Monitoring
The tenant's inference metrics (RPS, error rate, p95 latency, data latency), training queue and GPU trends, and per-model performance and input distribution.
Required permission: View
This screen shows how the models in operation are doing. The top has inference metrics for the whole tenant (Prometheus), the middle the training queue and GPU usage trends, and the bottom the performance and input distribution of the model you choose. Every member sees it, and it is read-only.
| # | Area | Description |
|---|---|---|
| ① | Filters | Model (All models (모델 전체) or one model), edge (All edges (엣지 전체) or one device; only in tenants with the Edge feature on), range (1h · 24h · 7d · 30d, default 24h). The model list holds only one page of recent models and tells you how many more exist. Filters stay in the address, so a link shares them. |
| ② | Operational metrics | RPS (RPS) (requests per second) · Error rate (에러율) (%) · p95 latency (p95 지연) (ms) · Data latency (데이터 지연) (ms). Computed from inference requests that pass through the platform, refreshed every 30 seconds. The line below explains why a value is empty. Data latency (데이터 지연) is always — because nothing records this value yet. |
| ③ | Training queue (학습 큐) | Trend of this tenant's running and queued training runs, regardless of the model and edge filters. Queued runs wait for the server-wide concurrent training limit (1 by default, shared by all tenants). GPU slots are not a queueing reason; they are checked at submission. |
| ④ | GPU slots (GPU 슬롯) | Trend of wired slots and occupied slots for the whole tenant. Data accumulates from when the platform started recording. |
| ⑤ | Per-model card area | Once you choose a model, the model performance and input distribution cards appear. |
If the alerts feature is on and a KPI alert is firing, a KPI alerts firing (발생 중인 KPI 알림) line appears above ②. Click it to go to that alert.
When you choose a model
| # | Card | Description |
|---|---|---|
| ① | Model performance (모델 성능) | Trends of accuracy, detection rate, mean confidence and low-confidence ratio. Hover over a point to see its sample count. Always covers the whole period, regardless of the range filter (the card says "Range not applied (기간 선택 미적용)"). |
| ② | Input distribution / drift | Distribution per input feature, with PSI and KS values (fixed 7-day baseline). If there are many features, only 4 are shown. Drift verdicts and policies are handled on the Drift screen. |
Both cards read per-model aggregates in the platform DB. Nothing fills these aggregates yet (the screen says "The aggregation job is not ready yet (집계 작업은 아직 준비 중입니다)"), so they are empty except in tenants loaded with demo data. Drift measurements (PSI, KS) and verdicts are recorded separately and shown on the Drift screen.
Main tasks
When values are empty
| Notice shown | Meaning |
|---|---|
| The metrics source (Prometheus) is disabled, so values are empty (메트릭 소스(Prometheus)가 비활성화되어 값이 비어 있습니다) | The monitoring stack is not connected. Contact the system administrator. |
| No inference requests have reached this tenant yet (이 테넌트로 들어온 추론 요청이 아직 없습니다) | The connection is fine. Values fill in once requests arrive. |
| No inference requests reached the selected model or edge in this period (선택한 모델·엣지로 들어온 추론 요청이 이 기간에 없습니다) | No requests match the filter. Widen the range. |
| No recorded trend (기록된 추세 없음) | Trend metrics have not accumulated yet. |
| No performance data · No distribution data (성능 데이터 없음 · 분포 데이터 없음) | The selected model has no aggregates (see above). |