Timeline
Overview
The Timeline chart draws one horizontal bar per trial spanning its start and completion time, similar to a Gantt chart. It reveals how trials were scheduled over wall-clock time — how many ran in parallel, whether workers were idle, and where slow trials created bottlenecks.
Controls
This chart has no additional controls; the X-axis unit (seconds, minutes, or hours) is chosen automatically based on the total time span of the study.
Operations
Pan, zoom, and reset follow the standard 2D chart controls.
- Hover: Hover over a bar to highlight it and see a tooltip with the trial number, state, start, end, and duration. Other bars are dimmed while one is highlighted.
- CSV export: Exports the same intervals as the chart, in elapsed seconds, including provisional RUNNING intervals. Trials omitted from the chart are omitted from this CSV too.
How to Read
- X axis: elapsed time since the study's earliest recorded trial start (re-based to zero). The unit switches automatically — seconds for short studies, minutes beyond 10 minutes, hours beyond 2 hours.
- Y axis: trial number.
- Bar length: the trial's recorded execution duration. RUNNING trials with no completion timestamp extend to the latest known start or completion in the study, not the current clock time. Their end and duration (including in CSV) are provisional, not recorded completion times or final execution durations.
- Missing timestamps: trials without a start are omitted. COMPLETE, PRUNED, and FAIL trials without a completion timestamp are also omitted; their end times are not inferred from other trials.
- Color by trial state: complete, pruned, running, and failed trials are colored differently (see the legend).
- Overlapping bars: trials ran in parallel. The number of overlapping bars at any vertical slice approximates the number of concurrent workers.
- Gaps between bars on the time axis: workers may have been idle, e.g. waiting for a new trial or a shared resource.
- A few very long bars: some trials took much longer than others — worth checking for expensive hyperparameters or a stuck trial.
- Many short trials: individual bars may be smaller than a pixel in the overview. Zoom into a smaller time and trial range to distinguish intervals; the overview does not imply that trials all started at zero.