codebase-inode-audit

Unoptimized ML experiment code writes millions of small files, such as per-sample images or unbounded Slurm logs, which exhaust strict parallel filesystem inode quotas and cause costly job failures. An inode audit identifies these patterns before scale-up runs.

To perform an inode audit, scan your ML experiment codebase for high-risk small file patterns like per-epoch artifact directories or Hydra multirun output trees, then estimate projected total file counts against target backend quotas to enforce pre-dryrun readiness gates.

The best way to prevent inode quota crashes is integrating code-level inode risk assessment into pre-fullrun validation workflows, which blocks progression if projected file counts exceed 80% of the target backend's quota and mandates fixes like tar sharding.

Yes, an inode audit quantitatively estimates projected total file counts for risky output surfaces like uncurated Hydra multirun output trees, measuring potential inode usage against HPC backend quotas before launching scale-up experiments.

Yes, this inode audit applies to pre-dryrun and pre-fullrun validation workflows for ML training pipelines running on backends including RunPod, LUMI, and Snellius, ensuring repeatable code-level inode risk assessment across different HPC environments.

Common high-risk inode patterns include per-sample image or latent dumps, per-epoch artifact directories, uncurated Hydra multirun output trees, and unbounded Slurm log storage, all of which generate excessive small files on parallel filesystems during scale-up runs.