neiam/dms: 07a12d4

modified Cargo.lock

+11 −0

@@ -4375,6 +4375,16 @@ version = "1.3.0"

4375	4375	source = "registry+https://github.com/rust-lang/crates.io-index"
4376	4376	checksum = "0fda2ff0d084019ba4d7c6f371c95d8fd75ce3524c3cb8fb653a3023f6323e64"
4377	4377
	4378	+[[package]]
	4379	+name = "signal-hook-registry"
	4380	+version = "1.4.8"
	4381	+source = "registry+https://github.com/rust-lang/crates.io-index"
	4382	+checksum = "c4db69cba1110affc0e9f7bcd48bbf87b3f4fc7c61fc9155afd4c469eb3d6c1b"
	4383	+dependencies = [
	4384	+ "errno",
	4385	+ "libc",
	4386	+]
	4387	+
4378	4388	[[package]]
4379	4389	name = "signature"
4380	4390	version = "2.2.0"

@@ -4887,6 +4897,7 @@ dependencies = [

4887	4897	"libc",
4888	4898	"mio",
4889	4899	"pin-project-lite",
	4900	+ "signal-hook-registry",
4890	4901	"slab",
4891	4902	"socket2 0.6.0",
4892	4903	"tokio-macros",

modified Cargo.toml

+1 −1

@@ -19,7 +19,7 @@ path = "src/lib.rs"

19	19	[dependencies]
20	20	axum = { version = "0.8.5", features = ["multipart"] }
21	21	log = "0.4.28"
22		−tokio = { version = "1.47.1", features = ["rt-multi-thread", "macros", "sync"] }
	22	+tokio = { version = "1.47.1", features = ["rt-multi-thread", "macros", "sync", "process", "time"] }
23	23	tower-http = { version = "0.6.6", features = ["fs", "trace"] }
24	24	tracing-subscriber = { version = "0.3.19", features = ["env-filter"] }
25	25	anyhow = "1.0.100"

modified Dockerfile

+4 −1

@@ -77,8 +77,11 @@ RUN curl -fL --retry 3 -o text-detection.rten https://ocrs-models.s3-accelerat

77	77	# Stage 4 — runtime: minimal Debian slim with the binaries + assets + templates.
78	78	# ---------------------------------------------------------------------------
79	79	FROM debian:bookworm-slim AS runtime
	80	+# ffmpeg is here for the `poster` worker, which lifts a still frame out of a
	81	+# video so the image-only categorizers have something to read. It is the
	82	+# heaviest thing in this layer; drop it only if video uploads go away.
80	83	RUN apt-get update \
81		− && apt-get install -y --no-install-recommends ca-certificates libssl3 \
	84	+ && apt-get install -y --no-install-recommends ca-certificates libssl3 ffmpeg \
82	85	&& rm -rf /var/lib/apt/lists/* \
83	86	&& groupadd --system --gid 10001 dms \
84	87	&& useradd --system --uid 10001 --gid dms --home /app --shell /usr/sbin/nologin dms \

modified Justfile

+1 −1

@@ -44,7 +44,7 @@ migrate:

44	44	serve:
45	45	cargo run --bin webservice
46	46
47		−# Run a categorizer worker. Usage: just worker thumbnail \| ocr \| describe \| extract_tags
	47	+# Run a categorizer worker. Usage: just worker poster \| thumbnail \| ocr \| describe \| extract_tags
48	48	worker kind:
49	49	cargo run --bin queue-worker -- --kind {{kind}}
50	50

modified README.md

+29 −3

@@ -111,9 +111,11 @@ All env-driven — defaults make sense for local dev.

111	111	\| `AUTO_MIGRATE` \| `true` \| Run pending migrations on webservice boot. Disable in environments that prefer a separate migration step. \|
112	112	\| `REGISTRATION_ENABLED` \| `false` \| When off, plain `/register` 404s; `/register?token=…` invite links still work. \|
113	113	\| `DMS_PUBLIC_URL` \| (unset) \| Override what the QR pairing URL encodes when behind a proxy. Falls back to the request's Host header, then the bind address. \|
114		−\| `WORKER_KIND` \| (required for workers) \| `thumbnail` / `ocr` / `describe` / `extract_tags`. Same as `--kind`. \|
	114	+\| `WORKER_KIND` \| (required for workers) \| `poster` / `thumbnail` / `ocr` / `describe` / `extract_tags`. Same as `--kind`. \|
115	115	\| `WORKER_PREFETCH` \| `4` \| Per-worker concurrency (RabbitMQ basic_qos). Tune CPU-bound (OCR) lower than I/O-bound. \|
116	116	\| `WORKER_MAX_ATTEMPTS` \| `100` \| Hard cap on per-(upload, kind) DB-tracked attempts. Once exceeded, the worker DLQs the next redelivery instead of running it — protects the queue from a single poisoned message hogging consumers. Set `0` to disable. \|
	117	+\| `FFMPEG_BIN` \| `ffmpeg` \| Binary the `poster` worker shells out to for video frame extraction. \|
	118	+\| `POSTER_TIMEOUT_SECS` \| `120` \| Hard cap on one frame extraction, so a video that hangs the decoder fails the job rather than parking a worker. \|
117	119	\| `OCRS_DETECTION_MODEL` \| `/opt/dms/models/text-detection.rten` \| Bundled in the image; override to swap models. Required for OCR — outside Docker (e.g. `cargo run`), set this to a downloaded `.rten` file or the worker DLQs jobs with a clear log line. \|
118	120	\| `OCRS_RECOGNITION_MODEL` \| `/opt/dms/models/text-recognition.rten` \| Same as above. \|
119	121

