Media Glossary · User-Generated Platforms

ABR (adaptive bitrate) streaming: a player switching between a video's renditions as the viewer's bandwidth drops or recovers, which is why every video is encoded into a ladder of rungs (Video on Demand).
Ad break: a cue in a live stream that inserts ads, recorded with its requester, requested duration, actual duration and the vendor's reference so creator ad revenue, advertiser billing and disputes reconcile against it (Live Latency, Playback & VODs).
Alt text: a description of an image for screen readers, stored on the post's use of the file (content_media) rather than on the file, because one image can illustrate two posts that need different descriptions (Images at Scale).
Anonymous upload: a file with no owner account (a null owner_account_id), which nobody can manage or be notified about, and so is the first candidate for a retention purge (Storage, Dedup & Delivery Economics).
Archive tier: storage classes such as S3 Glacier Flexible Retrieval and Deep Archive that retrieve in minutes to hours, with 90 and 180 day minimums, for evidence and originals that are never served to a viewer (Storage, Dedup & Delivery Economics).
Audio track: an alternate audio rendition of a video, such as a dubbed language or an audio-description track, carried in HLS as an EXT-X-MEDIA audio rendition with a source, a language and a state (Video on Demand).
Automatic captions: captions generated by speech recognition, which YouTube publishes when ready and lets creators edit or remove; an automatic track should never overwrite one the creator uploaded (Video on Demand).
AV1: a video codec published in 2018 by AOMedia under a royalty-free patent policy, about 30% more efficient than VP9 and HEVC by Meta's measurement, but expensive to encode and heavy to decode on older phones (Video on Demand).
AVIF: an image format that gives the smallest files for photographs but is slow to encode, so it pays off for images viewed many times and often not for the long tail (Images at Scale).
Blob: the stored object under one or more media assets: every asset points at a blob, exact duplicates share one, and it is purged only when no referencing asset is live and no legal hold covers any of them (Storage, Dedup & Delivery Economics).
Blocking scan: a check the scanning gate waits for, such as hash-list matching, which fails closed: if the service is down, uploads wait in scanning rather than publishing and scanning later (The Upload Path).
Byte-range playlist: an HLS playlist (EXT-X-BYTERANGE) that addresses each segment as a range of a larger file, so a rendition can be one object and a clip can be a playlist over a recording rather than a copy (Video on Demand).
C2PA Content Credentials: a signed manifest recording an asset's provenance (origin, edits, AI use), bound to the content by a cryptographic hash so any alteration breaks the binding; validation says who signed it, not that the content is true (Images at Scale).
Cache purge: a request to evict a file's cached copies from a CDN, every variant and format included; a takedown isn't done until a re-fetch confirms the cache no longer serves it (Storage, Dedup & Delivery Economics).
Channel points: a channel-scoped points currency viewers earn by watching and spend on that channel's rewards, polls and predictions; run as an append-only ledger and kept non-monetary (never bought, transferred or cashed out) (Live Chat & the Interactive Layer).
Channel type (IVS): Amazon IVS's per-channel processing choice (full transcode, transcode to a 720p or 480p ceiling, or transmux only), each priced per input hour, with prices differing about tenfold between the cheapest and dearest (Live Streaming: Ingest & Transcoding).
Chat mode: a channel-wide gate the broadcaster or a moderator sets on live chat, such as slow mode, followers-only, subscribers- or members-only, emote-only, unique chat or approved users only (Live Chat & the Interactive Layer).
Clear-key encryption: segment encryption with HLS EXT-X-KEY (AES-128 or SAMPLE-AES) and a key URL behind authentication, which stops casual downloading of segments (Video on Demand).
Client-side multi-encode: having the creator's broadcast software encode the whole ladder on their own GPU and send each rendition as a track over Enhanced RTMP, so the platform transmuxes instead of transcoding (Twitch Enhanced Broadcasting, IVS multitrack video) (Live Streaming: Ingest & Transcoding).
Clip: a short cut of a live stream made by a viewer or the streamer, published as its own content item with its own author, reports and life, pointing back to the source stream and able to outlive its recording (Live Latency, Playback & VODs).
CMAF (Common Media Application Format): ISO/IEC 23000-19: one set of fragmented MP4 segments referenced by both an HLS playlist and a DASH manifest, so every device requests the same object under the same cache key (Video on Demand).
Codec portfolio: the set of codecs (H.264, VP9, HEVC, AV1) a platform keeps so it can serve each device the best one it decodes; each codec is another ladder and more storage, so better codecs are gated on watch time (Video on Demand).
Cold tier: millisecond-retrieval storage classes with a 90-day minimum, such as S3 Glacier Instant Retrieval, for long-tail originals that almost never load (Storage, Dedup & Delivery Economics).
Common Encryption (CENC): ISO/IEC 23001-7, which defines four encryption schemes of which two are used: cenc (AES-CTR) and cbcs (AES-CBC with a subsample pattern); cbcs support across FairPlay, Widevine and PlayReady lets one CMAF set serve all three (Video on Demand).
Content class: a classification of an image (photo, graphic, screenshot, artwork) recorded with the asset to tune quality, since screenshots and line art need a near-lossless path and photographs take aggressive perceptual quality (Images at Scale).
Content disarm: publishing pixels the platform decoded and re-encoded instead of the uploaded bytes, which removes most of what a crafted file carries and strips metadata, though the decoder doing it still needs a sandbox (The Upload Path).
Content-addressed storage: storing an object under the cryptographic hash of its bytes, so byte-identical uploads share one stored object; it saves nothing for re-encoded, resized or re-saved copies (Storage, Dedup & Delivery Economics).
Convex hull: in per-title encoding, the upper-left boundary of a video's trial encodes plotted as quality against bitrate, where no other encode gives better quality for fewer bits; ladder rungs are picked from it (Video on Demand).
Cost per thousand views: a surface's delivery cost per thousand views (roughly bytes served per view times the egress price, plus per-request and per-transformation fees, divided by how often the CDN hits), tracked per surface rather than as a monthly total (Storage, Dedup & Delivery Economics).
Cost per viewer-hour: what it costs to encode, store and deliver an hour of watched video, the unit that ties the video pipeline to the business; delivery is usually the largest part at scale (Video on Demand).
Creator export: a creator downloading their own uploads, which returns a rendition rather than the original if the platform discarded the source, and is refused for removed or struck content (Video on Demand).
Crypto-shredding: deleting a file by destroying its own encryption key, which makes copies in erasure-coded volumes and backups unreadable without rewriting them, as Meta's f4 did per BLOB (Storage, Dedup & Delivery Economics).
Cue: something that has to happen at a point in the video rather than at a wall-clock time (an ad break, timed metadata, a marker, a chapter, a stage change), recorded as a row for reconciliation (Live Latency, Playback & VODs).
Custom reward: a channel-points reward a channel defines (Twitch allows up to 50), with a cost of at least 1 point, optional per-stream and per-viewer limits and a cooldown; a redemption stays unfulfilled until fulfilled or canceled, and canceling refunds the points (Live Chat & the Interactive Layer).
DASH (MPEG-DASH): an adaptive streaming standard ratified by ISO in April 2012 and owned by no one company, with an XML manifest (the MPD), templated segment URLs, common encryption and multiple base URLs for multi-CDN delivery (Video on Demand).
Decompression bomb: a small file that decodes into an enormous image, such as a uniform-color PNG declaring huge dimensions, guarded against by checking declared pixels against a cap before allocating anything (The Upload Path).
Delivery restriction: a dated row the CDN layer reads to refuse or limit delivery (a geo-block, a referrer allowlist, signed-only access, an embed block), scoped to the platform, a community or an asset and carrying its legal basis (Storage, Dedup & Delivery Economics).
Direct-to-storage upload: the pattern in which the API grants a scoped, expiring credential for one platform-named key and the client sends bytes straight to storage or a vendor, so the application server never carries them (The Upload Path).
DRM (digital rights management): encryption whose keys are released only to a trusted decryption module on the device, through Widevine, PlayReady and FairPlay license servers; a cost and a constraint, used where a contract or paid content requires it (Video on Demand).
Drops: watch-to-earn campaigns in which a game studio grants an in-game reward of no monetary value to viewers who watched participating channels (or subscribed) once their platform and game accounts are linked (Live Chat & the Interactive Layer).
Dual format: sending a horizontal and a vertical layout of one broadcast at once over a multitrack connection, so desktop viewers see landscape and a phone held upright sees vertical (a Twitch beta from October 2025) (Live Streaming: Ingest & Transcoding).
