# Spend telemetry budget on evidence you can use

High-volume traces and unbounded labels can become expensive without improving reliability decisions. Separate essential measurement from sampled diagnosis.

By Cobnex editorial. Published 2026-09-10. Updated 2026-09-11.

## Identify the measurement that must remain complete

For an operation-based indicator, determine which counts or durable outcomes are required to calculate the result accurately. Do not casually sample them and then report an exact population ratio.

Detailed traces may be sampled while aggregate counters remain complete, depending on the design. Explain the relationship so operators know why a counted failure may not have a full trace.

Keep the telemetry pipeline's own loss or delay visible.

## Estimate label growth

A metric with dimensions for operation, status and region can remain manageable when each has a small known set of values. Adding request identifiers can create a new series for every request.

Measure actual cardinality and retention cost before broad rollout. The number of metric names alone does not describe the storage and query workload.

Use searchable diagnostic records for high-cardinality identifiers instead of forcing them into every chart.

## Choose sampling for the question

Sample traces to understand latency and dependency behaviour, with appropriate treatment of errors and rare operations. A simple uniform sample may miss the unusual path the team needs to investigate.

Avoid claiming that a small sample proves the absence of rare failures. Use controlled tests and complete outcome measures where the consequence requires stronger evidence.

For a real budget, use current pricing and measured event sizes. Compression and indexing can make raw payload estimates differ substantially from the bill.

## Review operational value

Identify dashboards and fields nobody uses, but check incident and reporting needs before removing them. Low routine usage does not always mean low value during recovery.

Prioritise telemetry that connects a user-impact signal to a diagnosable operation. Collecting every payload is not a substitute for well-chosen identifiers and outcomes.

The cost goal is a measurement and investigation system the team can sustain. Reducing spend by losing the denominator or hiding failures defeats the purpose, while unlimited collection can create both expense and unnecessary data exposure.

## Sources

- [Google SRE: implementing service objectives](https://sre.google/workbook/implementing-slos/)