Each label combination creates a measurement identity

In Prometheus, a metric name and its complete label set identify a series. If a metric has several bounded label dimensions, multiplying their possible value counts gives a full cross-product estimate. It is exact only when every combination actually exists and is exported. Correlations can reduce the count: a route may support only one method, and a region may contain only some services. Conversely, hidden target labels such as instance or environment can multiply a seemingly small application-level estimate. Inspect the final exported label set rather than counting only arguments in one instrumentation call.

Unbounded identifiers such as request IDs, raw URLs containing object IDs, and user emails can create an identity per event or person. Hashing such a value does not reduce its cardinality when distinct inputs remain distinct outputs. Prefer a bounded route template or operation category for recurring metrics, with event-specific investigation in appropriately controlled logs or traces. A trace identifier in a metric exemplar serves a different role from making it a series label. Define an owner and expected maximum for each new dimension so growth is detected before it consumes the telemetry budget.

A classic histogram expands one label combination into multiple bucket series plus its count and sum. Include the positive-infinity bucket in the chosen bucket count so it is not forgotten or counted twice. A native histogram has a different representation and cost profile, so reuse neither the classic component count nor a guessed fixed byte cost blindly. Storage cost also includes label indexes, active head state, write-ahead logging, compression, retention, replication, and query work. Per-sample payload is only one part of the budget.

Active cardinality and churn answer different questions. A workload can keep roughly the same active count while rapidly replacing instance or deployment identities, creating substantial historical indexes and ingestion work. Long retention accumulates many identities even when each lives briefly. Measure new-series rate and historical unique sets alongside current active series. Aggregating after ingestion can make a dashboard cheaper to query without refunding the cost already paid to create and store the original identities. Decide whether dimensions should be removed or normalized before the expensive boundary when the use case permits it.

Before adding a label, write the aggregation or alert it enables and estimate its full combination count. Test production-like route counts, instance counts, and deployment churn in a controlled environment. Enforce SDK or collector cardinality controls according to their actual overflow semantics, then expose overflow or dropped attributes as evidence rather than silently interpreting the reduced metric as complete. A limit can preserve process health while changing which distinctions remain observable. Include those consequences in dashboard and alert reviews so the query still measures its claimed population.

The quadratic lab demonstrates only one cross-product pattern. Reset, set both coefficient a=1 and position x=1, predict y=1, then Play or Step. Reset and re-enter a=1 and x=2, predict y=4. Interpret x as a normalized size of each of two fully populated independent label sets, so their product grows quadratically. The controls are limited to this fixed y=ax² model and do not enumerate label sets, simulate churn, or estimate storage bytes. The actual series budget needs the explicit dimensions and component counts above.