Q: A team wants very accurate latency percentiles but their classic Prometheus histograms create too many bucket time series. What options do you discuss?

Classic histograms multiply time series by every bucket and every label combination. If a metric has 20 buckets and many labels, cost and memory grow quickly.

• Fix bucket boundaries: Use fewer buckets that match real SLO boundaries, such as 100ms, 300ms, 1s, and 3s, instead of many generic buckets.
• Reduce labels: Remove high-cardinality labels from histogram metrics, especially user IDs, raw paths, tenant IDs, and pod UIDs.
• Evaluate native histograms: Native histograms can represent distributions more compactly and with dynamic buckets, but the team must verify support across Prometheus, remote storage, Grafana, alert rules, and client libraries.

I would discuss three options: The decision is a tradeoff: enough precision for SLOs, without turning latency measurement itself into the largest source of telemetry cost.