DVR (live rewind): letting viewers rewind within a stream while it is still live, set per stream; YouTube limits it on streams over 12 hours and Mux doesn't recommend it over four hours (Live Latency, Playback & VODs).
Eager variant: an image variant made when the upload finishes, before the asset goes ready, giving predictable latency and storage for the sizes every new post needs (Images at Scale).
Edge scanning: a CDN comparing content served through its cache against known child-safety lists, as Cloudflare's CSAM Scanning Tool does, blocking matches where it can and notifying the customer daily (Storage, Dedup & Delivery Economics).
Egress: data transferred out to the internet, which most clouds bill (Amazon S3 after the first 100 GB a month) and Cloudflare R2 doesn't, so the provider's meter decides which behaviors cost money (Storage, Dedup & Delivery Economics).
Encoding ladder: the set of renditions a video is encoded into, with a resolution, bitrate and frame rate per rung, kept as versioned data so a player can step down and up as bandwidth changes (Video on Demand).
Enhanced RTMP: the Veovera extension of RTMP (v2 dated January 31, 2026) adding FourCC signaling for codecs such as HEVC, VP9 and AV1, HDR, multiple tracks in one connection and a server-initiated reconnect request (Live Streaming: Ingest & Transcoding).
Erasure coding: storing data as coded fragments (Reed-Solomon within a datacenter, XOR across datacenters in Meta's f4) instead of full replicas, cutting the effective replication factor from 3.6 to 2.8 or 2.1 for warm data (Storage, Dedup & Delivery Economics).
EventSub: Twitch's event subscription catalog (chat, holds, raids, rewards, polls, predictions, Hype Train, Shield Mode, ad breaks, drop grants and more), delivered at least once over webhooks, WebSockets or conduits with a message ID for de-duplication (Live Chat & the Interactive Layer).
Eviction order: the order in which a retention cap deletes recordings or uploads, such as least viewed first (which protects what audiences value) or oldest first (simpler, and harder on long-time creators) (Live Latency, Playback & VODs).
Exact dedup: deduplication keyed on a cryptographic hash of the original's bytes, done on the server after the full upload arrives so neither the response nor its timing reveals that someone else uploaded the same file (Storage, Dedup & Delivery Economics).
EXIF Orientation: the metadata tag (values 1 to 8) recording how a phone's stored frame should be rotated or mirrored; the pipeline applies it to the pixels before stripping the tag, or the variant comes out sideways (Images at Scale).
Extension: third-party web code in the viewing experience, shown as a panel below the video, an overlay over it or a component inside part of it, running in a sandboxed iframe on a separate origin (Live Chat & the Interactive Layer).
FairPlay Streaming: Apple's DRM for HLS on Apple platforms, for which Apple approves production credentials only for teams providing a consumer streaming service (Video on Demand).
Fanout: delivering one chat message to everyone in a room, published once to the channel's topic and written by each edge node to its local sockets, so cost and load scale with deliveries rather than sends (Live Chat & the Interactive Layer).
Fast start: an MP4 with its moov index at the front, so the platform can probe duration, codecs and dimensions from the first bytes and begin work before the whole file is read (The Upload Path).
Format negotiation: choosing an image format per viewer, either in markup (<picture> alternatives) or on the server from the request's Accept header, which then requires the cache to vary by it (Images at Scale).
Geo-block: refusing delivery by the viewer's country at the edge; on a media host it blanks the platform's files on every embedding site in that country, and it is served as a 451 and recorded with its legal basis (Storage, Dedup & Delivery Economics).
GIFV: Imgur's .gifv extension, introduced on October 9, 2014, for GIF uploads converted to WebM or MP4 video, the same move as serving animation as silent looping video at platform scale (Images at Scale).
Glass to glass: live latency measured from the camera capturing a frame to a viewer seeing it, stated by platforms as what most viewers get because the viewer's network and player buffer are part of it (Live Latency, Playback & VODs).
H.264 (AVC): the video codec that decodes on essentially every device, the universal fallback and the format of upload-time encodes, and the least efficient in the portfolio (Video on Demand).
Hard binding: the part of a C2PA credential that binds the manifest to the content by a cryptographic hash, so a re-encoded variant no longer matches the original's binding (Images at Scale).
Hate raid: many new or throwaway accounts arriving in a channel at once to flood its chat with abuse, answered with a preset lockdown and a mass action keyed on what the attackers did (Live Chat & the Interactive Layer).
HEIC (HEIF): the high-efficiency photo format iPhones from the iPhone 7 capture when the camera is set to High Efficiency; only Safari renders it, so platforms accept it as input, keep the original and serve every variant in web formats (Images at Scale).
Held message: a chat message kept invisible to the room by a term list, classifier or restriction until a moderator allows or denies it, recorded with the layer, model version and decision (Live Chat & the Interactive Layer).
HEVC (H.265): Apple's high-efficiency and HDR video path, with its own ladder in Apple's HLS specification, allowed only in fragmented MP4 in HLS and not covered by a royalty-free policy like AV1's (Video on Demand).
Highlight: a longer edit of a past broadcast made by the creator and kept after the past broadcast expires; Twitch caps highlights and uploads at 100 hours per channel since 2025 (Live Latency, Playback & VODs).
HLS (HTTP Live Streaming): Apple's adaptive streaming protocol, published as RFC 8216 in August 2017, in which a multivariant playlist lists variant streams and each media playlist lists one rendition's segments (Video on Demand).
Hold-back: the HLS rule (HOLD-BACK) that a player starts no closer to the live edge than at least three target durations, which with six-second segments puts standard HLS around 25 to 30 seconds behind once encoding and CDN time are added (Live Latency, Playback & VODs).
Hot tier: storage classes with millisecond retrieval and no minimum duration (S3 Standard, R2 Standard), for new uploads, active variants and anything a page may request (Storage, Dedup & Delivery Economics).
Hotlinking: another site's page loading a file directly from the platform's URL, which costs egress without an ad impression; platforms leave it open, limit it by referrer or require signed URLs, by content visibility (Storage, Dedup & Delivery Economics).
HTTP 451: the status code RFC 7725 defines as "Unavailable For Legal Reasons," returned for a file blocked by a legal demand with a Link header (rel="blocked-by") naming the blocking entity (Storage, Dedup & Delivery Economics).
Hype Train: a time-boxed group event on a channel in which contributions (cheers, subscriptions and gifted subscriptions) add points toward levels, expiring if contributions stop and followed by a cooldown (Live Chat & the Interactive Layer).
I-frame playlist: an HLS playlist of I-frames only (EXT-X-I-FRAME-STREAM-INF) that players use for scrubbing and fast-forward, required by Apple's authoring specification (Video on Demand).
IDR interval: the spacing of IDR (keyframe) frames, set to two seconds and aligned across every rendition with scene-cut keyframes disabled so segment boundaries line up and a player can switch at any segment (Live Streaming: Ingest & Transcoding).
Ingest edge: the point of presence where a creator's encoder connection lands, close to the creator, which terminates the protocol and forwards the stream to an origin (Live Streaming: Ingest & Transcoding).
Ingest routing: the per-stream decision, made at connect and again when capacity changes, that sends a stream from its edge to an origin with spare transcoding and backbone capacity (Twitch's Intelligest) (Live Streaming: Ingest & Transcoding).
Ingest session: one encoder connection feeding a stream, recorded with its protocol, edge, origin, codecs and outcome; a stream survives several sessions through reconnects and takeovers (Live Streaming: Ingest & Transcoding).
Ingredient (C2PA): a source file recorded in a C2PA manifest when an edit is made in a C2PA-aware tool, building a tree of provenance (Images at Scale).
Input envelope: the maximum resolution, frame rate and bitrate a channel's tier accepts at ingest; a feed outside it is refused with a reason rather than accepted and transcoded badly (Live Streaming: Ingest & Transcoding).
IPTC Digital Source Type: a controlled vocabulary written into image metadata describing where an image came from (digitalCapture, computationalCapture, trainedAlgorithmicMedia and others), with terms retired over time (Images at Scale).
JPEG XL: an image format supported in Chrome 155, Edge 155, Firefox 158 and Safari 17 per MDN as of October 2026, which MDN still advises pairing with an AVIF, WebP or JPEG alternative (Images at Scale).
Just-in-time encoding: encoding that waits for the first viewer instead of running at upload, as Mux offers, so videos nobody watches never pay for their encodes (Video on Demand).
Latency tier: a named per-stream live latency setting (standard, reduced, low, real time) chosen before going live, each giving up something in quality ceiling, robustness, device reach, features or cost (Live Latency, Playback & VODs).
