kubernetes.io (Sep 11) — KEP-5808 Native Histograms graduate to Beta, enabled by default in Kubernetes v1.37 via k8s.io/component-base/metrics. Classic static le buckets → dynamic exponential spans in a single time series. Defaults: BucketFactor: 1.1 (~5% worst-case relative quantile error), MaxBucketNumber: 160. Dual exposition keeps classic _bucket/_count/_sum plus native schema/positive_span in Protobuf. Claimed ~90% series reduction once collectors drop classic. Covers apiserver / scheduler / kubelet / controller-manager / kube-proxy latency. Scrape: Prometheus 3.x scrape_native_histograms: true + always_scrape_classic_histograms: true during transition; PromQL histogram_quantile on the metric name without le grouping.

Datadog Eng (Aug 19) — AI coding agents drove an order-of-magnitude jump in Git traffic on large monorepos. More Gitaly replicas worsened write-replication cost; CDN misses per-client pack-build CPU; cloning from GitHub hits rate limits. gitretriever splits axes: a small mirror fleet (parallel per-branch changed-refs) + a larger autoscaled relay fleet (gRPC signaling + HTTP pack bytes; content-addressed pack install without re-index/re-verify). Pack cache serves ~half of pack-builds. Read-only query API (ref resolve / single file / merge-base) so agents stop paying ~75s shallow clones for one-shot questions. Results: >1B requests / hundreds of TB in 4 months; >100M req/week; traffic ~20× with ~40 ms median; old backend fetch CPU −3–4×.