Telemetry Sampling
Keeping a subset of traces or events to bound observability cost, chosen so the ones that matter survive.
At scale, retaining every trace costs more than the compute being traced. Sampling is not optional; the question is which strategy.
Head-based decides at the start of the request, usually at a fixed percentage. Cheap and simple, and it discards slow and failed traces at the same rate as everything else — so the traces you actually want are mostly gone.
Tail-based buffers all spans of a trace and decides after it completes, so it can keep 100% of errors and slow requests and 1% of the fast successful ones. This is what you want, and it costs a collector tier that holds spans in memory until the trace finishes.
Whatever the strategy, sampling must be consistent across services — a decision made at the edge and propagated — or traces arrive with holes in the middle, which is worse than not having them.