rock
1f3439c9e0
New Grafana dashboard (llm-observability) with 27 panels across 5 rows:
TTFT, inter-token latency, batching/scheduling, KV cache, throughput.
Both metric sources are shown side by side because they answer different
questions and neither is complete alone:
vLLM 0.11.0 (reasoning) — engine truth: kv_cache_usage_perc,
num_requests_running/waiting, num_preemptions_total, prefill vs
decode time, prefix cache, per-request histogram quantiles.
TEI 1.8.2 (embeddings, reranker) — te_batch_next_size, te_queue_size,
te_embed_duration, te_embed_queue_duration.
api-gateway — the ONLY per-request latency source covering Ollama
models (ornith:35b, qwen2.5:3b-instruct), which expose no metrics.
It is derived from the SSE byte stream, so it is streaming-only and
counts SSE chunks rather than true tokens. vLLM's own histograms are
authoritative for reasoning.
All 44 panel queries were executed against live Prometheus before commit —
all return success. vLLM/TEI panels have live series; gateway panels show
no data until the netpol fix in homelab-frontend PR #62 merges.
Also fixes a dead alert. alerts/cag-kv-cache-rules.yaml derived
vllm:prefix_cache_hit_rate from vllm:cache_query_hit_total and
vllm:cache_query_miss_total. Those metric names do not exist in vLLM
0.11.0 — the real series are vllm:prefix_cache_hits_total and
vllm:prefix_cache_queries_total. The recording rule produced no data, so
VLLMPrefixCacheHitRateLow could never fire.
Not included: real GPU hardware metrics (SM utilisation, framebuffer,
power, temperature). Only nvidia-device-plugin is deployed, which exposes
none of them; DCGM exporter is required. vllm:kv_cache_usage_perc and
preemption rate are the only GPU-pressure proxies available today.
77 lines
2.9 KiB
YAML
77 lines
2.9 KiB
YAML
apiVersion: monitoring.coreos.com/v1
|
|
kind: PrometheusRule
|
|
metadata:
|
|
name: cag-kv-cache-rules
|
|
namespace: monitoring
|
|
labels:
|
|
release: prometheus
|
|
spec:
|
|
groups:
|
|
# Recording rules: derive vLLM prefix cache hit rate from raw counters.
|
|
# Metric names verified against live vLLM 0.11.0 scrape (2026-09-29).
|
|
# The previous names (vllm:cache_query_hit_total / vllm:cache_query_miss_total)
|
|
# do not exist in this version — vLLM exposes prefix_cache_hits_total and
|
|
# prefix_cache_queries_total — so this rule produced no data and
|
|
# VLLMPrefixCacheHitRateLow could never fire.
|
|
- name: vllm.recording
|
|
rules:
|
|
- record: vllm:prefix_cache_hit_rate
|
|
expr: |
|
|
rate(vllm:prefix_cache_hits_total[5m])
|
|
/
|
|
clamp_min(
|
|
rate(vllm:prefix_cache_queries_total[5m]),
|
|
0.001
|
|
)
|
|
labels:
|
|
job: "reasoning-predictor"
|
|
|
|
# Alert rules: CAG pipeline health
|
|
- name: cag.alerts
|
|
rules:
|
|
# KV warmup consistently failing — cache never gets warm
|
|
- alert: KVWarmupFailureHigh
|
|
expr: |
|
|
rate(memory_kv_cache_warmup_error_total[5m]) > 0.01
|
|
for: 5m
|
|
labels:
|
|
severity: warning
|
|
team: poimen
|
|
annotations:
|
|
summary: "KV warmup errors elevated ({{ $value | humanize }}/s)"
|
|
description: |
|
|
The KV cache warmer is failing to reach vLLM at a rate > 0.01/s
|
|
for 5 consecutive minutes. The prefix KV cache is not being warmed;
|
|
agents will see full prefill latency on every call.
|
|
Check reasoning-predictor health and poimen-memory NetworkPolicy.
|
|
|
|
# Warmup misses: vLLM reachable but not caching
|
|
- alert: KVWarmupMissHigh
|
|
expr: |
|
|
rate(memory_kv_cache_warmup_miss_total[5m]) > 0.01
|
|
for: 5m
|
|
labels:
|
|
severity: warning
|
|
team: poimen
|
|
annotations:
|
|
summary: "KV warmup cache misses elevated ({{ $value | humanize }}/s)"
|
|
description: |
|
|
vLLM is reachable but cached_tokens=0 on warmup responses.
|
|
Possible causes: prefix caching disabled in vLLM, max-num-seqs
|
|
saturated during warmup, or context format changed.
|
|
|
|
# Prefix cache hit rate low after warmup should be established
|
|
- alert: VLLMPrefixCacheHitRateLow
|
|
expr: |
|
|
vllm:prefix_cache_hit_rate < 0.70
|
|
for: 10m
|
|
labels:
|
|
severity: warning
|
|
team: poimen
|
|
annotations:
|
|
summary: "vLLM prefix cache hit rate low ({{ $value | humanizePercentage }})"
|
|
description: |
|
|
vLLM prefix cache hit rate below 70% for 10 minutes.
|
|
Expected >= 85% in steady state with CAG warming active.
|
|
Check KVWarmupFailureHigh / KVWarmupMissHigh alerts for root cause.
|