Kubernetes v1.37: Native Histograms Graduates to Beta

If you’ve ever tried to troubleshoot performance issues in Kubernetes, you know the pain of piecing together metrics to understand what’s actually happening. You might see average latency is fine, but suspect something’s wrong with outliers. Or you’re drowning in cardinality from high-dimensional metrics that make your monitoring system groan. Kubernetes v1.37 addresses this head-on by graduating native histogram support to Beta—and enabling it by default. This is a meaningful shift in how you’ll be able to observe your clusters.

Native histograms, introduced as Alpha in v1.36, fundamentally change how Prometheus-compatible metrics record distribution data. Instead of the traditional approach where each latency bucket becomes a separate metric series, native histograms store histogram data as a single, structured object. This means you get granular percentile information (p50, p95, p99, even p99.9) without the cardinality explosion that typically comes with detailed observability. Technically, the kubelet and other Kubernetes components now emit these histograms for key metrics like API server request latencies and container runtime operations. Your existing monitoring tools—Prometheus, Grafana, Datadog—will automatically recognize and work with these when you upgrade, with no configuration changes needed.

The practical benefits show up immediately in real scenarios. Consider a microservices platform where you’re tracking HTTP request duration across hundreds of endpoints. With traditional metrics, setting up per-endpoint latency buckets explodes your cardinality and can actually degrade Prometheus performance. With native histograms, you capture rich distribution data while keeping cardinality flat. Another common case: detecting resource contention. If your average CPU looks fine but p99 latency spikes during peak hours, native histograms make that relationship visible without requiring custom instrumentation. You’re also reducing storage overhead—histograms take less disk space than equivalent bucket series.

Since this is now Beta and enabled by default, you should expect it to be production-ready in the next release cycle. If you’re running v1.37, you’ll want to verify your monitoring stack handles native histograms correctly—most modern versions do, but older Prometheus instances might need a minor update. The feature is backward compatible, so you won’t break anything by upgrading. Start paying attention to these metrics in your dashboards; they’ll give you better signal about what’s actually happening in your cluster without the noise.