its invalidation (one tx, models made executor-generic).
Previously the delete committed in its own tx and the regeneration in a second: a dropped
handler future between them left the taken-down photo permanently downloadable, and
nothing could notice — the keepsake still looked complete and the host could no longer
even find the upload to retry.
NO MORE TAKEDOWN STORM
Every removal spawned two uncancellable full exports. A superseded worker was INERT, not
STOPPED: it still ran every ffmpeg spawn and image resize and wrote a whole archive before
discovering it had lost. Five deletes left ten workers alive, each holding a
full-gallery-sized temp, in a 1 GB container.
Now: a debounce before `claim_job` (a superseded worker fails its claim and does ZERO
work, collapsing a burst into one export), plus `update_progress` — which already ran
exactly the right liveness predicate and threw the answer away — returning it so both file
loops bail the instant they are retired. Moderating a comment no longer rebuilds the
multi-GB ZIP either: the ZIP is media-only, so it is carried forward, with prune taught to
protect any file a live row still references.
AN ESCAPE HATCH
A failed export was terminal at runtime: release_gallery refuses an already-released event,
recovery only runs at boot, and there was no retry route — the only outs were restarting the
container or reopening uploads to every guest. Added POST /host/export/rebuild. Also widened
mark_failed's guard to `status IN ('running','pending')`: when claim_job itself ERRORED, the
row was still `pending`, so the failure was never recorded and the job sat at 0% forever.
SILENT INVALIDATION
Retiring the epoch 404s the download instantly, but nothing told the clients — guests kept
seeing an enabled download button through the whole rebuild. Now broadcasts an invalidation.
And the download was a top-level <a href>: a 404 (or a 429 from the per-IP export limit that
everyone on the venue WiFi shares) NAVIGATED THE USER OUT OF THE PWA onto raw JSON. It now
targets a hidden iframe, so an error response is harmless.
ALSO
- The HTML export still had the exists()-then-open TOCTOU the ZIP path was fixed to remove,
at three sites, two with a hard `?` that failed the ENTIRE viewer. It is reachable: the
compression worker hard-deletes an original when a transcode fails and can still be running
when the gallery is released.
- Crash-orphaned .tmp files were immortal (prune deliberately skips .tmp). Swept at boot,
where it is unambiguously safe.
- export_status read the epoch in one statement and used it in the next; now one query.
- DiskCache::snapshot IGNORED its path argument on a cache hit, returning whatever filesystem
was measured last. Latent only because both callers pass media_path — it would have silently
misreported the moment anyone measured the exports volume. Now keyed by path.
- Corrected two comments that asserted safety properties the code does not have (claim_job
does NOT prevent a retired-epoch claim — retirement is enforced at READ time; and a
multi-replica reap is NOT contained, it can corrupt the keepsake). Added a rollback runbook
to migration 014, the repo's first destructive migration.
TESTS
The regression guard for the headline FOR SHARE fix was probabilistic — it could pass on
broken code if the interleaving fell the wrong way. It is now STRUCTURAL: the upload is held
open mid-body with its pre-flight check already passed, so the release provably lands inside
the exact window. Verified 5/5 FAILING (201 instead of 403) with the fix reverted, 5/5
passing with it.
Verified: backend 40 tests, svelte-check 0 errors, e2e typecheck clean,
e2e 160 passed / 1 skipped on chromium-desktop.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
||
|
fabi
|
8e906d7866 |
refactor(export): single epoch authority; fix upload-path TOCTOU losing photos
A deep investigative review of the export concurrency (four parallel adversarial
reviewers) found that three rounds of race fixes had been chasing the wrong bug, and
that the model itself was the root cause. This replaces the model and fixes the real
data loss.
THE ACTUAL BUG (upload path, not the state machine)
`upload` checked `uploads_locked_at` BEFORE streaming the body — minutes, for a large
video — then committed the row without re-checking. A release landing mid-upload let the
row commit AFTER the export snapshot: the photo appeared in the live feed and was
PERMANENTLY missing from the keepsake, with nothing to regenerate it. release_gallery's
comment claims to prevent exactly this; it never did. Every prior round fixed the DB
state machine, and the leak was upstream of it.
