Outcome#
You can choose durable, working, scratch, and node-local locations, preserve production run identity, and prevent mutable files from changing under queued or running jobs.
Concept#
$SCRATCH is not a backup and $TMPDIR disappears with the job. Slurm captures
the batch script, not every external code, configuration, model, or input file
it references.
Worked Example#
The plan separates Git, approved Euler storage, scratch, and durable results and records enough inputs to rerun the job.
A correct example uses these decisions:
- What is Euler scratch for? Temporary high-throughput files that can be recreated.
- What must a reproducible run identify? Code revision, environment, immutable inputs, parameters, resources, and outputs.
Common Trap#
Keeping the only copy in scratch, processing high-I/O workloads directly on an external NAS mount, or changing inputs in place.
If Blocked#
Do not invent permissions or recursively change a shared tree. Ask the data owner about authoritative storage and use Euler storage for lifecycle and collaboration recovery.
Useful references:
Understand Before Accepting AI Output#
Inspect every source and destination before copying or deleting. An agent must not assume a temporary path is backed up or that a successful transfer is a backup.