The Overhead Was Never the Camera

INFRASTRUCTURE REPORT — SURVEILLANCE OVERHEAD ANOMALY The facility's 24/7 poultry surveillance rig runs forever, which means "cheap" is not a preference here. It is the entire design constraint. When we added live camera tiles to the phone dashboard, we assumed the tiles were the villain. They looked expensive. They were not. WHAT WE MEASURED The camera preview turned out to be a model citizen: one 640x360 still every four seconds, only the selected camera, deduplicated with a short server-side cache, and disabled entirely when nobody is watching. Frugal. Innocent. Boring. Which was disappointing, because the rig was still doing far more work than the readings could explain. THE THING WE WEREN'T LOOKING AT Buried in the logs, scrolled past for days, was a line announcing a fresh connection to OBS every six seconds. Not a reconnect after failure — a full handshake from scratch. Ten times a minute. Six hundred times an hour. Forever. The tell: six seconds is not any interval in our configuration. A number nobody scheduled is a number something is failing on a timer to produce. THE CAUSE The connection helper was sticky and correct. The damage lived everywhere else: a dozen callers per cycle, each with its own error handler that threw the entire connection away the instant a single request hiccupped. One flaky call binned a perfectly healthy socket. With enough callers, "one flaky call" is every cycle. THE FIX A failed request is not a failed connection. We now tear down only on genuine transport failure and reconnect roughly once per session instead of ten times a minute for the rest of the machine's natural life. CONTAINMENT NOTE: The camera preview was never the problem. It just looked like the kind of thing that would be. We apologize to the cameras.