NVLink domains per-node GPU islands · grouped tiles auto 30s ↻ refresh

GPU trends last 3h · per node

All GPUs util · framebuffer · thermals · NVLink · workload

Per-node allocation requests · limits · active usage against allocatable — bars share each node's track so they read against the same 100% auto 30s ↻ refresh

Node trends last 3h · per node · shown nodes only

All shown nodes cores · bytes · % of allocatable · pods

Saved queries (this browser)

Recent queries

Alerts all states firing pending ↻ refresh

Rules what is monitored, and the exact expr behind every alert all states firing pending inactive ↻ refresh

What the MCP tools are good for all read-only · same API this page uses

A Kubernetes API (the K8s MCP) sees state: pods, events, logs, instantaneous top. Prometheus sees behavior over time: rates, trends, quantiles, alert state. Together they complete an ops agent — the K8s MCP answers what happened, Prometheus MCP answers since when, how fast, and why. Click any example to load it in the Query tab.

Triage pairing — the K8s MCP + Prometheus MCP workflow

1. K8s MCP: describe pod → lastState: OOMKilled, restart count
2. K8s MCP: logs --previous → what it was doing
3. Prometheus MCP: prom_query_range on container_memory_working_set_bytes{pod=…}[6h] → sawtooth = leak, plateau-at-limit = undersized
4. Evidence-backed verdict instead of a guess

LLM-safe shaping why tool answers stay small

Range results are downsampled to evenly-spaced points (the shape — ramp/spike/sawtooth — survives), floats are rounded to 4 significant digits, and series counts are capped with a truncation footer. Caps are env-tunable (PROM_MAX_SERIES, PROM_MAX_POINTS, PROM_MAX_LABEL_VALUES), and this page reuses the exact same caps and client as the tools.