Lazy variant: an image variant made through the pipeline on first request and then written to media_variants and storage, so the second request is a read even after the CDN evicts it (Images at Scale).
Lead Moderator: a Twitch channel role, assignable from November 2025, that can add and remove moderators and change all moderation settings (Live Chat & the Interactive Layer).
LL-HLS (Low-Latency HLS): HLS with partial segments, preload hints, blocking playlist reloads, delta updates and rendition reports, which brings live latency under about five seconds (Live Latency, Playback & VODs).
Lockdown: a preset "under attack" bundle of chat settings (long followers-only, no first-time chatters, stricter holds, slow mode) applied in one click and logged as one decision; Twitch's version, Shield Mode, launched November 30, 2022 (Live Chat & the Interactive Layer).
Long tail: the mass of rarely requested files that miss the CDN and are served from storage, the storage system's real customer and the reason on-demand variants requested there are encodes rather than reads (Storage, Dedup & Delivery Economics).
Message review handler: Amazon IVS Chat's hook that calls the platform's function for every message and waits at most 200 ms, applying a preset fallback result (allow or deny) on a timeout or error (Live Chat & the Interactive Layer).
Metadata stripping: removing EXIF and XMP metadata (GPS, capture time, serial numbers) from every served variant, keeping only an allowlist, after orientation and color profile have been applied to the pixels (Images at Scale).
Mezzanine file: the high-quality master a platform transcodes from; platforms publish a mezzanine spec (YouTube's recommended upload settings) because the input sets the quality ceiling and how fast processing starts (The Upload Path).
Minimum storage duration: the minimum period a colder storage class bills for (30, 90 or 180 days), so moving young or soon-to-be-purged files into it pays the minimum anyway (Storage, Dedup & Delivery Economics).
Multipart upload: S3's protocol for large objects: parts numbered 1 to 10,000 sent in any order and completed with a list; incomplete uploads have no expiry and are billed until completed or aborted (The Upload Path).
Multivariant playlist: the top-level HLS playlist (the RFC's master playlist) listing each variant stream with its peak BANDWIDTH, codecs and resolution, plus alternate audio, subtitle and caption renditions (Video on Demand).
Near match: a perceptual near-duplicate pair within the platform's distance threshold, stored as a lead for repost collapsing, takedown sweeps and re-upload detection, never as an automatic verdict (Storage, Dedup & Delivery Economics).
Non-moderator chat delay: holding every non-moderator chat message for a few seconds (Twitch offers 2, 4 or 6) so moderators and bots can remove it before viewers see it (Live Chat & the Interactive Layer).
On-demand variant: an image variant made the first time a URL asks for it and cached afterwards, so redesigns need no backfill, at the risk of an encoding storm when the cache is cold (Images at Scale).
Opaque ID: the viewer identifier an extension receives instead of the account (on a platform's own build, a keyed hash of extension and account, so it can't be joined across extensions); the real ID is released only after the viewer accepts an identity-sharing prompt (Live Chat & the Interactive Layer).
Origin (live): where a live stream's video is transcoded, chosen for spare transcoding capacity and backbone headroom rather than nearness to the creator (Live Streaming: Ingest & Transcoding).
Orphan: stored data nothing points at, such as abandoned uploads, unattached assets, variants of purged originals and anonymous uploads, which a media pipeline produces constantly (Storage, Dedup & Delivery Economics).
P2P assist (peer-to-peer assist): viewers sharing video segments with each other to offload the CDN, at the cost of viewers' upload bandwidth and a more complex player; Twitch tested it in Korea (Live Latency, Playback & VODs).
Partial segment: an LL-HLS subdivision of a segment (EXT-X-PART) no longer than a part target, which a player can request before the full segment exists (Live Latency, Playback & VODs).
Past broadcast: the recording of a live stream kept as an on-demand video for a period that, on Twitch, has historically depended on the channel's program status (Live Latency, Playback & VODs).
Per-title encoding: computing an encoding ladder for each video instead of using a fixed table, because the right bitrate depends on the content; Netflix moved to it in December 2015 (Video on Demand).
Perceptual hash: a hash of an image's or video's content that stays close for the same picture recompressed, resized or lightly cropped, used to find near-duplicates and to match against hash lists (Storage, Dedup & Delivery Economics).
Perceptual quality: a quality target set per image to hold a perceived level (Cloudinary's q_auto, Cloudflare's perceptual levels) instead of one fixed number, since the same setting means different things per image and encoder (Images at Scale).
Persistent license: a DRM license stored on the device for offline playback, with an expiry and a play window (Video on Demand).
Pipeline version: the version of encoder settings recorded on every variant, so a change in quality, format or stripping rolls out by regenerating stale variants in the background rather than invalidating everything (Images at Scale).
Platform events: live state updates (a poll's tally, a goal's progress, a viewer count) distributed on a separate path from chat lines and coalesced to the latest value, since a client only needs the newest state (Live Chat & the Interactive Layer).
PlayReady: Microsoft's DRM system, one of the three (with Widevine and FairPlay) that cover the market (Video on Demand).
Poll (live): a question with 2 to 5 choices run in a channel for a fixed window, one at a time, where viewers vote once and a channel may allow extra votes paid in channel points (Live Chat & the Interactive Layer).
Prediction: a question with 2 to 10 outcomes on which viewers stake channel points during a window; the broadcaster locks, resolves (paying those who picked the winner) or cancels it with refunds, and Twitch cancels it after 24 hours unresolved (Live Chat & the Interactive Layer).
Presigned URL: a URL signed to grant one storage operation until it expires, which Amazon calls a bearer token: reusable until expiry (up to 7 days from the SDKs) and bound to the credentials that signed it (The Upload Path).
Probe: the record made of every video on arrival (container, codecs, resolution, frame rate, bitrate, fast start), used to accept fixable deviations, reject the rest and schedule the transcode (The Upload Path).
Processing tier: which server-side processing a channel gets (guaranteed, priority, as available or source only) when transcoding capacity is rationed, recorded with the stream (Live Streaming: Ingest & Transcoding).
QoE (quality of experience): per-playback measurements players report (whether playback started, startup time, rebuffering, rungs played, upscaling), measured per rendition to find a wrong ladder or device rule (Video on Demand).
Quarantine: the media asset state a hash-list match moves an upload into: the original is kept in restricted storage under a legal hold and a moderation case opens, in one transaction, rather than deleting evidence (The Upload Path).
Raid: a broadcaster sending their viewers to another live channel at the end of a stream after a countdown (90 seconds on Twitch), with the target's consent setting and the sender accountable for what arrives (Live Chat & the Interactive Layer).
Real-time stage: a small set of participants talking in real time over WebRTC through an SFU (IVS: up to 12 hosts, under 300 ms), with a composition broadcast to the audience at a higher latency tier (Live Latency, Playback & VODs).
Reconnect window: the time a platform holds a stream open after the encoder disconnects, so a reconnect continues the same playlist and recording (Mux: 60 seconds by default for standard latency, adjustable up to 1,800) (Live Streaming: Ingest & Transcoding).
Recording merge: writing an encoder's reconnect within a window into the same recording when resolution, frame rate and codecs match, so a broadcast yields one recording rather than a fragment per Wi-Fi drop (Live Latency, Playback & VODs).
Referrer-limited delivery: a CDN rule denying image requests whose Referer doesn't name an allowed domain; easy to bypass, and it also hides images from image search and link previews (Storage, Dedup & Delivery Economics).
Rejected (asset state): the terminal state for an upload that failed validation or whose session expired: its bytes are deleted and the row stays with a rejected_reason (The Upload Path).
Rendition: one encoded version of a video at a given resolution, bitrate and codec, stored as a row in media_variants; a rung of the encoding ladder (Video on Demand).
Rendition (C2PA): the C2PA term for a copy made by non-editorial transformations such as re-encoding or scaling; a platform that signs as a claim generator can record the original as an ingredient of its variant (Images at Scale).
Repackaging: rewriting existing encodes into new containers and manifests instead of re-encoding; Instagram cut the cost of its basic adaptive encodings 94% this way (Video on Demand).
Replication factor: the stored bytes per byte of data; Meta's Haystack ran an effective 3.6 (three copies on RAID-6), which f4's erasure coding cut to 2.8 or 2.1 for warm data (Storage, Dedup & Delivery Economics).
Replication tree: Twitch's hierarchy of data centers in which a viewer's request goes to the nearest one and is forwarded up until a node has the stream, collapsing a million identical requests into one per node (Live Latency, Playback & VODs).
