One open port, fifty sites, and every product limit in writing

A network team never opens with what the camera detects: it opens with which ports have to be opened. IRIS listens on fourteen, but only one crosses the firewall, 443. The other thirteen listen inside the machine only, and that is not buried in the program: it is a policy you can see and edit from the web interface. For the customer that means thirteen exceptions nobody has to request, document, justify to an auditor or review every year, and thirteen doors nobody will leave open long after the component that needed them is gone. Across a distributed installation it also means thirteen fewer negotiations with each branch, and branches tend to have their own rules; anyone who has ever fought a network team knows exactly what that is worth. And when something goes wrong, the entire network check comes down to one question: does 443 get through.

A single installation runs servers spread across sites, countries and time zones, and they are handled as if they were one. Organisations nest to any depth — group, country, region, customer, site — and access is delegated accordingly: a site manager touches their own site, the centre sees everything. Every organisation, every server and every camera carries its own time zone, so comparing the rush hour of four countries never skews a series. And here is what matters most: this is not a central brain that surveillance depends on. Each site captures its cameras, detects, applies its rules, records and drives its relays locally, and only events and measurements travel to the centre, never frames. If the link drops, the site keeps working and publishes what it stored the moment the link is back.

The product states more than one hundred thousand servers and more than five million cameras, but the number on its own proves nothing: anyone can write it. What holds it up is that all eight product limits are published, each with the unit it is shared out by, and that three are solved by adding machines while five are a hard wall you have to respect. One of the shareable ones: sixty-four cameras fit per capture process, so five thousand cameras are seventy-nine processes spread over as many machines as it takes. That is an engineering answer, and publishing your own walls is the strongest proof of confidence in a product there is: anyone can write scalable, almost nobody publishes their eight limits. And the arithmetic does not live in a sales spreadsheet: there is a calculator inside the product that asks twelve questions, states what hardware is needed, flags which resource runs out first, and declares a sizing unfeasible — not risky — when it crosses a limit that cannot be shared out. Every number it returns carries where it came from, and it says in writing what has not yet been measured in that particular installation.

Whatever the system measures can be taken out of it. Ask it for its data catalogue and it returns everything it knows how to calculate and everything you can group by, names already translated: the same vocabulary the screens use, so there is no second-class interface for outsiders. The limits are stated up front, which is what anyone writing code against it appreciates: a date range is mandatory, a response carries at most fifty thousand rows, and a query that is too broad is rejected with the reason spelled out rather than quietly degrading. Outwards, the system talks to cameras from any manufacturer over standard protocols, drives barrier, turnstile and siren relays, alerts by email and messaging, calls external systems and exports to a spreadsheet. And all of it runs under the same user identity and over the same port: there is no separate user directory for integrations.

All eight limits, one by one, because that is exactly what an integrator will ask in a demonstration. Three share out: 64 cameras per capture process, 8,192 live connections at the entry point and the 1,000 a single instance serves at once; all three are solved by adding processes or entry points. The other five are a hard wall: 64 consumers per camera — every recorder, every detector and every viewer takes one — 16,384 temporary ports from the operating system, which is not even our limit and is the reason for the previous one, 8 frames per batch in the detection engine, 600 frames held per consumer, the multiplier that sends memory through the roof as soon as resolution goes up, and one graphics card per process, because a model is not split in half between two: either it fits whole or you need two copies. Those five are not fixed by buying more machines, and that is exactly why they are written down instead of hidden. With this list and the calculator, an integrator can size an installation without calling us.

The analytics is not a separate module copying data: it is the same source the screens use, aggregated. Seven metrics, sixteen dimensions to group and filter by, twenty-seven visualisation types and five export formats generated in the browser itself, offline. And here is the detail that gives away someone who has actually done this: the metrics that cannot be re-aggregated without lying — distinct objects, dwell time and time-weighted average occupancy — are marked as such and served from detail, because a sum of sums means nothing. The programming interface returns the same thing under the same rules: a date range is mandatory, pagination does not degrade as you advance, and a permissions error, a vocabulary error and a range error are never confused with one another. And the scope of what can be seen is resolved on the server from the permissions of whoever is asking: a request can narrow what it sees, never widen it.