The streaming format war resolved the way infrastructure wars usually do: quietly, in the container layer. HLS and DASH persist as manifest dialects, CMAF unified the segments beneath them, and the remaining decisions are practical rather than tribal.
What actually differs now
HLS is Apple’s lineage: m3u8 playlists, mandatory for reach into Apple platforms, extended over the years with everything DASH pioneered. DASH is the ISO standard: XML manifests (MPD), template-based addressing, codec-agnostic by design, historically the choice for DRM flexibility. With CMAF, both can reference identical fMP4 segments, so the war shrank to manifest generation, player support matrices, and DRM plumbing.
The CMAF consolidation
One encode, one segment set, two thin manifests: that is the modern packaging target. It halves storage, doubles effective cache efficiency (both audiences hit the same segment objects), and simplifies the origin. The residual complexity is DRM: CENC common encryption with mode differences (historically CBCS for Apple’s FairPlay, CTR elsewhere) meant some estates still dual-encrypt; mode convergence has steadily reduced that tax, and greenfield builds should design for single-encrypt from day one.
A cache-arithmetic example shows why unification matters commercially: a dual-format library stores every rendition twice and splits audience demand across two object sets, so each set’s long tail is colder, hit ratios sag, and origin egress rises, the multi-CDN fragmentation effect, self-inflicted inside one CDN. Estates that completed CMAF unification report the boring, excellent trifecta: storage roughly halved, measurable hit-ratio gains on the long tail, and one packaging pipeline to operate. The strategic lesson generalizes across delivery: object-set unification is cache efficiency, and every layer that forks your objects, formats, encodings, unnecessary variants, quietly forks your economics.
Delivery consequences
Manifests and segments want opposite cache treatment: segments are immutable, cache-forever objects; live manifests mutate every few seconds and need short TTLs with stale-while-revalidate to avoid stampedes at refresh boundaries. Cache-key hygiene matters doubly because token auth rides these URLs (our token article applies verbatim). And segment duration remains the master latency knob: shorter segments cut glass-to-glass delay and multiply request rates, which is a pricing conversation on request-priced networks.
In practice
Greenfield: CMAF single-encode, HLS+DASH manifests generated at packaging, segments immutable with long TTLs, manifests short with SWR, tokens path-scoped per stream. Legacy dual-format estates: the migration pays for itself in storage and hit ratio, and it is a packaging-layer project, not a player rewrite. Either way, put manifest QPS and segment hit ratio on separate dashboard lines; they fail differently.
Streaming architecture reviews here start at the packager, because that is where the delivery bill is actually written.