Understanding run output

Every run writes everything it produces under one folder:

<output_dir>/<run_id>/

sflow run prints that path when the run ends — on success and on failure, as soon as the folder exists — together with the summary path and any command-log paths, so you can jump straight to diagnostics. (A --dry-run creates nothing; it only prints the paths it would use.)

This page is a map of what lives in there and which one answers your question. Each row links to the page with the full detail.

Start with the summary

sflow_summary.log is the single best entry point. It is written live during the run and finalized when the workflow exits, and it collects in one place:

• workflow status, start/end time, duration, output directory and task counts
• runtime details — package version, binary and Python paths, install mode, repo path, and git branch/commit when available
• a task duration timeline and a task event timeline
• probe traces (the last attempt of every readiness/failure probe) when any task defines probes
• GPU and node usage charts when resource placement data exists
• failure hints — task name, attempts, reason, and the path to that task's log

See Outputs & logs → Execution summary for an annotated example.

What do you want to know?

Question	Look at	Details
Did the run succeed, and how long did each task take?	sflow_summary.log	Outputs & logs
What did my script actually print?	<task>/<task>.log	Outputs & logs
What did sflow itself do — scheduling, status transitions?	sflow.log	Outputs & logs
What exact command was launched (srun / docker / kubectl)?	*_cmds.log	Outputs & logs
What metrics did my benchmark produce?	<task>/result.json, results.json	Results
Why did readiness never fire?	probe traces in sflow_summary.log	Probes
How busy were the GPUs / CPU / network?	sflow_monitor.log, sflow_monitor/	Monitor
Which GPUs and nodes did each task get?	usage charts in sflow_summary.log	Resources
How do I get all this off the cluster?	storage targets + uploads:	Uploads

The output tree at a glance

<output_dir>/<run_id>/
├── sflow_summary.log      # start here: status, timings, probe traces, failure hints
├── sflow.log              # orchestration + command/status lines (no task stdout)
├── *_cmds.log             # launch commands, grouped by family (bash/slurm/docker/ssh/python)
├── results.json           # workflow-level metric index      (only with `result:`)
├── sflow_monitor.log      # resource overview                (only with `monitor:`)
├── sflow_monitor/         # raw samples + per-task reports   (only with `monitor:`)
└── <task>/
    ├── <task>.log         # full per-task stdout/stderr
    ├── result.json        # canonical per-task metrics       (only with `result:`)
    └── ...                # anything your scripts write

A task's stdout/stderr always goes to its own <task>/<task>.log and is deliberately kept out of sflow.log, so the orchestration log stays readable. See Outputs & logs for the full contract, including how Kubernetes log streaming is reconciled before anything parses it.

Triaging a failure

1. sflow_summary.log — the failure hint names the task, the attempt count, the reason, and the path to that task's log.
2. <task>/<task>.log — the task's own stdout/stderr, the actual error.
3. *_cmds.log — confirm the launch command and its flags were what you expected (wrong mounts, wrong --gpus, missing env).
4. probe traces in the summary — if the task hung rather than crashed, these show the last readiness/failure probe attempt. See Probes.
5. sflow_monitor/ — if it was slow rather than broken, check whether the GPUs were actually busy. See Monitor.

Re-running with --dry-run is often the fastest way to confirm a planning problem (node/GPU placement, resolved variables, mounts) without consuming cluster resources.

Machine-readable output

Two files are meant to be consumed by other tools rather than read by a human:

• <task>/result.json — the canonical per-task result, written after the task succeeds.
• results.json — the workflow-level index of every task's results.

Both appear only when a task declares result:. They are the stable contract for downstream tasks and external tooling — see Results.

Beyond the local folder

• Monitor — hardware sampling and reports. Raw CSV samples are always written when monitor: is set; the per-task charts and summaries need report: { enabled: true }.
• Uploads — declare storage targets and per-task uploads: to ship files to S3 as each task completes, so partial results survive a cancelled run.