Progressive MP4 and byte ranges
Progressive MP4 remains useful for broad compatibility and simple workflows, but tube behavior makes correct 206 Partial Content handling essential. Viewers seek, replay, and abandon. The CDN must cache and combine useful byte ranges without repeatedly pulling full files from origin.
Range caching is subtler than it looks. RFC 9111 permits a cache to store an incomplete 200 response or a 206 as an incomplete entry and to combine several ranged responses into one stored representation, so a cache can legitimately hold a partial object. That is only safe when the object carries a strong validator such as a stable ETag or Last-Modified; without one, a POP can stitch or retain a fragment it should never serve. Strong validators are a prerequisite, not an optimization.
Design for skipping, not sequential playback
Tube traffic does not stream front to back. In public measurements of a large adult portal, roughly three-quarters of video requests involved at least one non-adjacent byte-range skip, and most video fetches returned fewer than ten chunks. Abandonment is driven by early skipping, not by viewers finishing. Contiguous read-ahead buffering and "fetch the first N segments" strategies therefore spend origin and edge bytes on data viewers never watch. Predictive prefetch and end-to-end range serving matter more than aggressive buffering. Homepage and search tiles load only a small fraction of the bytes they request, so a per-referrer prefetch policy that defers segment prefetch on those entry points cuts egress on the highest-abandonment traffic without touching the video-page funnel.
Adaptive HLS and DASH
Adaptive streaming can provide a better experience across mobile networks and variable bandwidth. It also adds packaging, manifests, more objects, encoding-ladder decisions, and player complexity. Use it where adaptive bitrate, content protection, or device behavior justifies the operational cost.
Measure bytes rather than vanity ratios
A high request-hit ratio can hide expensive origin traffic if small thumbnails hit cache while large video ranges miss. Track byte-hit ratio, origin bytes, shield efficiency, range behavior, startup time, seek delay, rebuffering, playback errors, and delivered cost by region and ASN.
The two ratios diverge sharply on adult portals. In a public CDN trace of a major adult site, images were about 80% of requests but only about 37% of bytes, while video was about 63% of bytes but only about 20% of requests. Optimizing to request-hit ratio can therefore leave origin egress, usually the dominant cost, almost untouched. Byte-offload (origin bytes avoided ÷ bytes served) is the cache KPI that actually tracks the bill, and shielding distorts classic hit-ratio accounting enough that it cannot be used as a proxy.
Tier the cache for a deep catalogue
Leading adult platforms operate with multi-petabyte hot-cache footprints. Use edge tiers for high-demand ranges, deeper cache capacity for broader catalogue retention, and dedicated origin-shield clusters engineered for extended adult libraries. This keeps a viral title from evicting the useful long tail and prevents geographically scattered misses from repeatedly reaching storage.
Popularity is also unusually concentrated, which is why edge topology beats broad catalogue caching: the top decile of videos can account for the large majority of accesses, and a simulated city-wide edge cache has been shown to remove more than 90% of backhaul in dense metros. Put the engineering into hot-set pre-warm and tier placement, not into trying to cache everything.
The thumbnail plane can take down the filesystem
Images dominate request count even though they are a minority of bytes, so the failure mode is not bandwidth; it is inode and object-count exhaustion. Millions of thumbnails, hover sprites, and preview frames can fill a filesystem's inode table while df -h still shows free space, so capacity monitoring has to watch df -i, object counts, and processing queue depth as first-class signals. Keep the thumbnail and video planes on separate storage and cache paths so one cannot starve the other.
Do not let archive storage break instant playback
The long tail is exactly what operators want to archive, but archive classes impose minimum storage durations, minimum billable object sizes, and retrieval delays: archive objects often must be explicitly restored before they can be served. Tiering cold renditions into archive therefore trades storage cost for playback latency and can cost more than a warmer class if the "cold" tail is requested more than expected. Tier by rendition (cold low-bitrate and preview assets only) and keep anything reachable on demand in a low-latency class.
Abandoned uploads are a silent storage line item
Resumable multipart uploads have no expiry until they are completed or explicitly aborted, and every part is billed for its lifetime. Creator, studio, and partner uploads that die halfway quietly accrue storage spend and can fail integrity checks in confusing ways if part numbering is not consecutive. Wire an abort-incomplete-multipart lifecycle rule and idempotent part numbering into the ingest pipeline from day one.
Read the adult video CDN architecture guide