Inference System Performance on Live Inference Service Kubernetes production environment
3Latency P50 (ms)LSTM
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| LSTMLagged steps=30, Forecast steps=15, Batch size=256, Interval=1-minute2026.05 | 3 | 4.8 | 726 | 0.1 | |
| TFTLagged steps=30, Forecast steps=15, Batch size=256, Interval=1-minute2026.05 | 3 | 4.8 | 779 | 0.2 | |
| STARIXNetLagged steps=30, Forecast steps=15, Batch size=256, Interval=1-minute2026.05 | 3 | 4.8 | 346 | 0.1 | |
| ARIMAAutoregressive lag=30, Interval=1-minute2026.05 | 3.2 | 232 | 17,406 | 3.1 | |
| DeepARLagged steps=30, Forecast steps=15, Batch size=256, Interval=1-minute2026.05 | 3.2 | 4.8 | 584 | 0.2 |