Question

Design a general-purpose telemetry collection pipeline (OpenTelemetry-style) that receives metrics, logs, and traces from 100k agents totaling 4GB/s, normalizes/enriches them, and fans them out to multiple backends (a metrics TSDB, a log store, a trace store, and a cold archive in object storage). One backend (the trace store) is regularly slow or briefly down. The pipeline must not lose data when a backend is degraded and must not let a slow backend stall ingestion of the others. Design the stages, buffering, and per-backend isolation.