Repost collapsing: folding reposts of the same image in a feed or search result into one, crediting the earliest upload, using perceptual near matches (Storage, Dedup & Delivery Economics).
Request coalescing (edge): an edge making one upstream fetch per segment however many viewers ask at once, which live delivery needs when a raid or a final multiplies an audience in a minute (Live Latency, Playback & VODs).
Reserve and reconcile: holding a declared size or duration against an account's quota when an upload is granted, releasing the rest at confirm and everything if the session expires, as Cloudflare Stream does with maxDurationSeconds (The Upload Path).
Resumable session (Google-style): the YouTube and Google Cloud Storage upload protocol in which a session URI receives chunked PUTs and a 308 Resume Incomplete reports the confirmed range; the session URI itself acts as an authentication token (The Upload Path).
Resumable upload: an upload protocol that survives a dropped connection by letting the client ask the server how much it received and continue from that offset: tus, Google-style resumable sessions and S3 multipart (The Upload Path).
Retention purge: a rule applied by a job that removes assets after a period, in a stated order, with exemptions, notice and skips for holds, recorded asset by asset (Storage, Dedup & Delivery Economics).
RTMP and RTMPS: the legacy TCP contribution protocol every encoder and platform speaks; RTMPS adds TLS, encrypting the stream key and the video (Live Streaming: Ingest & Transcoding).
Save-Data: a request header (Save-Data: on) a browser sends when the user asked to reduce data use; experimental and not Baseline per MDN, and any response shaped by it must list it in Vary (Images at Scale).
Scanning gate: the scanning state between an upload being received and anything being published, which an asset leaves only when every blocking scan (validation, exact hash, hash lists and the like) has passed (The Upload Path).
SCTE-35: splice signals marking where an ad break goes out and comes back in, carried as markers in recorded VOD playlists and mapped by the HLS draft into EXT-X-DATERANGE attributes (Live Latency, Playback & VODs).
Shared chat: a combined chat across collaborating live channels in which each message keeps its source channel and is moderated under that channel's rules and moderators; the room is a view, the message belongs to one community (Live Chat & the Interactive Layer).
Shoutout: a broadcaster recommending another channel in chat, a lighter form of raid, rate-limited on Twitch to once every 2 minutes and to the same channel once every 60 minutes (Live Chat & the Interactive Layer).
Signed playback: replacing a video's ID with a short-lived signed token, minted per viewer session after the entitlement check and verified on every manifest and segment request, optionally with country and IP rules (Video on Demand).
Signed URL: a URL (or cookie) carrying an expiring signature the CDN verifies, used for private, members-only and paid media; CloudFront recommends signed cookies for sets of files such as HLS segments (Storage, Dedup & Delivery Economics).
Simulcasting: broadcasting one stream to several services at once, which Twitch allows on conditions unless an exclusivity agreement applies, and which some vendors offer as a restreaming feature (Live Streaming: Ingest & Transcoding).
Soft binding: an invisible watermark or fingerprint that lets a C2PA credential be found again after it has been stripped from the file (Images at Scale).
Source cap: a ceiling on source quality applied at ingest where delivery is expensive, as Twitch capped source at 720p in South Korea before closing its business there on February 27, 2024 (Live Streaming: Ingest & Transcoding).
SRT: a UDP contribution protocol that retransmits lost packets within a latency window, drops late ones, encrypts with AES and carries a stream ID; its IETF draft is an expired individual submission, not a standard (Live Streaming: Ingest & Transcoding).
SSAI (server-side ad insertion): the platform signaling an ad break and an ad server stitching ad segments into each viewer's playlist; on IVS it requires the MPEG transport stream container (Live Latency, Playback & VODs).
Storage cap: a limit on what a free account or channel keeps (Flickr's 1,000 photos and videos from 2019, Twitch's 100 hours of highlights and uploads from 2025), set with an explicit order, exemptions and notice (Storage, Dedup & Delivery Economics).
Storyboard: a sprite sheet of frames at regular intervals with a WebVTT or JSON index of tile positions and times, which drives the preview on timeline hover (Video on Demand).
Stream key: a long-lived bearer credential that lets whoever holds it broadcast on a channel: one active key per channel, stored as an HMAC for lookup, revealed only after step-up and rotatable in one click (Live Streaming: Ingest & Transcoding).
Stream takeover: Amazon IVS's mechanism in which a new connection with a higher priority on the stream key replaces the running one without disconnecting viewers or splitting the recording (Live Streaming: Ingest & Transcoding).
Suspicious user: a Twitch chatter status (ACTIVE_MONITORING or RESTRICTED) that flags or holds a possible ban evader's messages without banning a possibly innocent account (Live Chat & the Interactive Layer).
Takedown sweep: queuing the perceptual near matches of a copy removed for a policy or legal reason for review, not automatic removal (Storage, Dedup & Delivery Economics).
Timed metadata: data placed at a point in the video rather than in wall-clock time (ID3 tags in segments on IVS, EXT-X-DATERANGE in HLS) so every viewer gets it at the same moment of the stream; it should carry an identifier, not content (Live Latency, Playback & VODs).
Tombstone (URL): the response a removed file's URL returns afterwards (HTTP 410, or a placeholder image for embeds) so caches stop refilling from the origin and embeds show a clear removal (Storage, Dedup & Delivery Economics).
Transcode farm: the job system that encodes video: a queue of work units (chunks cut at keyframes), a pool of CPU, GPU or ASIC encoders, and a scheduler that gives first playable renditions priority over backfills (Video on Demand).
Transcoding (live): decoding a creator's live feed and re-encoding it into several renditions at lower resolutions and bitrates, the largest compute bill of a live platform (Live Streaming: Ingest & Transcoding).
Transformation spec: the canonical, ordered form of an image variant's parameters (width snapped to an allowlist, fit, quality), so two URLs that mean the same transform hit the same row and cache key (Images at Scale).
Transmuxing: repackaging a creator's encoded video into the delivery format (RTMP in, HLS segments out) without re-encoding: almost free in compute, but viewers get exactly the bitrate the creator sent (Live Streaming: Ingest & Transcoding).
tus: an open resumable upload protocol: the client creates an upload with POST, asks the server's offset with HEAD and continues with PATCH from exactly that offset, with extensions for expiry, checksums, termination and concatenation (The Upload Path).
Unique transformation: a distinct image transform Cloudflare bills per calendar month ($0.50 per 1,000 beyond the included 5,000, as of October 2026), excluding format and quality, one reason widths are snapped to an allowlist (Images at Scale).
Upload session: one grant to send bytes for one asset, carrying the protocol, the byte and duration caps from the purpose and plan, the server's received offset and an expiry after which a sweeper cleans up (The Upload Path).
Variant: a derived file of a media asset (a size, crop, format and quality made from the original), the only thing a platform serves; images, posters and video renditions all live in media_variants (Images at Scale).
Vary: Accept: the response header telling caches a response depends on the request's Accept header, required for server-side format negotiation and best normalized into a few capability buckets (Images at Scale).
VCU (video coding unit): Google's Argos transcoding ASIC, deployed across tens of thousands of servers, giving 8x (H.264) to 20x (VP9) the throughput of software for offline two-pass encoding (Video on Demand).
Verified bot: a Twitch bot status that raises chat limits (7,500 messages per 30 seconds, 2,000 joins per 10 seconds) but doesn't exempt the bot from AutoMod or grant the chat bot badge (Live Chat & the Interactive Layer).
VP9: Google's video codec, which YouTube's VCU accelerates alongside H.264, absent from Apple's HLS authoring specification (Video on Demand).
Warm tier: millisecond-retrieval storage classes with a 30-day minimum and a retrieval fee (S3 Standard-IA, R2 Infrequent Access), for originals of older media that must still load on request (Storage, Dedup & Delivery Economics).
Warning (chat): a Twitch moderation action that blocks a user from chatting until they acknowledge it, with a reason of up to 500 characters, an early rung below timeouts and bans (Live Chat & the Interactive Layer).
Watch-time gate: running expensive encodes (per-title ladders, better codecs) only once a video's projected watch time crosses a threshold, so videos nobody watches never pay for them (Video on Demand).
WebP: an image format supported by every current major browser, the safe modern default for still and animated images (Images at Scale).
WHIP (WebRTC-HTTP Ingestion Protocol): RFC 9725 (March 2025): publishing a WebRTC stream with an SDP offer in an HTTP POST and a bearer token, allowing in-browser go-live with one audio and one video track per session (Live Streaming: Ingest & Transcoding).
Widevine: Google's DRM system, built on Common Encryption with a hardware root of trust provisioned at the factory (Video on Demand).

