Trace
When a test fails, execution records help explain what happened.
A trace records the operations involved in an execution so their sequence and relationships can be inspected. In distributed tracing, individual operations are represented as spans with timing and other recorded attributes.
[OpenTelemetry]In practice
A support request has a model-call span and a database-lookup span. Their timings show that the lookup caused most of the delay.
[OpenTelemetry]A little deeper
A parent operation can contain child spans for downstream work. A trace helps locate where time or errors occurred; a successful operation status is not a judgment that an AI answer was useful. [OpenTelemetry]
A common mix-up
A trace is automatically a quality score.
Recorded operations describe execution; quality needs separate criteria. [OpenTelemetry]