LLM Ops on VCF

Watching a vLLM server from VCF Operations when there is no adapter for it. Part 1 decides which metrics are worth shipping (not the histogram), part 2 makes the pipeline watch itself so you hear when it stops. Start with part 1.

Telegraf runs the script every 60 s; the script's stdout is one integer; the integer is the health of the integration

vLLM metrics into VCF Operations, part 2: monitor the monitor

Part 1’s design only helps if it runs every minute, for ever, and says when it stops. Self-monitoring through Telegraf, batched pushes, the four alerts for day one, and an offline test mode.

Prometheus histogram buckets in; flat P50/P95/P99 gauges out

vLLM metrics into VCF Operations, part 1: don't ship the histogram

vLLM exposes about 200 Prometheus series per model, and pushing them raw into VCF Operations is a cardinality bomb. A pipeline that computes the useful numbers client-side and drops what can’t be graphed.