Vedrfolnir
Queue predictions and observed error, over time.
Live jobs
Start and finish estimates use currently visible work and capacity.
| Job execution | State / pool | Wait | Expected start | Expected finish | Forecast / unavailable reason |
|---|
Prediction error over time
First observation per job and status · grouped by completion day (UTC).
Negative bias means the forecast was too early. P90 error measures misses; it is not a completion-by guarantee. Job uncertainty is not calibrated.
Error by workload
| Runner pool | Eligible / evaluated | Scored | MAE | P90 error | Bias | Evidence |
|---|
Uses the full recent sample, independent of the chart window. Groups use pool, job type and capacity evidence saved at capture. Load is queued jobs / effective pool slots: idle < 0.5, busy < 2, saturated ≥ 2. Zero capacity is paused; missing evidence is unknown. Inferred slots are lower bounds. Small samples (< 40 scored) are flagged; larger samples remain descriptive, not proof of accuracy.
Recent outcomes
Saved forecasts matched to actual completion. Cancelled or invalid outcomes are excluded from accuracy metrics.
| Job execution | Forecast captured | Observed state | Outcome | Wait error | Finish error |
|---|
Workflow forecasts and accuracy
Workflow targets are separate from individual job targets. P90 bounds may be absent with insufficient calibration.
| Plan | State | Expected finish | P90 finish | Method |
|---|
| Load | Method | Completed plans | MAE | Bias |
|---|