Fix: re-assert the lock under `SELECT ... FOR SHARE` INSIDE the commit tx. That row lock
conflicts with release_gallery's `UPDATE event`, so either the upload commits first (and
the release — hence the snapshot — is ordered after it) or the release wins and the
upload is rejected as `uploads_locked` (reversible; the client keeps the blob).
Regression test verified NON-VACUOUS: with the fix reverted it fails, reporting accepted
uploads missing from the keepsake.
THE MODEL (migration 014)
"Which generation is current?" had no single answer: it was assembled from three
separately-written pieces across two tables (`export_released_at`, a per-job
`release_seq`, and cached `export_{zip,html}_ready` flags). Keeping them in agreement
took ~10 hand-written guards; each review round found one more path that slipped through.
Now: ONE monotonic `event.export_epoch`, bumped in the same UPDATE as any change to
`export_released_at`. Each job carries a copy. Downloadable IFF
`released AND job.epoch = event.export_epoch AND status = 'done'` — readiness DERIVED,
never stored. A retired-epoch worker is INERT BY CONSTRUCTION: losing a race can no
longer corrupt state, only waste work.
Deleted: both ready-flag columns, both ready-flip blocks, `if ready { continue }`, the
reopen seq-bump, open_event's transaction, and the cross-table claim guard.
That last one was also UNSOUND: under READ COMMITTED, a blocked UPDATE re-evaluates its
WHERE against the updated target row but answers subqueries on OTHER tables from the
original snapshot — so the previous `EXISTS (SELECT ... FROM event ...)` claim guard
could admit a claim against an already-reopened event while RETURNING the post-bump seq.
Every guard is now same-row (`epoch = $n`), which EPQ re-evaluates correctly.
OTHER REAL DEFECTS FIXED
- release_gallery committed the release and THEN awaited the enqueue inside the handler
future. Axum drops that future on client disconnect → event released, uploads locked,
ZERO export_job rows, host unable to retry ("bereits freigegeben"), restart-only
recovery. Now one tx; workers spawn (detached) after commit.
- No flush/fsync before the tmp→final rename. async_zip's close() never flushes the inner
file and tokio's File drop DISCARDS write errors — so a full disk silently produced a
truncated archive that was then marked done and published, after which the last good
generation was pruned. Now flush + sync_all, which also surfaces ENOSPC.
- Download read the ready flag and the file path in TWO queries; a reopen between them
served a keepsake that should 404. Now one read through the `export_current` view.
- `if !src.exists() { continue }` then `open()?` — a TOCTOU that turned a tolerated gap
into a hard failure of the ENTIRE keepsake (unretryable while released). Now open-first,
warn, and skip.
- Recovery was flag-driven and never checked the disk: a `done` row whose archive was gone
was a permanent dead end needing manual DB surgery. Now verifies the file and re-arms.
- Temp files/dirs leaked on every error path (the leak is what fills the disk that
corrupts the next archive). Now cleaned up.
- Export archives were not event-scoped (`viewer_tmp_` already was), so a second event
would collide on paths and one event's prune would delete another's live keepsake.
- Host deletes after release never regenerated the keepsake — a photo removed on request
lived on in the archive forever. Now bumps the epoch and rebuilds without it.
- claim_job swallowed DB errors as "someone else won", stranding a job pending at 0%.
- export_status reported retired-epoch rows (a progress bar that never moves).
- The startup sweep's single-instance assumption is now documented, not hidden.
Verified: backend 40 tests, svelte-check 0 errors, e2e typecheck clean,
e2e 160 passed / 1 skipped on chromium-desktop (incl. 2 new regressions; the
upload-acceptance one confirmed to fail without the fix).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
||
|
fabi
|
9f3712894d |
fix(export): pin export workers to a live release epoch; make reopen atomic
Adversarial re-review of |
||
|
fabi
|
d643256f36 |
fix(rereview): close residual export finalize↔flip double-race at reopen
Adversarial re-review of
|