ABR (adaptive bitrate) streaming: a player switching between a video's renditions as the viewer's bandwidth drops or recovers, which is why every video is encoded into a ladder of rungs (Video on Demand).
Ad break: a cue in a live stream that inserts ads, recorded with its requester, requested duration, actual duration and the vendor's reference so creator ad revenue, advertiser billing and disputes reconcile against it (Live Latency, Playback & VODs).
Alt text: a description of an image for screen readers, stored on the post's use of the file (content_media) rather than on the file, because one image can illustrate two posts that need different descriptions (Images at Scale).
Anonymous upload: a file with no owner account (a null owner_account_id), which nobody can manage or be notified about, and so is the first candidate for a retention purge (Storage, Dedup & Delivery Economics).
Archive tier: storage classes such as S3 Glacier Flexible Retrieval and Deep Archive that retrieve in minutes to hours, with 90 and 180 day minimums, for evidence and originals that are never served to a viewer (Storage, Dedup & Delivery Economics).
Audio track: an alternate audio rendition of a video, such as a dubbed language or an audio-description track, carried in HLS as an EXT-X-MEDIA audio rendition with a source, a language and a state (Video on Demand).
Automatic captions: captions generated by speech recognition, which YouTube publishes when ready and lets creators edit or remove; an automatic track should never overwrite one the creator uploaded (Video on Demand).
AV1: a video codec published in 2018 by AOMedia under a royalty-free patent policy, about 30% more efficient than VP9 and HEVC by Meta's measurement, but expensive to encode and heavy to decode on older phones (Video on Demand).
AVIF: an image format that gives the smallest files for photographs but is slow to encode, so it pays off for images viewed many times and often not for the long tail (Images at Scale).
Blob: the stored object under one or more media assets: every asset points at a blob, exact duplicates share one, and it is purged only when no referencing asset is live and no legal hold covers any of them (Storage, Dedup & Delivery Economics).
Blocking scan: a check the scanning gate waits for, such as hash-list matching, which fails closed: if the service is down, uploads wait in scanning rather than publishing and scanning later (The Upload Path).
Byte-range playlist: an HLS playlist (EXT-X-BYTERANGE) that addresses each segment as a range of a larger file, so a rendition can be one object and a clip can be a playlist over a recording rather than a copy (Video on Demand).
C2PA Content Credentials: a signed manifest recording an asset's provenance (origin, edits, AI use), bound to the content by a cryptographic hash so any alteration breaks the binding; validation says who signed it, not that the content is true (Images at Scale).
Cache purge: a request to evict a file's cached copies from a CDN, every variant and format included; a takedown isn't done until a re-fetch confirms the cache no longer serves it (Storage, Dedup & Delivery Economics).
Channel points: a channel-scoped points currency viewers earn by watching and spend on that channel's rewards, polls and predictions; run as an append-only ledger and kept non-monetary (never bought, transferred or cashed out) (Live Chat & the Interactive Layer).
Channel type (IVS): Amazon IVS's per-channel processing choice (full transcode, transcode to a 720p or 480p ceiling, or transmux only), each priced per input hour, with prices differing about tenfold between the cheapest and dearest (Live Streaming: Ingest & Transcoding).
Chat mode: a channel-wide gate the broadcaster or a moderator sets on live chat, such as slow mode, followers-only, subscribers- or members-only, emote-only, unique chat or approved users only (Live Chat & the Interactive Layer).
Clear-key encryption: segment encryption with HLS EXT-X-KEY (AES-128 or SAMPLE-AES) and a key URL behind authentication, which stops casual downloading of segments (Video on Demand).
Client-side multi-encode: having the creator's broadcast software encode the whole ladder on their own GPU and send each rendition as a track over Enhanced RTMP, so the platform transmuxes instead of transcoding (Twitch Enhanced Broadcasting, IVS multitrack video) (Live Streaming: Ingest & Transcoding).
Clip: a short cut of a live stream made by a viewer or the streamer, published as its own content item with its own author, reports and life, pointing back to the source stream and able to outlive its recording (Live Latency, Playback & VODs).
CMAF (Common Media Application Format): ISO/IEC 23000-19: one set of fragmented MP4 segments referenced by both an HLS playlist and a DASH manifest, so every device requests the same object under the same cache key (Video on Demand).
Codec portfolio: the set of codecs (H.264, VP9, HEVC, AV1) a platform keeps so it can serve each device the best one it decodes; each codec is another ladder and more storage, so better codecs are gated on watch time (Video on Demand).
Cold tier: millisecond-retrieval storage classes with a 90-day minimum, such as S3 Glacier Instant Retrieval, for long-tail originals that almost never load (Storage, Dedup & Delivery Economics).
Common Encryption (CENC): ISO/IEC 23001-7, which defines four encryption schemes of which two are used: cenc (AES-CTR) and cbcs (AES-CBC with a subsample pattern); cbcs support across FairPlay, Widevine and PlayReady lets one CMAF set serve all three (Video on Demand).
Content class: a classification of an image (photo, graphic, screenshot, artwork) recorded with the asset to tune quality, since screenshots and line art need a near-lossless path and photographs take aggressive perceptual quality (Images at Scale).
Content disarm: publishing pixels the platform decoded and re-encoded instead of the uploaded bytes, which removes most of what a crafted file carries and strips metadata, though the decoder doing it still needs a sandbox (The Upload Path).
Content-addressed storage: storing an object under the cryptographic hash of its bytes, so byte-identical uploads share one stored object; it saves nothing for re-encoded, resized or re-saved copies (Storage, Dedup & Delivery Economics).
Convex hull: in per-title encoding, the upper-left boundary of a video's trial encodes plotted as quality against bitrate, where no other encode gives better quality for fewer bits; ladder rungs are picked from it (Video on Demand).
Cost per thousand views: a surface's delivery cost per thousand views (roughly bytes served per view times the egress price, plus per-request and per-transformation fees, divided by how often the CDN hits), tracked per surface rather than as a monthly total (Storage, Dedup & Delivery Economics).
Cost per viewer-hour: what it costs to encode, store and deliver an hour of watched video, the unit that ties the video pipeline to the business; delivery is usually the largest part at scale (Video on Demand).
Creator export: a creator downloading their own uploads, which returns a rendition rather than the original if the platform discarded the source, and is refused for removed or struck content (Video on Demand).
Crypto-shredding: deleting a file by destroying its own encryption key, which makes copies in erasure-coded volumes and backups unreadable without rewriting them, as Meta's f4 did per BLOB (Storage, Dedup & Delivery Economics).
Cue: something that has to happen at a point in the video rather than at a wall-clock time (an ad break, timed metadata, a marker, a chapter, a stage change), recorded as a row for reconciliation (Live Latency, Playback & VODs).
Custom reward: a channel-points reward a channel defines (Twitch allows up to 50), with a cost of at least 1 point, optional per-stream and per-viewer limits and a cooldown; a redemption stays unfulfilled until fulfilled or canceled, and canceling refunds the points (Live Chat & the Interactive Layer).
DASH (MPEG-DASH): an adaptive streaming standard ratified by ISO in April 2012 and owned by no one company, with an XML manifest (the MPD), templated segment URLs, common encryption and multiple base URLs for multi-CDN delivery (Video on Demand).
Decompression bomb: a small file that decodes into an enormous image, such as a uniform-color PNG declaring huge dimensions, guarded against by checking declared pixels against a cap before allocating anything (The Upload Path).
Delivery restriction: a dated row the CDN layer reads to refuse or limit delivery (a geo-block, a referrer allowlist, signed-only access, an embed block), scoped to the platform, a community or an asset and carrying its legal basis (Storage, Dedup & Delivery Economics).
Direct-to-storage upload: the pattern in which the API grants a scoped, expiring credential for one platform-named key and the client sends bytes straight to storage or a vendor, so the application server never carries them (The Upload Path).
DRM (digital rights management): encryption whose keys are released only to a trusted decryption module on the device, through Widevine, PlayReady and FairPlay license servers; a cost and a constraint, used where a contract or paid content requires it (Video on Demand).
Drops: watch-to-earn campaigns in which a game studio grants an in-game reward of no monetary value to viewers who watched participating channels (or subscribed) once their platform and game accounts are linked (Live Chat & the Interactive Layer).
Dual format: sending a horizontal and a vertical layout of one broadcast at once over a multitrack connection, so desktop viewers see landscape and a phone held upright sees vertical (a Twitch beta from October 2025) (Live Streaming: Ingest & Transcoding).
