Move from flat probabilistic sampling to tail-based sampling in the OTel Collector. Keep every error and slow trace, cut health-check noise to 1%, and check that the Collector is doing what you think.

A flat 5% sampling rate sounds like a sensible trade between cost and coverage. It isn't. A random slice of your traffic is mostly the requests you'll never look at, and it throws away the rare ones you need at the same rate.

Metrics tell you how the crowd is doing. Logs tell you what one service saw. A trace tells you what one request went through, and at 2 a.m. that is usually the question you are asking.

At 1.5 million log events per second you cannot keep, batch, or ship everything the way you did at moderate scale. Content-aware sampling, OTel exporter tuning, Collector-side batching, and cheaper bytes on the wire.