A marquee live event is delivery’s final exam: an audience that arrives at once, watches the same thing, and leaves — with no second chance at any moment of it. The engineering is the live-streaming setup you already run; what the big night adds is arithmetic, contracts and rehearsal. Teams that look calm at kickoff did the multiplication months earlier and found the surprises on a Tuesday that did not matter.
The maths: concurrency × bitrate, honestly
Peak egress is one multiplication: peak concurrent viewers × average delivered bitrate. The honesty lives in the inputs. Concurrency: forecast from comparable events and marketing reach, then plan capacity at a multiple of the central estimate — events that matter are exactly the ones that beat their forecasts. Average bitrate: not your top rung but the audience-weighted mix across the ladder — big-screen-heavy audiences pull it up sharply; work it from real ABR distribution on comparable content. A worked example: one million concurrent at a 5 Mbps weighted average is 5 Tbps of edge egress — a number that belongs in provider conversations, not just spreadsheets — and at, say, 60% peak-hour duty across a three-hour event, on the order of four to five petabytes delivered. Add the request-rate dimension (manifest polls per viewer per segment interval — a million viewers polling every three seconds is a third of a million manifest requests per second) because platforms have request ceilings as well as bandwidth ones. The full worksheet with regional splits is in our capacity-planning piece.
Booking capacity: providers, regions, writing
Numbers in hand, book — do not assume. Every major CDN runs an event process: notify weeks out with date, window, per-region peak estimates and hostnames; ask what is reserved versus best-effort, per region, in writing; get the named event contact and the escalation path for the night; and settle the commercial treatment ahead, because event terabits can dwarf a monthly commit. Regional splits matter more than totals — 5 Tbps aggregate is comfortable; 3 of it concentrated in one country at 9 p.m. local is the actual question. At genuine scale, multi-CDN is standard event practice: capacity insurance, regional strength-matching, and a live failover option — provided the second platform is configured to parity and the steering rehearsed, per multi-CDN failover, and provided your security posture rides along, per the parity discipline you already run.
The delivery configuration for one big night
The event stream is your standard live setup with the dials moved toward robustness: segment TTLs and manifest TTLs verified per cache class; request collapsing and origin shield proven, because at event concurrency the origin must see per-object fan-in of one; redundant encoder-packager pipelines with a tested switch; and the ladder reviewed for the night — a capped top rung is a defensible event-day choice, since every Mbps of weighted average multiplies straight through section one’s equation. Decide latency posture deliberately: standard segments are the robust default; if the event demands low latency, stage it with the escape hatch from the low-latency guide rather than betting the night on the fragile mode. Pre-warm what can be pre-warmed (slates, pre-roll, the player itself), and freeze changes: config freezes days out, with the event build of everything already the production build.
The dress rehearsal: finding the surprise early
Rehearse twice. A load rehearsal: synthetic clients driving realistic manifest-poll and segment-fetch patterns at a meaningful fraction of forecast — coordinated with providers, since a good load test is indistinguishable from an attack — watching the numbers that predict the night: per-region hit ratio, origin fan-in, manifest p99, error rates. And a failure rehearsal: kill the primary encoder mid-stream, fail a region to the second CDN, break decisioning if you run SSAI, and confirm each failover behaves as designed with real players attached. The rehearsal’s output is a fix list and — equally valuable — measured switch times: knowing failover takes ninety seconds changes what you attempt live. A rehearsal that finds nothing was too gentle; run it again harder.
Event night: the war room and the ladder
Run the night like the incident it is pre-committed to almost being: one delivery lead with authority to act, provider contacts in a live channel, and a single dashboard — concurrency, per-region egress and hit ratio, origin fan-in, error and rebuffer rates, stitcher health if applicable. Decisions come off the pre-written ladder, not off debate: thresholds at which you cap the top rung, shift traffic between CDNs, fall back from low-latency to standard, or drop to slate — each with its trigger, owner and rollback, exactly in the runbook spirit of the DDoS playbook, because a viral audience and an attack look identical for the first five minutes. Afterwards, while the graphs are fresh: actual peak versus forecast, which ladder rungs carried the night, what the rehearsal missed — written down, because the next event’s section one starts from this event’s truth.