What the Alteox Media Server does

• In-house MPEG-TS demuxer and muxer; every parser is fuzz-tested and returns errors instead of failing.
• Every channel component runs with panic recovery: it is restarted, logged and counted — the other channels keep running.
• A slow viewer, a bad input or a crashing transcoder affects only itself: publishers never block, lagging consumers are dropped.
• The transcoder runs as a separate, supervised process with restart backoff and an output-stall watchdog.
• Subtitles, teletext and SCTE-35 are re-muxed by the Go core, never by the transcoder — their oddities cannot crash it.
• Tested against a corpus of damaged streams: dropped packets, CC errors, truncated packets, garbage bytes, PMT/PID changes, PCR/PTS jumps and 33-bit wraps.

• SRT caller (pull) and listener (push), on a shared listener with stream ids or on dedicated per-channel ports with their own passphrase.
• UDP unicast and multicast including source-specific multicast, raw TS or RTP-wrapped TS.
• HTTP(S) MPEG-TS pull, and HLS pull with TS and fMP4 segments.
• Custom request headers and User-Agent for HTTP and HLS inputs.
• copy:// takes another channel’s output in-process, without copying data.

• SRT on a shared listener (stream id read:<channel>), on dedicated per-channel play ports, or pushed as SRT caller.
• UDP MPEG-TS unicast/multicast with configurable TTL and interface, paced at constant rate by PCR.
• HTTP MPEG-TS at /<channel>/mpegts.
• HLS with CMAF/fMP4 segments (optionally TS segments) and a multivariant playlist for ABR; DASH from the same CMAF segments.
• New SRT and HTTP-TS viewers start at a keyframe (fast start).

• SRT: streamid=read:<channel>/<rendition> or read:<channel>/source; a dedicated play port can be fixed to one rendition.
• HTTP-TS: /<channel>/<rendition>/mpegts and /<channel>/source/mpegts.
• The auth hook sees the requested rendition; the channel page lists every per-rendition URL with a copy button.

• Frames never leave the GPU: NVDEC → CUDA deinterlacer → split → scale_cuda → NVENC.
• All renditions come from one decode; every audio track is decoded and encoded once, not once per rendition.
• Dynamic filters: the deinterlacer runs only for interlaced sources, letterbox/blur backgrounds only when the aspect ratio differs; a source change re-plans the filters.
• GPU admission control: nvidia-smi monitoring, a capacity model per GPU model, least-loaded GPU selection, NVENC session limits, reject/queue/CPU fallback when full.
• Also Intel Quick Sync, VAAPI and CPU encoding.

Measured capacity: NVIDIA RTX 4000 SFF Ada Generation

3 renditions: 720p, 576p, 360p (H.264 NVENC) + AAC audio; measured next to another media server on the same GPU, which itself used about 22–35 % of NVENC/NVDEC.

Utilisation at the end of each run, including the other media server on the same GPU. Ramp: one channel added about every 90 s, stop when a transcoder falls behind real time or NVENC/NVDEC reach 85 %. Your numbers depend on sources, ladder and filters — calibrate on your hardware.

• An input fails after no valid video for loss_timeout (default 2 s); the channel switches to the best healthy input, on a keyframe when possible, and returns to the primary after return_after (default 30 s).
• Timestamps stay monotonic and PIDs stable across switches; players see one continuous stream.
• Per-input media analysis (codec, resolution, scan, frame rate, tracks, languages) and a 5-minute stability window of errored seconds.
• select: quality ranks inputs by complete tracks, error rate and resolution class; max_connected limits how many sources are pulled at once, with probing and exponential cooldown.
• Manual switch to any input from the UI or API.

• on_demand: inputs and transcoder start with the first viewer and sleep after on_demand_idle (default 30 s) without viewers.
• transcode_on_demand: the source keeps running for UDP/TS customers; the transcoder and its GPU only run while a transcoded output is watched.
• Waiting viewers are held until data flows (SRT, HTTP-TS) or the first segments exist (HLS/DASH). Sleeping channels raise no alarms.

• Templates hold outputs, failover, transcoding, track mapping, access control, labels and on-demand settings.
• Changing a template restarts exactly the channels whose effective settings changed.
• The channel editor shows inherited fields with Override / Reset to template.

• A request for an unknown channel name asks the lookup server; the answer becomes an on-demand channel.
• Works on SRT, HTTP-TS and HLS/DASH entry requests; names may contain "/".
• Refreshed periodically; a failing lookup server never stops a running channel.

• Importer for flussonic.conf: CLI, REST API (dry run by default) and web UI, with a report of everything mapped differently.
• auth.backend_protocol: flussonic — existing Flussonic auth backends work unchanged, including multiple backends and X-AuthDuration re-checks.
• Flussonic-style event sinks (play_started, play_closed, source events).
• A Flussonic HTTP API v3 subset under /streamer/api/v3, so control planes can monitor the server.

• Dashboard with card and dense list views for many channels, bulk actions by label, live updates over server-sent events.
• Channel editor with a visual track mapper, transcode profile editor, sessions view with kick and block.
• REST API with API keys or users (admin / viewer roles), JSON Merge Patch, If-Match concurrency, dry runs.
• Prometheus metrics at /metrics; health endpoints for load balancers and a Nagios-format health check.

• In-memory history since start: 1 hour at 10 s, 24 hours at 1 minute, 7 days at 10 minutes.
• Errors per channel and per source (an input URL shared by several channels), CPU, memory and per-GPU utilisation, NVENC, NVDEC, temperature and power.
• ETSI TR 101 290 priority 1–3 indicators per input, with a channel drill-down.

• Cold standby: surviving nodes take over a failed node’s channels.
• Hot standby per channel: a second node runs inputs, transcoder and packager with muted UDP outputs and identical HLS/DASH segment numbers.
• Measured in loopback end-to-end tests: cold takeover about 4 s, hot standby 1.9–2.5 s to serving.
• GPU-aware placement, drain for maintenance, fencing of nodes that lose the leader, TLS 1.3 between nodes.

• Flap suppression, batching, rate limits, retries and bounded queues: raising an alarm never blocks.
• Alarms in the UI, the API, server-sent events and Prometheus metrics.
• In a cluster, cluster-wide alarms come from the leader only.

• Selectors with fallbacks (e.g. German audio, else the first one), fixed bindings and required tracks with an alarm when missing.
• Output PIDs depend only on the configuration, so an input switch or a broadcaster PID change causes no PID churn.
• One source track can feed several outputs, e.g. AC-3 copied plus the same track as AAC.

• Results are cached; on hook failure the default is deny (configurable).
• SRT is authorised before the handshake completes; HLS/DASH sessions are tracked across playlist and segment requests.
• Kick sessions by id, user, token or IP, optionally with a temporary block.