DVR (live rewind): letting viewers rewind within a stream while it is still live, set per stream; YouTube limits it on streams over 12 hours and Mux doesn't recommend it over four hours (Live Latency, Playback & VODs).
Eager variant: an image variant made when the upload finishes, before the asset goes ready, giving predictable latency and storage for the sizes every new post needs (Images at Scale).
Edge scanning: a CDN comparing content served through its cache against known child-safety lists, as Cloudflare's CSAM Scanning Tool does, blocking matches where it can and notifying the customer daily (Storage, Dedup & Delivery Economics).
Egress: data transferred out to the internet, which most clouds bill (Amazon S3 after the first 100 GB a month) and Cloudflare R2 doesn't, so the provider's meter decides which behaviors cost money (Storage, Dedup & Delivery Economics).
Encoding ladder: the set of renditions a video is encoded into, with a resolution, bitrate and frame rate per rung, kept as versioned data so a player can step down and up as bandwidth changes (Video on Demand).
Enhanced RTMP: the Veovera extension of RTMP (v2 dated January 31, 2026) adding FourCC signaling for codecs such as HEVC, VP9 and AV1, HDR, multiple tracks in one connection and a server-initiated reconnect request (Live Streaming: Ingest & Transcoding).
Erasure coding: storing data as coded fragments (Reed-Solomon within a datacenter, XOR across datacenters in Meta's f4) instead of full replicas, cutting the effective replication factor from 3.6 to 2.8 or 2.1 for warm data (Storage, Dedup & Delivery Economics).
EventSub: Twitch's event subscription catalog (chat, holds, raids, rewards, polls, predictions, Hype Train, Shield Mode, ad breaks, drop grants and more), delivered at least once over webhooks, WebSockets or conduits with a message ID for de-duplication (Live Chat & the Interactive Layer).
Eviction order: the order in which a retention cap deletes recordings or uploads, such as least viewed first (which protects what audiences value) or oldest first (simpler, and harder on long-time creators) (Live Latency, Playback & VODs).
Exact dedup: deduplication keyed on a cryptographic hash of the original's bytes, done on the server after the full upload arrives so neither the response nor its timing reveals that someone else uploaded the same file (Storage, Dedup & Delivery Economics).
EXIF Orientation: the metadata tag (values 1 to 8) recording how a phone's stored frame should be rotated or mirrored; the pipeline applies it to the pixels before stripping the tag, or the variant comes out sideways (Images at Scale).
Extension: third-party web code in the viewing experience, shown as a panel below the video, an overlay over it or a component inside part of it, running in a sandboxed iframe on a separate origin (Live Chat & the Interactive Layer).
FairPlay Streaming: Apple's DRM for HLS on Apple platforms, for which Apple approves production credentials only for teams providing a consumer streaming service (Video on Demand).
Fanout: delivering one chat message to everyone in a room, published once to the channel's topic and written by each edge node to its local sockets, so cost and load scale with deliveries rather than sends (Live Chat & the Interactive Layer).
Fast start: an MP4 with its moov index at the front, so the platform can probe duration, codecs and dimensions from the first bytes and begin work before the whole file is read (The Upload Path).
Format negotiation: choosing an image format per viewer, either in markup (<picture> alternatives) or on the server from the request's Accept header, which then requires the cache to vary by it (Images at Scale).
Geo-block: refusing delivery by the viewer's country at the edge; on a media host it blanks the platform's files on every embedding site in that country, and it is served as a 451 and recorded with its legal basis (Storage, Dedup & Delivery Economics).
GIFV: Imgur's .gifv extension, introduced on October 9, 2014, for GIF uploads converted to WebM or MP4 video, the same move as serving animation as silent looping video at platform scale (Images at Scale).
Glass to glass: live latency measured from the camera capturing a frame to a viewer seeing it, stated by platforms as what most viewers get because the viewer's network and player buffer are part of it (Live Latency, Playback & VODs).
H.264 (AVC): the video codec that decodes on essentially every device, the universal fallback and the format of upload-time encodes, and the least efficient in the portfolio (Video on Demand).
Hard binding: the part of a C2PA credential that binds the manifest to the content by a cryptographic hash, so a re-encoded variant no longer matches the original's binding (Images at Scale).
Hate raid: many new or throwaway accounts arriving in a channel at once to flood its chat with abuse, answered with a preset lockdown and a mass action keyed on what the attackers did (Live Chat & the Interactive Layer).
HEIC (HEIF): the high-efficiency photo format iPhones from the iPhone 7 capture when the camera is set to High Efficiency; only Safari renders it, so platforms accept it as input, keep the original and serve every variant in web formats (Images at Scale).
Held message: a chat message kept invisible to the room by a term list, classifier or restriction until a moderator allows or denies it, recorded with the layer, model version and decision (Live Chat & the Interactive Layer).
HEVC (H.265): Apple's high-efficiency and HDR video path, with its own ladder in Apple's HLS specification, allowed only in fragmented MP4 in HLS and not covered by a royalty-free policy like AV1's (Video on Demand).
Highlight: a longer edit of a past broadcast made by the creator and kept after the past broadcast expires; Twitch caps highlights and uploads at 100 hours per channel since 2025 (Live Latency, Playback & VODs).
HLS (HTTP Live Streaming): Apple's adaptive streaming protocol, published as RFC 8216 in August 2017, in which a multivariant playlist lists variant streams and each media playlist lists one rendition's segments (Video on Demand).
Hold-back: the HLS rule (HOLD-BACK) that a player starts no closer to the live edge than at least three target durations, which with six-second segments puts standard HLS around 25 to 30 seconds behind once encoding and CDN time are added (Live Latency, Playback & VODs).
Hot tier: storage classes with millisecond retrieval and no minimum duration (S3 Standard, R2 Standard), for new uploads, active variants and anything a page may request (Storage, Dedup & Delivery Economics).
Hotlinking: another site's page loading a file directly from the platform's URL, which costs egress without an ad impression; platforms leave it open, limit it by referrer or require signed URLs, by content visibility (Storage, Dedup & Delivery Economics).
HTTP 451: the status code RFC 7725 defines as "Unavailable For Legal Reasons," returned for a file blocked by a legal demand with a Link header (rel="blocked-by") naming the blocking entity (Storage, Dedup & Delivery Economics).
Hype Train: a time-boxed group event on a channel in which contributions (cheers, subscriptions and gifted subscriptions) add points toward levels, expiring if contributions stop and followed by a cooldown (Live Chat & the Interactive Layer).
I-frame playlist: an HLS playlist of I-frames only (EXT-X-I-FRAME-STREAM-INF) that players use for scrubbing and fast-forward, required by Apple's authoring specification (Video on Demand).
IDR interval: the spacing of IDR (keyframe) frames, set to two seconds and aligned across every rendition with scene-cut keyframes disabled so segment boundaries line up and a player can switch at any segment (Live Streaming: Ingest & Transcoding).
Ingest edge: the point of presence where a creator's encoder connection lands, close to the creator, which terminates the protocol and forwards the stream to an origin (Live Streaming: Ingest & Transcoding).
Ingest routing: the per-stream decision, made at connect and again when capacity changes, that sends a stream from its edge to an origin with spare transcoding and backbone capacity (Twitch's Intelligest) (Live Streaming: Ingest & Transcoding).
Ingest session: one encoder connection feeding a stream, recorded with its protocol, edge, origin, codecs and outcome; a stream survives several sessions through reconnects and takeovers (Live Streaming: Ingest & Transcoding).
Ingredient (C2PA): a source file recorded in a C2PA manifest when an edit is made in a C2PA-aware tool, building a tree of provenance (Images at Scale).
Input envelope: the maximum resolution, frame rate and bitrate a channel's tier accepts at ingest; a feed outside it is refused with a reason rather than accepted and transcoded badly (Live Streaming: Ingest & Transcoding).
IPTC Digital Source Type: a controlled vocabulary written into image metadata describing where an image came from (digitalCapture, computationalCapture, trainedAlgorithmicMedia and others), with terms retired over time (Images at Scale).
JPEG XL: an image format supported in Chrome 155, Edge 155, Firefox 158 and Safari 17 per MDN as of October 2026, which MDN still advises pairing with an AVIF, WebP or JPEG alternative (Images at Scale).
Just-in-time encoding: encoding that waits for the first viewer instead of running at upload, as Mux offers, so videos nobody watches never pay for their encodes (Video on Demand).
Latency tier: a named per-stream live latency setting (standard, reduced, low, real time) chosen before going live, each giving up something in quality ceiling, robustness, device reach, features or cost (Live Latency, Playback & VODs).
