codex/bound-transcode-segments
5
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
567cdd1a15 |
feat(playback): balance transcode sessions across multiple GPUs (#425)
* feat(playback): balance transcode sessions across multiple GPUs
playback.hw_device now accepts a comma-separated render-device list (e.g.
"/dev/dri/renderD128,/dev/dri/renderD129"). Each transcode session resolves
the list to one concrete device at spawn — the present device with the
fewest active GPU sessions, ties keeping list order — and holds that device
for its whole lifetime (seek/audio restarts reuse it); the reservation
releases on session shutdown, idempotently, including early spawn-failure
paths. Software-accel sessions never reserve, so they cannot skew the
balance.
A single configured value keeps the historical pass-through contract and an
empty value still auto-detects, so existing deployments are unaffected.
PickRenderDevice is list-aware too, picking least-loaded without reserving,
which lets the non-session consumers (chapter thumbnails, download
artifacts, transcode nodes) spread load best-effort when given a list.
Motivation: hosts with two identical media GPUs (e.g. dual Arc A310)
previously pinned every session to one device while the second sat idle.
* feat(admin): GPU device picker for playback hw_device
The hw-accel detection endpoint now reports render_device_details — each
render device with a human label derived from its sysfs PCI vendor/device
ids ("Intel GPU (0x56a6)") — and the Playback settings page renders them as
per-device toggles instead of requiring a hand-typed device path. No
selection means auto (first available device); one selection pins every
session; multiple selections balance least-loaded. The stored
playback.hw_device value stays the comma-separated list, written in stable
detection order regardless of click order, and a configured-but-undetected
device stays visible so a temporarily missing GPU is not silently dropped
on save.
* fix(playback): make GPU selection and reservation atomic
Review follow-up: resolveSessionHWDevice previously selected the
least-loaded device and incremented its count in two separate critical
sections, so concurrent session starts could all pick the same device
before any reservation landed. Device presence checks now happen outside
the lock and selection + reservation share one critical section; a
concurrency test asserts an exact split across two devices for eight
simultaneous starts, which the two-step version cannot guarantee.
* refactor(playback): one typed GPU acquisition boundary, release on process exit
Replace the CSV-handling spread across resolveSessionHWDevice and
PickRenderDevice with HWDeviceSet + AcquireHWDevice in hwdevice.go: every
GPU workload resolves exactly one device immediately before spawn.
Balancing is explicitly QSV/VAAPI-only — NVENC addresses GPUs by CUDA
index/UUID, so a multi-entry list warns and uses the first entry instead
of collapsing through the path-presence filter. Sessions now release
their reservation only after ffmpeg has been reaped (shutdown waits on
done first), closing the window where a new start could pick a device
the old process still occupied. Render-device sysfs descriptions move to
gpudetect.go so the allocator file owns only selection/reservation.
* fix(downloads): prepared downloads acquire a GPU through the shared pool
PrepareFile resolves the configured hw_device list to one concrete
device via AcquireHWDevice and holds the reservation until ffmpeg exits
(Run is synchronous, so the deferred release is the process-exit
boundary). Download encodes now participate in the same active-load
accounting as streaming sessions instead of best-effort spreading.
* fix(chapterthumbs): resolve hw_device list per extraction via the shared pool
ExtractFrame acquires one concrete device from AcquireHWDevice for the
hardware attempt (released when the attempt finishes) instead of passing
the raw comma-separated value to ffmpeg as a single device path. The
service stops pre-resolving and caching a device at first use — the raw
configured value flows through and each extraction resolves it.
* fix(transcodenode): fresh starts use this node's configured hw_device
/transcode/start constructed TranscodeOpts with an empty HWDevice, so
fresh sessions auto-detected the first GPU and bypassed the configured
list while reconstructed sessions honored it. Both paths now feed the
node-local config value into StartTranscode's shared resolution.
* feat(admin): node-aware GPU inventory on /admin/system/hw-accel
playback.hw_device is one cluster-wide value consumed by every transcode
node, but the endpoint probed only whichever healthy node had the fewest
jobs — an admin could configure devices that don't exist on the other
nodes. The endpoint now probes every healthy node concurrently and
returns a nodes array (URL, name, resolved accel, devices, or probe
error) alongside the backward-compatible flat fields, and the config doc
states the homogeneous-path contract explicitly.
* feat(admin): GPU picker survives empty detection, warns on node divergence
The picker rows are now the union of detected devices and configured
entries, so configured-but-missing devices stay visible (and
deselectable) when detection returns nothing or an older node omits
render_device_details (plain render_devices paths fall back to a generic
label). Per-node inventories from the hw-accel endpoint drive two
warnings: a banner when responding nodes report different device sets,
and a per-row note listing nodes missing that device. The multi-select
is hidden for NVENC — balancing is QSV/VA-API only — with a notice when
a multi-device value is already stored.