@@ -184,8 +186,9 @@ hosts, which would put those gigabytes back in RAM.

184	186
185	187	Per entry:
186	188
187		−- Only entries whose name resolves to an image type are stored. Everything
188		− else is counted and skipped.
	189	+- Only entries whose name resolves to an image or video type are stored.
	190	+ Everything else, such as an export's stray `.json` or `.txt`, is counted and
	191	+ skipped rather than becoming an upload nobody asked for.
189	192	- Directories, `__MACOSX/` resource forks, and dotfiles are ignored. Nested
190	193	paths flatten to the basename — storage is content-addressed, so the
191	194	archive's layout doesn't survive.

@@ -228,6 +231,29 @@ a namespace wildcard. Bare values match across whichever sources are enabled.

228	231	The whole pipeline is covered by unit tests for the parser plus integration tests
229	232	that exercise the SQL against a real Postgres (`tests/search_integration.rs`).
230	233
	234	+## Video
	235	+
	236	+A video upload cannot go straight through the categorizers: thumbnail, OCR and
	237	+describe all work on a decoded still, and none of them can open an `.mp4`. So a
	238	+video gets one job first, `poster`, which shells out to ffmpeg for a single
	239	+representative frame, stores it as `<stem>.poster.jpg` beside the original, and
	240	+then enqueues thumbnail/ocr/describe itself. From their point of view the video
	241	+has become an image; `workers::source_image_relative` is what quietly hands them
	242	+the frame instead of the original.
	243	+
	244	+ffmpeg's `thumbnail` filter picks the frame, scoring a window rather than taking
	245	+frame 0, which on a GIF-turned-mp4 is routinely a blank fade-in. Captions burned
	246	+into the picture survive, so OCR on an imgur-style export finds real text.
	247	+
	248	+Run the worker like any other: `just worker poster`. It needs ffmpeg on PATH
	249	+(`FFMPEG_BIN` overrides the binary, `POSTER_TIMEOUT_SECS` the 120s cap); the
	250	+runtime image installs it.
	251	+
	252	+Every kind that exists must appear in `queue::ALL_KINDS`. That list drives the
	253	+queue topology, and a kind missing from it has no queue, so anything published
	254	+for it is dropped by the broker without an error. A unit test keeps the list
	255	+exhaustive.
	256	+
231	257	## Pipelines
232	258
233	259	Every upload row keeps a `processing_status`. Per-upload, per-kind state

modified assets/f/js/src.js

+16 −0

Click to load diff…

modified compose.yml

+7 −0

Click to load diff…

modified src/bin/queue-worker.rs

+3 −2

Click to load diff…

modified src/bin/webservice.rs

+9 −0

Click to load diff…

modified src/config.rs

+8 −0

Click to load diff…

modified src/handlers/failed.rs

+1 −7

Click to load diff…

modified src/handlers/uploads.rs

+46 −26

Click to load diff…

modified src/queue/messages.rs

+60 −0

Click to load diff…

modified src/queue/mod.rs

+3 −8

Click to load diff…

modified src/queue/publisher.rs

+1 −8

Click to load diff…

modified src/storage/local.rs

+15 −10

Click to load diff…

modified src/storage/mod.rs

+29 −1

Click to load diff…

modified src/telemetry.rs

+1 −1

Click to load diff…

modified src/workers/describe.rs

+11 −3

Click to load diff…

modified src/workers/mod.rs

+22 −1

Click to load diff…

modified src/workers/ocr.rs

+1 −1

Click to load diff…

added src/workers/poster.rs

+171 −0

Click to load diff…

modified src/workers/thumbnail.rs

+20 −2

Click to load diff…

modified templates/search.html

+6 −0

Click to load diff…

modified templates/uploads/list.html

+6 −0

Click to load diff…

modified templates/uploads/show.html

+7 −0

Click to load diff…

neiam /dms

Ingest video, and run it through the pipeline via an extracted frame

Files changed