Lazy variant: an image variant made through the pipeline on first request and then written to media_variants and storage, so the second request is a read even after the CDN evicts it (Images at Scale).
Lead Moderator: a Twitch channel role, assignable from November 2025, that can add and remove moderators and change all moderation settings (Live Chat & the Interactive Layer).
LL-HLS (Low-Latency HLS): HLS with partial segments, preload hints, blocking playlist reloads, delta updates and rendition reports, which brings live latency under about five seconds (Live Latency, Playback & VODs).
Lockdown: a preset "under attack" bundle of chat settings (long followers-only, no first-time chatters, stricter holds, slow mode) applied in one click and logged as one decision; Twitch's version, Shield Mode, launched November 30, 2022 (Live Chat & the Interactive Layer).
Long tail: the mass of rarely requested files that miss the CDN and are served from storage, the storage system's real customer and the reason on-demand variants requested there are encodes rather than reads (Storage, Dedup & Delivery Economics).
Message review handler: Amazon IVS Chat's hook that calls the platform's function for every message and waits at most 200 ms, applying a preset fallback result (allow or deny) on a timeout or error (Live Chat & the Interactive Layer).
Metadata stripping: removing EXIF and XMP metadata (GPS, capture time, serial numbers) from every served variant, keeping only an allowlist, after orientation and color profile have been applied to the pixels (Images at Scale).
Mezzanine file: the high-quality master a platform transcodes from; platforms publish a mezzanine spec (YouTube's recommended upload settings) because the input sets the quality ceiling and how fast processing starts (The Upload Path).
Minimum storage duration: the minimum period a colder storage class bills for (30, 90 or 180 days), so moving young or soon-to-be-purged files into it pays the minimum anyway (Storage, Dedup & Delivery Economics).
Multipart upload: S3's protocol for large objects: parts numbered 1 to 10,000 sent in any order and completed with a list; incomplete uploads have no expiry and are billed until completed or aborted (The Upload Path).
Multivariant playlist: the top-level HLS playlist (the RFC's master playlist) listing each variant stream with its peak BANDWIDTH, codecs and resolution, plus alternate audio, subtitle and caption renditions (Video on Demand).
Near match: a perceptual near-duplicate pair within the platform's distance threshold, stored as a lead for repost collapsing, takedown sweeps and re-upload detection, never as an automatic verdict (Storage, Dedup & Delivery Economics).
Non-moderator chat delay: holding every non-moderator chat message for a few seconds (Twitch offers 2, 4 or 6) so moderators and bots can remove it before viewers see it (Live Chat & the Interactive Layer).
On-demand variant: an image variant made the first time a URL asks for it and cached afterwards, so redesigns need no backfill, at the risk of an encoding storm when the cache is cold (Images at Scale).
Opaque ID: the viewer identifier an extension receives instead of the account (on a platform's own build, a keyed hash of extension and account, so it can't be joined across extensions); the real ID is released only after the viewer accepts an identity-sharing prompt (Live Chat & the Interactive Layer).
Origin (live): where a live stream's video is transcoded, chosen for spare transcoding capacity and backbone headroom rather than nearness to the creator (Live Streaming: Ingest & Transcoding).
Orphan: stored data nothing points at, such as abandoned uploads, unattached assets, variants of purged originals and anonymous uploads, which a media pipeline produces constantly (Storage, Dedup & Delivery Economics).
P2P assist (peer-to-peer assist): viewers sharing video segments with each other to offload the CDN, at the cost of viewers' upload bandwidth and a more complex player; Twitch tested it in Korea (Live Latency, Playback & VODs).
Partial segment: an LL-HLS subdivision of a segment (EXT-X-PART) no longer than a part target, which a player can request before the full segment exists (Live Latency, Playback & VODs).
Past broadcast: the recording of a live stream kept as an on-demand video for a period that, on Twitch, has historically depended on the channel's program status (Live Latency, Playback & VODs).
Per-title encoding: computing an encoding ladder for each video instead of using a fixed table, because the right bitrate depends on the content; Netflix moved to it in December 2015 (Video on Demand).
Perceptual hash: a hash of an image's or video's content that stays close for the same picture recompressed, resized or lightly cropped, used to find near-duplicates and to match against hash lists (Storage, Dedup & Delivery Economics).
Perceptual quality: a quality target set per image to hold a perceived level (Cloudinary's q_auto, Cloudflare's perceptual levels) instead of one fixed number, since the same setting means different things per image and encoder (Images at Scale).
Persistent license: a DRM license stored on the device for offline playback, with an expiry and a play window (Video on Demand).
Pipeline version: the version of encoder settings recorded on every variant, so a change in quality, format or stripping rolls out by regenerating stale variants in the background rather than invalidating everything (Images at Scale).
Platform events: live state updates (a poll's tally, a goal's progress, a viewer count) distributed on a separate path from chat lines and coalesced to the latest value, since a client only needs the newest state (Live Chat & the Interactive Layer).
PlayReady: Microsoft's DRM system, one of the three (with Widevine and FairPlay) that cover the market (Video on Demand).
Poll (live): a question with 2 to 5 choices run in a channel for a fixed window, one at a time, where viewers vote once and a channel may allow extra votes paid in channel points (Live Chat & the Interactive Layer).
Prediction: a question with 2 to 10 outcomes on which viewers stake channel points during a window; the broadcaster locks, resolves (paying those who picked the winner) or cancels it with refunds, and Twitch cancels it after 24 hours unresolved (Live Chat & the Interactive Layer).
Presigned URL: a URL signed to grant one storage operation until it expires, which Amazon calls a bearer token: reusable until expiry (up to 7 days from the SDKs) and bound to the credentials that signed it (The Upload Path).
Probe: the record made of every video on arrival (container, codecs, resolution, frame rate, bitrate, fast start), used to accept fixable deviations, reject the rest and schedule the transcode (The Upload Path).
Processing tier: which server-side processing a channel gets (guaranteed, priority, as available or source only) when transcoding capacity is rationed, recorded with the stream (Live Streaming: Ingest & Transcoding).
QoE (quality of experience): per-playback measurements players report (whether playback started, startup time, rebuffering, rungs played, upscaling), measured per rendition to find a wrong ladder or device rule (Video on Demand).
Quarantine: the media asset state a hash-list match moves an upload into: the original is kept in restricted storage under a legal hold and a moderation case opens, in one transaction, rather than deleting evidence (The Upload Path).
Raid: a broadcaster sending their viewers to another live channel at the end of a stream after a countdown (90 seconds on Twitch), with the target's consent setting and the sender accountable for what arrives (Live Chat & the Interactive Layer).
Real-time stage: a small set of participants talking in real time over WebRTC through an SFU (IVS: up to 12 hosts, under 300 ms), with a composition broadcast to the audience at a higher latency tier (Live Latency, Playback & VODs).
Reconnect window: the time a platform holds a stream open after the encoder disconnects, so a reconnect continues the same playlist and recording (Mux: 60 seconds by default for standard latency, adjustable up to 1,800) (Live Streaming: Ingest & Transcoding).
Recording merge: writing an encoder's reconnect within a window into the same recording when resolution, frame rate and codecs match, so a broadcast yields one recording rather than a fragment per Wi-Fi drop (Live Latency, Playback & VODs).
Referrer-limited delivery: a CDN rule denying image requests whose Referer doesn't name an allowed domain; easy to bypass, and it also hides images from image search and link previews (Storage, Dedup & Delivery Economics).
Rejected (asset state): the terminal state for an upload that failed validation or whose session expired: its bytes are deleted and the row stays with a rejected_reason (The Upload Path).
Rendition: one encoded version of a video at a given resolution, bitrate and codec, stored as a row in media_variants; a rung of the encoding ladder (Video on Demand).
Rendition (C2PA): the C2PA term for a copy made by non-editorial transformations such as re-encoding or scaling; a platform that signs as a claim generator can record the original as an ingredient of its variant (Images at Scale).
Repackaging: rewriting existing encodes into new containers and manifests instead of re-encoding; Instagram cut the cost of its basic adaptive encodings 94% this way (Video on Demand).
Replication factor: the stored bytes per byte of data; Meta's Haystack ran an effective 3.6 (three copies on RAID-6), which f4's erasure coding cut to 2.8 or 2.1 for warm data (Storage, Dedup & Delivery Economics).
Replication tree: Twitch's hierarchy of data centers in which a viewer's request goes to the nearest one and is forwarded up until a node has the stream, collapsing a million identical requests into one per node (Live Latency, Playback & VODs).