* style: gofmt touched files
* fix(playback): release GPU reservations on process exit
---------
Co-authored-by: rxwatcher <rxwatcher@users.noreply.github.com>
Co-authored-by: Quick <31828688+Quick104@users.noreply.github.com>
|
||
|
|
854d07cf8f |
feat(playback): add protocol v3 planning and recovery (#398)
* docs(playback): plan protocol v3 server implementation * docs(playback): incorporate protocol v3 review * feat(playback): implement protocol v3 server * fix(playback): persist empty route diagnostics * feat(playback): harden protocol v3 HDR routing * feat(playback): complete protocol v3 client contract * fix(playback): harden protocol v3 recovery * fix(playback): restore dovi_rpu strip filter for DV remuxes The v3 work renamed the Dolby Vision strip recipe to a dovi_split=mode=bl bitstream filter that does not exist in stock FFmpeg or jellyfin-ffmpeg; the probe failed closed on every deployment, disabling the new validated DV7-to-HDR10 route and regressing the previously working dovi_rpu=strip=1 remux path from main. Restore dovi_rpu across the probe, remux and HLS copy arguments, and the recipe-card constant. Also from review: validate the remux DV mode for every profile (garbage modes on non-P7 sources silently no-opped), reject preserve mode for P7 outright (a base-layer-only remux cannot preserve dual-layer DV), tag dvhe sample entries only for the explicit v3 preserve recipe so legacy web/jellycompat remuxes keep their pre-v3 hev1 labeling, and honor the token-frozen DV mode in the proxy remux path instead of legacy-auto. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(playback): correct v3 planner policy and contract validation Review fixes to the v3 planner and wire contracts: - Bar Profile 7 sources from the non-strip progressive remux route: a base-layer-only remux can never deliver native dual-layer DV, so the planner no longer emits plans claiming validated Dolby Vision while the executed remux drops the enhancement layer. - Accept the device-quirks feature flag from either capability location, matching every other dual-location feature check. - Treat legacy hdr_unknown rows as HDR10 for HDR10-capable clients with a degradation warning instead of leaving them unplayable under v3. - Honor bandwidth_cap_kbps as a hard ceiling in every quality mode and wire the previously dead Metered signal into conservative auto rungs. - Degrade to the validated source-quality route instead of a terminal when only an implicit quality reduction demanded an unsupported transcode; explicit user-selected rungs keep terminal behavior. - Bound inner capability lists and strings; compare attempt keys exactly instead of case-folded; make ParseTrackIDV3 strict about canonical numerics; accept dvdsub/pgssub/dvbsub aliases and stop promising burn-in for unknown subtitle codecs; probe every h264 encoder rather than requiring libx264; normalize the file-level bitrate fallback. - Evaluate subtitle renderability against the engine each candidate route executes on, not always media3_direct. - Pin the with-quirks attempt-key preimage arity in the cross-language fixture so the Kotlin client stays in lockstep. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(playback): harden v3 control-plane reliability Review fixes to the v3 session, store, and handler layer: - Bound concurrent replans with a slot semaphore: each replan pins a pooled connection for its advisory lock while issuing further store queries from the same pool, so an unbounded recovery storm could turn every connection into a lock holder and deadlock the server. - Make CompleteReplan a real compare-and-swap (base-revision predicate, ErrReplanSupersededV3) and map BeginReplan insert races to a replay instead of a raw unique violation. - Fingerprint start requests (request_digest column): an attempt ID reused with different input is now a 409-style conflict rather than a silent replay, and both replay paths check session liveness so dead sessions surface as retryable terminals. - Pre-delete expired attempt rows on SaveAttempt so a retry during the cleanup window cannot wedge on an unreachable conflict. - Align the in-memory store's semantics with Postgres and add DB-backed planstore tests (SILO_TEST_DATABASE_URL), including a regression test inserting every route-event name against the real CHECK constraint. - Session manager: v3 route-set updates own RemuxDVMode outright so a replan onto an SDR source clears a stale strip mode; replacement reservations survive unrelated legacy stream updates; replacement admission excludes the replaced session explicitly instead of decrementing totals it may no longer be part of; the admission CAS loop is bounded and decider errors are logged. - Map transient store failures to 500s instead of terminal 404/403s; authorize route events via identity-only projections after the rate limiter; keep sanitized diagnostics deterministic. - Merge the server-computed durable plan key into replan exclusions so unreproducible client history cannot re-select the failed route. - Remap tracks only when the effective edition changes (a same-file replan no longer switches audio to a lookalike track) and remap ID-only subtitle selections on edition fallback. - Cache the v3/shadow feature flags for five seconds instead of one settings SELECT per playback request; stop remote transports best-effort when the start call times out; carry dvm/tid claims and the transport-scoped job identity through the legacy audio-change re-mint; index playback_route_events(received_at) for the retention delete; run store maintenance for DB-less deployments too. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(transcode): reap idle node jobs and gate WebVTT conversion - Add an idle reaper to the transcode node: a job untouched by manifest or segment requests for ten minutes is closed and unregistered. After a v3 replan retires a transport ID, a stale in-flight stream token could resurrect the old job via reconstruct and encode to end-of-file for nobody; jobs waiting on readiness count registration as access and are never reaped mid-wait, and reaping keeps the recipe so a still-valid token reconstructs on the next hit. - Reject bitmap subtitle tracks (PGS) on the .vtt conversion path with 415 before headers are written instead of spawning an ffmpeg command that always fails mid-response, and make the extract-format override fall back to source-driven mapping for bitmap codecs. - Drain error bodies on non-202 node responses so the HTTP transport can reuse connections. - Pin the transcode-dir cleanup separator-boundary semantics with a regression test (a session ID sharing another's prefix must not retain foreign directories). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(playback): close v3 planner policy gaps from review - Clamp the final transcode bitrate to bandwidth_cap_kbps: the ladder has no rung below 480p/1500kbps, so lower caps were silently exceeded even though the cap is documented as a hard delivery ceiling. - Treat video-only media as audio-compatible instead of forcing an AAC conversion (or an audio_conversion_unsupported terminal) onto a file with no audio stream. Tracks whose codec failed to probe keep the gate. - Only promise a bitmap subtitle sidecar for embedded PGS with an engine that renders embedded bitmap: external/downloaded bitmap and embedded DVD/DVB published artifact URLs that always failed at fetch. They now fall through to burn-in or its terminal. - Accept client_video_transformations_v1 from either client_features or the nested context when validating client-executor transformations, matching the planner's dual-source reads. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(playback): probe and execute DV remuxes with one ffmpeg binary The v3 transformation registry probed the configured playback.ffmpeg_path while progressive remux execution resolved the process-global discovery path, so a deployment where only one binary carries dovi_rpu could plan a server_dv7_to_hdr10 route and then fail it at stream time. Resolution now goes through a shared ResolveFFmpegPath (configured path first, discovery fallback — the same rule the transcode pipeline already used), the dovi_rpu probe is cached per binary path, and the stream handler and proxy worker pass their configured path into ServeRemuxWithDVMode. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(playback): harden v3 replan identity and control-plane limits - Seed failure-replan track selections from the durable current plan before overlaying the request: after an alternate-version fallback the normalized request still carries requested-edition track IDs, so a replan omitting unchanged tracks was rejected as a track/file mismatch. - Remap ID-only audio selections across edition changes (parse the ID to an index like the subtitle remap already does) instead of leaving a stale file-bound ID to fail validation. - Release the node planner reservation when a prepared remote transport rolls back after the node accepted the job; repeated failed starts could otherwise pin max-job/bandwidth budgets for the full reservation age. - Size the replan semaphore below the PostgreSQL pool via a store capacity advisor: with max_connections at or below the fixed bound, advisory-lock holders could starve the inner store queries they need to finish. - Contain shadow-planner panics with a recover boundary; it runs on a bare goroutine where an escaped panic kills the process for what is telemetry-only work. Document why the memory store's session lock is deliberately a no-op. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(transcode): serialize node job teardown against reconstructs - Look up and touch manifest/segment sessions in one critical section so the idle reaper cannot unregister a job between the lookup and its liveness refresh. - Re-validate each reap candidate under the per-session lifecycle lock before closing it: Close removes the output directory, and without the lock it could race a token reconstruct and wipe the segments the fresh ffmpeg is writing. - Take the lifecycle lock in handleStop so a stop racing a RequireReady start's readiness wait blocks until registration and tears the job down, instead of 404ing and orphaning the ffmpeg until the reaper. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
|
203a18ae83 |
feat(observability): OpenTelemetry logs+traces with secret redaction and slog standardization (#290)
