Metric semantics

Each metric is classified by how its raw samples are aggregated over a time range. The classification comes from hpa-agent-calypso‘s MetricHelper call sites, not from a guess based on the metric name - with one documented exception: the two SQL backend-tracer metrics (Average Result Processing Time (ms), Average Result Rows Processed) are Introscope’s own built-in JDBC instrumentation, not a MetricHelper call site, so their classification is operator-verified rather than sourced from agent code.

• GAUGE: point-in-time reading. Average or max over a window is meaningful; sum is not. MetricHelper: sendMetricAverage / sendMetricPercentage.
• DELTA: per-interval delta, reset at every harvest. Sum over a window is meaningful. MetricHelper: sendMetricInterval.
• STATE_COUNTER: retained running value that moves in discrete increments/decrements - neither a plain gauge nor a monotonic-since-start counter. Aggregated by max over a window; watched for restart resets. MetricHelper: sendMetricCumulator / sendMetricLongCumulator - despite the method name, this is not the same as CUMULATIVE below.
• CUMULATIVE: monotonic counter since process start. Aggregated with an end-minus-begin delta; watched for restart resets. No MetricHelper method currently maps here - reserved for a confirmed case once one is found.
• UNKNOWN: not found in the classification table - typically a built-in/broker-native metric collected outside MetricHelper, or a fully dynamic name with no fixed literal component. The series shape is inspected before any aggregation is assumed.