Repost collapsing: folding reposts of the same image in a feed or search result into one, crediting the earliest upload, using perceptual near matches (Storage, Dedup & Delivery Economics).
Request coalescing (edge): an edge making one upstream fetch per segment however many viewers ask at once, which live delivery needs when a raid or a final multiplies an audience in a minute (Live Latency, Playback & VODs).
Reserve and reconcile: holding a declared size or duration against an account's quota when an upload is granted, releasing the rest at confirm and everything if the session expires, as Cloudflare Stream does with maxDurationSeconds (The Upload Path).
Resumable session (Google-style): the YouTube and Google Cloud Storage upload protocol in which a session URI receives chunked PUTs and a 308 Resume Incomplete reports the confirmed range; the session URI itself acts as an authentication token (The Upload Path).
Resumable upload: an upload protocol that survives a dropped connection by letting the client ask the server how much it received and continue from that offset: tus, Google-style resumable sessions and S3 multipart (The Upload Path).
Retention purge: a rule applied by a job that removes assets after a period, in a stated order, with exemptions, notice and skips for holds, recorded asset by asset (Storage, Dedup & Delivery Economics).
RTMP and RTMPS: the legacy TCP contribution protocol every encoder and platform speaks; RTMPS adds TLS, encrypting the stream key and the video (Live Streaming: Ingest & Transcoding).
Save-Data: a request header (Save-Data: on) a browser sends when the user asked to reduce data use; experimental and not Baseline per MDN, and any response shaped by it must list it in Vary (Images at Scale).
Scanning gate: the scanning state between an upload being received and anything being published, which an asset leaves only when every blocking scan (validation, exact hash, hash lists and the like) has passed (The Upload Path).
SCTE-35: splice signals marking where an ad break goes out and comes back in, carried as markers in recorded VOD playlists and mapped by the HLS draft into EXT-X-DATERANGE attributes (Live Latency, Playback & VODs).
Shared chat: a combined chat across collaborating live channels in which each message keeps its source channel and is moderated under that channel's rules and moderators; the room is a view, the message belongs to one community (Live Chat & the Interactive Layer).
Shoutout: a broadcaster recommending another channel in chat, a lighter form of raid, rate-limited on Twitch to once every 2 minutes and to the same channel once every 60 minutes (Live Chat & the Interactive Layer).
Signed playback: replacing a video's ID with a short-lived signed token, minted per viewer session after the entitlement check and verified on every manifest and segment request, optionally with country and IP rules (Video on Demand).
Signed URL: a URL (or cookie) carrying an expiring signature the CDN verifies, used for private, members-only and paid media; CloudFront recommends signed cookies for sets of files such as HLS segments (Storage, Dedup & Delivery Economics).
Simulcasting: broadcasting one stream to several services at once, which Twitch allows on conditions unless an exclusivity agreement applies, and which some vendors offer as a restreaming feature (Live Streaming: Ingest & Transcoding).
Soft binding: an invisible watermark or fingerprint that lets a C2PA credential be found again after it has been stripped from the file (Images at Scale).
Source cap: a ceiling on source quality applied at ingest where delivery is expensive, as Twitch capped source at 720p in South Korea before closing its business there on February 27, 2024 (Live Streaming: Ingest & Transcoding).
SRT: a UDP contribution protocol that retransmits lost packets within a latency window, drops late ones, encrypts with AES and carries a stream ID; its IETF draft is an expired individual submission, not a standard (Live Streaming: Ingest & Transcoding).
SSAI (server-side ad insertion): the platform signaling an ad break and an ad server stitching ad segments into each viewer's playlist; on IVS it requires the MPEG transport stream container (Live Latency, Playback & VODs).
Storage cap: a limit on what a free account or channel keeps (Flickr's 1,000 photos and videos from 2019, Twitch's 100 hours of highlights and uploads from 2025), set with an explicit order, exemptions and notice (Storage, Dedup & Delivery Economics).
Storyboard: a sprite sheet of frames at regular intervals with a WebVTT or JSON index of tile positions and times, which drives the preview on timeline hover (Video on Demand).
Stream key: a long-lived bearer credential that lets whoever holds it broadcast on a channel: one active key per channel, stored as an HMAC for lookup, revealed only after step-up and rotatable in one click (Live Streaming: Ingest & Transcoding).
Stream takeover: Amazon IVS's mechanism in which a new connection with a higher priority on the stream key replaces the running one without disconnecting viewers or splitting the recording (Live Streaming: Ingest & Transcoding).
Suspicious user: a Twitch chatter status (ACTIVE_MONITORING or RESTRICTED) that flags or holds a possible ban evader's messages without banning a possibly innocent account (Live Chat & the Interactive Layer).
Takedown sweep: queuing the perceptual near matches of a copy removed for a policy or legal reason for review, not automatic removal (Storage, Dedup & Delivery Economics).
Timed metadata: data placed at a point in the video rather than in wall-clock time (ID3 tags in segments on IVS, EXT-X-DATERANGE in HLS) so every viewer gets it at the same moment of the stream; it should carry an identifier, not content (Live Latency, Playback & VODs).
Tombstone (URL): the response a removed file's URL returns afterwards (HTTP 410, or a placeholder image for embeds) so caches stop refilling from the origin and embeds show a clear removal (Storage, Dedup & Delivery Economics).
Transcode farm: the job system that encodes video: a queue of work units (chunks cut at keyframes), a pool of CPU, GPU or ASIC encoders, and a scheduler that gives first playable renditions priority over backfills (Video on Demand).
Transcoding (live): decoding a creator's live feed and re-encoding it into several renditions at lower resolutions and bitrates, the largest compute bill of a live platform (Live Streaming: Ingest & Transcoding).
Transformation spec: the canonical, ordered form of an image variant's parameters (width snapped to an allowlist, fit, quality), so two URLs that mean the same transform hit the same row and cache key (Images at Scale).
Transmuxing: repackaging a creator's encoded video into the delivery format (RTMP in, HLS segments out) without re-encoding: almost free in compute, but viewers get exactly the bitrate the creator sent (Live Streaming: Ingest & Transcoding).
tus: an open resumable upload protocol: the client creates an upload with POST, asks the server's offset with HEAD and continues with PATCH from exactly that offset, with extensions for expiry, checksums, termination and concatenation (The Upload Path).
Unique transformation: a distinct image transform Cloudflare bills per calendar month ($0.50 per 1,000 beyond the included 5,000, as of October 2026), excluding format and quality, one reason widths are snapped to an allowlist (Images at Scale).
Upload session: one grant to send bytes for one asset, carrying the protocol, the byte and duration caps from the purpose and plan, the server's received offset and an expiry after which a sweeper cleans up (The Upload Path).
Variant: a derived file of a media asset (a size, crop, format and quality made from the original), the only thing a platform serves; images, posters and video renditions all live in media_variants (Images at Scale).
Vary: Accept: the response header telling caches a response depends on the request's Accept header, required for server-side format negotiation and best normalized into a few capability buckets (Images at Scale).
VCU (video coding unit): Google's Argos transcoding ASIC, deployed across tens of thousands of servers, giving 8x (H.264) to 20x (VP9) the throughput of software for offline two-pass encoding (Video on Demand).
Verified bot: a Twitch bot status that raises chat limits (7,500 messages per 30 seconds, 2,000 joins per 10 seconds) but doesn't exempt the bot from AutoMod or grant the chat bot badge (Live Chat & the Interactive Layer).
VP9: Google's video codec, which YouTube's VCU accelerates alongside H.264, absent from Apple's HLS authoring specification (Video on Demand).
Warm tier: millisecond-retrieval storage classes with a 30-day minimum and a retrieval fee (S3 Standard-IA, R2 Infrequent Access), for originals of older media that must still load on request (Storage, Dedup & Delivery Economics).
Warning (chat): a Twitch moderation action that blocks a user from chatting until they acknowledge it, with a reason of up to 500 characters, an early rung below timeouts and bans (Live Chat & the Interactive Layer).
Watch-time gate: running expensive encodes (per-title ladders, better codecs) only once a video's projected watch time crosses a threshold, so videos nobody watches never pay for them (Video on Demand).
WebP: an image format supported by every current major browser, the safe modern default for still and animated images (Images at Scale).
WHIP (WebRTC-HTTP Ingestion Protocol): RFC 9725 (March 2025): publishing a WebRTC stream with an SDP offer in an HTTP POST and a bearer token, allowing in-browser go-live with one audio and one video track per session (Live Streaming: Ingest & Transcoding).
Widevine: Google's DRM system, built on Common Encryption with a hardware root of trust provisioned at the factory (Video on Demand).