* feat(observability): OpenTelemetry logs+traces with secret redaction Part of #265. Adds opt-in OpenTelemetry (logs + traces) alongside the existing stderr + opslog pipeline, plus secret redaction on all sinks. Default-off: with no OTEL_* / SILO_OTEL_ENABLED config, behavior is unchanged. Bootstrap (internal/telemetry): - Setup() builds one shared resource, a TracerProvider (parent-based trace-id ratio sampler), a LoggerProvider, and the W3C TraceContext+Baggage propagator from env. It installs NO MeterProvider — metrics stay on Prometheus, and the built-in no-op global MeterProvider keeps the trace instrumentation libs from double-emitting. Shutdown is deferred with a flush timeout. - Logs are bridged via otelslog fan-out (slog.MultiHandler), level-gated by the shared LevelVar and best-effort so a failing collector can't break the console or DB branches. stderr + opslog stay untouched. Secret redaction (internal/logredact): - A slog.Handler masks secret-keyed attributes (password, token, api_key, authorization, cookie, ...) — including .With-bound attrs, nested groups, secret-keyed group subtrees, and values behind a LogValuer — on the console and OTLP sinks, with a no-op fast path when a record has no secret keys. opslog.shouldRedact delegates to logredact.SecretKey so all sinks share one marker list. Rotation is infra-managed (no custom file sink): container runtime for stderr, collector/backend for OTLP, opslog partition-pruning for the DB. Documented in docs/architecture/observability.md. Verification: go build ./..., go vet, gofmt -l — clean; go test ./internal/telemetry/ ./internal/logredact/ -race pass. AI-use disclosure: implemented with AI assistance (Claude Code), including adversarial reviews that hardened the bootstrap and fixed two redaction leak paths; reviewed by the author. * refactor(observability): slog context+component sweep, sloglint gate (phase 3) Part of #265. Builds on the OTel bootstrap + redaction commit. Standardizes every log call site onto the context-carrying slog variants so records correlate with the active OpenTelemetry trace, and locks the standard in with a machine gate so future code (human- or AI-authored) can't drift back. - Call-site sweep: converted the remaining slog.<Level>(...) calls to the slog.<Level>Context(ctx, ...) form wherever a context.Context is in scope (background/init calls with no ctx are left as-is), across 183 files. Applied via a type-aware AST codemod. Log levels and message strings are preserved verbatim; a component attr (canonical per-package name) is added to direct package-level slog calls. Bound-logger calls keep their existing .With bindings. The main.go and telemetry package conversions rode with their file in the previous commit to keep each file within a single commit. - Enforcement (.golangci.yml): enable sloglint with context=scope, static-msg, key-naming-case=snake, no-mixed-args. After the sweep all four report zero violations repo-wide (tests included), so make lint / CI now blocks any regression to the non-context form. The gate ships with the sweep because it cannot be green until the legacy sites are converted. Metrics remain on Prometheus; no behavior change to /metrics or Grafana. Verification: go build ./..., go vet ./..., gofmt -l — clean; sloglint (all 4 rules) 0 violations repo-wide; log levels verified unchanged. AI-use disclosure: implemented with AI assistance (Claude Code), including the codemod; reviewed by the author. * fix(observability): honor per-signal OTLP protocol and secret WithGroup names Two Codex review findings on PR #290: - telemetry: OTEL_EXPORTER_OTLP_{TRACES,LOGS}_PROTOCOL now override the generic OTEL_EXPORTER_OTLP_PROTOCOL per signal, so mixed collector setups (e.g. HTTP logs + gRPC traces) build the right exporter. - logredact: entering a group whose name is secret-bearing (e.g. WithGroup("authorization")) now masks every leaf in that subtree, matching how slog.Group("authorization", ...) is masked as a whole. * fix(observability): address review feedback on telemetry bootstrap - Telemetry setup failure no longer kills boot: Setup returns usable no-op providers alongside the error and main logs and continues with telemetry disabled, honoring the best-effort contract. - Honor OTEL_TRACES_SAMPLER (always_on/off, traceidratio, parentbased_* variants); unsupported values fall back to parentbased_traceidratio. - Attach node identity as semconv service.instance.id instead of the non-semconv node.name. - Rename opslog retention-scope log attrs to target_component/target_level so they no longer collide with the canonical component routing key, and tag those lines with component=opslog. - Fix stale levelGated comment casing; use WarnContext in the telemetry shutdown defer; document the LogValuer double-resolve on the redaction slow path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Quick <31828688+Quick104@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
|
978e1b4954 |
feat(playback): NVENC support for transcoding (#79)
* feat(playback): NVENC support for transcoding * fix(playback): probe nvenc before auto-selecting * fix(playback): use safe NVENC smoke probe dimensions --------- Co-authored-by: Quick <31828688+Quick104@users.noreply.github.com> |
||
|
Silo Server Migration
|
c085b12fd1 | Initial Silo migration |