X-Ray accounts recording traces at an uncapped rate above 10%

Observability priced by the trace

Tracing feels free until volume arrives: at $5.00 per million recorded traces, a high-throughput account sampling every third request records tens of millions of traces monthly for a bill nobody scrutinises. Sampling exists precisely because a representative fraction answers the same questions (latency distributions, dependency maps, error localisation) that exhaustive recording answers, at a twentieth of the cost.

The delta model, with both variables measured

The recommendation quantifies one move: reducing the configured fixed rate to 5%, AWS’s own default sampling rate. Reduced volume scales proportionally (target rate over current rate), and the saving is the trace-count difference times the per-million price. Both variables must be real: the configured rate comes from the account’s discovered sampling configuration, and the monthly volume from the trace-count series measured over at least 30 days; X-Ray publishes that series only when group-level metrics are enabled. Missing rate, missing volume, or thin coverage each abstain; the only fixed input is X-Ray’s published price, which is a rate card, not an estimate.

Account-scoped on purpose

X-Ray sampling rules match by service and host patterns, frequently a bare wildcard, so a high rate is not a property of any one instance or function; it applies account-and-region wide. The finding is therefore raised at the account grain, one per affected region, rather than scattered across every workload the rule happens to touch.

Read your sampling configuration

aws xray get-sampling-rules \
  --query 'SamplingRuleRecords[].SamplingRule.[RuleName,FixedRate,ReservoirSize,Priority]' \
  --output table

A FixedRate above 0.10 with no meaningful reservoir is the shape this rule prices.

Reduce without losing the traces that matter

Drop the fixed rate to about 5% and add a small reservoir of 1 request per second, so low-traffic services keep a floor of representative traces. Then protect the exception: keep a separate higher-priority rule sampling error and fault responses at 100%, because failure investigations want every bad trace and the bad traces are, by definition, rare. Watch recorded volume for a day and tune from measurement, which is the same standard the rule holds itself to.