Jobs, queues, watches, and dashboard

Cloud state, local submission records, and daemon work are separate views. The job, log, watch, and dashboard resources are Azure ML workspace APIs and therefore manage AML/Sing jobs. Volcano currently supports submission only.

Cloud jobs

aj job list
aj job list -n 100 -s Running -e training
aj job list --type Command --tag team
aj job show <job-or-aj-id>
aj job status <job-or-aj-id> --ws <workspace>
aj job logs <job-or-aj-id> --ws <workspace>
aj job cancel <job-or-aj-id> --ws <workspace>

job list reads the configured workspace unless --ws NAME is available and specified. Job type and tag filters are sent to Azure; status and experiment filters are applied by the daemon while fetching.

Short eight-character aj IDs are resolved through the local journal. A full Azure job name also works. Pass --ws when the template or compute resolves to a workspace other than the active project workspace.

aj job logs checks status, downloads the selected aggregate output and error text, and reports when a queued or provisioning job has no logs yet. The SDK and dashboard additionally support file lists and byte-range reads.

The CLI cancels but does not expose permanent deletion. In aj dash, d deletes a terminal job after confirmation; active jobs must be canceled first. The SDK exposes d.job.delete() for AML/Sing jobs.

Local records

aj list
aj list -n 50
aj list -t <template>

.azure_jobs/record.jsonl is append-only and newest-first when read. It records the built request, result status, backend name, portal URL, timestamp, and failure note. It is local history, not the authoritative cloud status.

Statistics and experiments

aj job stats
aj job stats --days 30
aj job stats --all

aj exp list
aj exp list --days 30 --all
aj exp show <experiment>

Statistics summarize duration, GPU-hours, success rate, experiment, compute, workspace, and user where data is available. --all fans out across visible workspaces and reports partial failures rather than hiding successful results.

Set the default experiment with:

aj config experiment training

Submission queue

aj run --queue -t <template> train.py
aj queue list
aj queue show <ticket>
aj queue wait <ticket> --timeout 3600
aj queue cancel <ticket>

Queue entries are daemon-owned and persisted per project/workspace context. Only pending entries can be canceled through the queue.

Watches

aj watch add <job>
aj watch list
aj watch listen
aj watch listen --no-desktop --timeout 600
aj watch remove <job>

The daemon keeps polling after the invoking shell exits. Watch journals survive daemon restart and are restored lazily when a later request reopens that project/workspace context. Watches are removed automatically after a terminal transition. watch listen receives server-sent events; desktop notifications use the platform notifier when available.

Dashboard

aj dash
aj dash -n 500 --page-size 40
aj dash --mouse

--last controls the initial fetch, not a permanent pagination cap. The TUI can request later Azure pages.

Important keys:

Keys	Action
up, down	select job
left, right	previous/next page
i, l	info/log view
r	refresh
c, d	cancel/delete
f, e, /, F	status, experiment, search, clear filters
w	switch workspace
o	select log file
L, s, Ctrl-S	live tail, auto-scroll, save buffer
Esc, q	manual/quit

The dashboard uses the same SDK and daemon as the CLI. Controllers perform I/O; stores own state; HTTP Range reads keep log tailing bounded.

See TUI dashboard design for state, concurrency, and destructive-action contracts.