Design Netflix, Hotstar / Amazon Prime Video - OTT Streaming Platform

Before drawing, I ask: on-demand movies, live sports, or both? Either Hotstar or Prime Video can represent either workload. The agreed scope, not the brand name, decides the architecture.

For VOD I emphasize asset processing, sessions and global delivery. For live sports I emphasize burst admission, fresh segments, redundant ingest and glass-to-glass delay.

I will support catalog lookup, authorized adaptive playback, VOD seek/resume and a bounded live DVR window. Video bytes bypass the catalog API. A subscription/geography check cannot be replaced with possession of an object URL.

A manifest must reference ready segments with the correct asset/key version. Sessions have explicit expiry/renewal behavior. Billing itself and rights acquisition are external to this scope.

Assume 4 Mb/s average delivered bitrate, not the top ladder bitrate. One million concurrent viewers is a bandwidth problem even if session creation is only thousands of requests per second.

Assume five million starts over ten minutes. This average hides sharper spikes, so I would load-test burst envelopes, license capacity and admission behavior.

A title differs from its immutable encoded asset version. Catalog metadata describes it, a manifest identifies playable outputs, and entitlements determine who can watch in which market/time window.

For screen limits I use authoritative per-account lease admission. A cached increment is unsafe under concurrent starts and leaks when clients disappear. Explicit release helps but expiry is required.

The player receives a scoped grant and manifest after authorization. The CDN validates supported authorization without routing every segment through my API. The DRM license exchange is a separate protected operation; signed URLs alone are not a DRM system.

For Prime Video as an on-demand prompt, I begin with this path. Upload an original, validate/transcode/package it asynchronously, and publish a playable version only when required outputs exist. The viewer fetches through a CDN.

A catalog/playback application, transactional metadata and queue-driven workers are a sufficient baseline. Separate services later according to scaling and ownership boundaries.

For a Hotstar match or Prime Video live sports, there is no completed movie to process before viewing. I need a continuously advancing stream with redundant input, encoding and packaging. Healthy HTTP endpoints do not prove the broadcast is advancing.

Smaller chunks can reduce latency but increase overhead and sensitivity to jitter. I budget capture/encode, packaging, network and player buffering rather than promising zero delay and perfect resilience.

Two devices may reach different API instances, but simultaneous admission must serialize through authoritative account lease state. A lost startup response replays the same operation instead of consuming another slot.

Once granted, media stays on the delivery path. Existing unexpired sessions can survive a catalog outage according to policy, but grant expiry and rights revocation still bound access.

At match start, authentication, lease creation, DRM and manifests spike before stable watching begins. Pre-scale and rehearse those dependencies together. Reserve capacity for existing viewers and use bounded retry-after with jitter.

Recommendations, thumbnails or analytics sampling can degrade. Entitlements cannot simply be bypassed and manifests cannot reference missing content.

I would disconnect an encoder, serve a stale manifest, remove a segment, slow licensing and drop a region. Success means a known player outcome within a declared recovery budget, not an HTTP 200.

Track startup percentiles, rebuffer/watch-time ratio, fatal errors, live-edge lag, segment freshness, origin offload and license success by device, region, rendition and CDN.

For “Prime Video movies,” my core story is immutable VOD preparation, authorized sessions, adaptive delivery and resume. For “Hotstar during a major match,” I spend the remaining time on ingest redundancy, freshness, bursts and player recovery. Either product may need both; I clarify rather than infer.

Senior depth connects control and byte paths, bandwidth, leases and user-visible failures. Staff depth adds regional budgets, device/codec rollouts, rights dependencies, CDN economics and scheduled-event rehearsals.