feat(stt): enrichit les sondes (host_health, cluster_status, metrics) (#52)

- host_health : ajoute CPU%, disque /%, température CPU, uptime ; pour gpu-01
  un bloc GPU complet (température, utilisation, VRAM%, puissance via
  rocm_scraper) en plus de llama-server. Requêtes concurrentes (asyncio.gather).
- cluster_status : ajoute le nb de nœuds k8s Ready et une ligne CrashLoopBackOff
  (pods en boucle de crash) — complète le filtrage des pods terminés.
- metrics_block (contexte grafana) : ajoute le résumé GPU (util/temp/VRAM).
- Toutes les PromQL validées contre le Prometheus in-cluster.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
ALI YESILKAYA 2026-06-22 23:15:49 +02:00 committed by GitHub
parent 421ec77ef9
commit 0d09dec3dc
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 88 additions and 34 deletions

View file

@ -137,6 +137,10 @@ async def alerts_block(client: httpx.AsyncClient) -> str:
async def cluster_block(client: httpx.AsyncClient) -> str:
try:
nodes = await _prom_query(client, 'up{job=~"storage-01|gpu-01-node"}')
knodes = await _prom_query(
client, 'count(kube_node_status_condition{condition="Ready",status="true"} == 1)'
)
ready = await _prom_query(client, 'count(kube_pod_status_ready{condition="true"} == 1)')
# « Non prêt » = ready false ET pod ENCORE actif (Running/Pending). Sans le filtre de
# phase, les pods de CronJob terminés (Succeeded/Failed — ex. sacrifice-assign-renfort)
# comptent comme « non prêts » → fausse alarme. Le join exclut les pods terminés.
@ -145,20 +149,25 @@ async def cluster_block(client: httpx.AsyncClient) -> str:
'kube_pod_status_ready{condition="false"} == 1'
' and on(namespace,pod) kube_pod_status_phase{phase=~"Running|Pending"} == 1',
)
ready = await _prom_query(client, 'count(kube_pod_status_ready{condition="true"} == 1)')
crash = await _prom_query(
client, 'kube_pod_container_status_waiting_reason{reason="CrashLoopBackOff"} == 1'
)
except httpx.HTTPError:
return "Données indisponibles : Prometheus injoignable."
hosts_up = sum(1 for _, v in nodes if v == 1)
lines = [f"Hôtes hors-cluster joignables : {hosts_up}/{len(nodes) or 2} (storage-01, gpu-01)."]
if knodes:
lines.append(f"Nœuds k8s Ready : {int(knodes[0][1])}.")
if ready:
lines.append(f"Pods prêts (cluster) : {int(ready[0][1])}.")
lines.append(f"Pods prêts : {int(ready[0][1])}.")
if not_ready:
names = ", ".join(
m.get("pod", "?") for m, _ in not_ready[:6] if m.get("pod")
)
names = ", ".join(m.get("pod", "?") for m, _ in not_ready[:6] if m.get("pod"))
lines.append(f"Pods NON prêts : {names or len(not_ready)}.")
else:
lines.append("Tous les pods scrutés sont prêts.")
lines.append("Tous les pods actifs sont prêts.")
if crash:
cnames = ", ".join(sorted({m.get("pod", "?") for m, _ in crash[:6]}))
lines.append(f"⚠️ CrashLoopBackOff : {cnames}.")
return "\n".join(lines)
@ -167,11 +176,23 @@ async def metrics_block(client: httpx.AsyncClient) -> str:
llama = await _prom_query(client, 'up{job="llama-server-gpu"}')
targets_up = await _prom_query(client, 'count(up == 1)')
targets_all = await _prom_query(client, 'count(up)')
gpu_util = await _prom_query(client, 'rocm_gpu_utilization_percent')
gpu_temp = await _prom_query(client, 'rocm_gpu_temperature_celsius')
gpu_vram = await _prom_query(client, '100*rocm_vram_used_bytes/rocm_vram_total_bytes')
except httpx.HTTPError:
return "Données indisponibles : Prometheus injoignable."
lines = []
if llama:
lines.append("llama-server (GPU) : " + ("en ligne" if llama[0][1] == 1 else "hors ligne") + ".")
if gpu_util or gpu_temp or gpu_vram:
g = []
if gpu_util:
g.append(f"utilisation {round(gpu_util[0][1])}%")
if gpu_temp:
g.append(f"{round(gpu_temp[0][1])}°C")
if gpu_vram:
g.append(f"VRAM {round(gpu_vram[0][1])}%")
lines.append("GPU RX 6700XT : " + ", ".join(g) + ".")
if targets_up and targets_all:
lines.append(f"Cibles Prometheus UP : {int(targets_up[0][1])}/{int(targets_all[0][1])}.")
return "\n".join(lines) if lines else "Aucune métrique disponible."