Staff analytics

Credential

A dashboard login is used automatically. Otherwise paste an access token; it is held in this tab only. An API key cannot reach this page, whoever issued it.

No credential set.

Range

Days are Pacific calendar days (America/Los_Angeles, PST or PDT by date — not a fixed offset). "Past day" is today so far; "past week" is the last seven days including today.

Daily trends

One bar per Pacific day. A job counts on the day it was dispatched (or submitted, if it never was), the same rule as the volume panel. Hover a bar for the per-pipeline breakdown; click a bar, or one colour within it, to list those jobs below. Any non-zero slice is drawn at least a few pixels tall so a single failure stays visible and clickable, so heights are exact only above that floor.

Red (failed) sits on top of green (succeeded). Canceled and still-running jobs are in neither colour; the tooltip counts them. Percent is failed or succeeded over the two together.

Time processed is the segment length of every job run that day, whatever its outcome (assets.duration_seconds, the unit quota is charged in). A LeRobot export counts the length of the segment it exported. The tooltip also gives the succeeded part.

Volume per service

Input seconds is the unit: the length of the recordings submitted (assets.duration_seconds), which is what quota is charged in. Node seconds is a different quantity — compute consumed (actual_node_seconds), which includes instance startup — and is listed separately, never added to input.

Job timeline

One bar per job. Length is the task's life: processing_tasks.started_at → completed_at, from start to completion. Start is when the worker claimed the job on its node; for a release without the worker start marker it is dispatch, which comes before the node exists and so includes EC2 launch and boot. Either way it includes input download and wrap-up (output upload, validation, publish) around the algorithm itself; the job list below splits the two. It still ends at completion, before the node shuts down, so it is shorter than the node's billable time; the two are never mixed here. Dashed, hatched bars are time spent waiting — refused a capacity slot, or queued while the node launches and boots — and are never drawn as running. Overlapping jobs are placed in separate lanes so each stays hoverable.

Uploads

One row per upload batch, newest first; click a row for its files. Upload time is the batch's own created_at → completed_at (last update while in flight; stalled after 30 minutes without one), and throughput is total size over that time, so it includes any pause on the client. Batches do not record who uploaded: Account shows the account's members. SLAM / Hand / Depth use the same per-segment check the dashboard does.