Lightning-AI/pytorch-lightning

MLFlow Logger Makes a New Run When Resuming from hpc Checkpoint

Open

#6,205 opened on Feb 25, 2021

View on GitHub
 (8 comments) (0 reactions) (0 assignees)Python (3,233 forks)batch import
bugcheckpointingenvironment: slurmhelp wantedlogger: mlflowpriority: 2won't fix

Repository metrics

Stars
 (26,687 stars)
PR merge metrics
 (Avg merge 9d 15h) (3 merged PRs in 30d)

Description

🐛 Bug

Currently the MLFlowLogger creates a new run when resuming from an hpc checkpoint, e.g., after preemption by slurm and requeuing. Runs are an MLFlow concept that groups things in their UI, so when resuming after requeue, it should really be reusing the run ID. I think this can be patched into the hpc checkpoint using the logger which I believe exposes the run ID. This can also be seen on the v_num on the progress bar which changes after preemption (in general that v_num probably shouldnt be changing in this case). I'm happy to attempt to PR this if the owners agree that it's a bug.

To Reproduce

Use MLFlowLogger on a slurm cluster and watch the mlflow UI when preemption happens, there will be a new run created.

Expected behavior

Runs are grouped neatly on the MLFlow UI

Environment

  • CUDA: - GPU: - available: False - version: 10.2
  • Packages: - numpy: 1.20.1 - pyTorch_debug: False - pyTorch_version: 1.7.1 - pytorch-lightning: 1.2.0 - tqdm: 4.57.0
  • System: - OS: Linux - architecture: - 64bit - ELF - processor: x86_64 - python: 3.8.1 - version: #1 SMP Thu Jan 21 16:15:07 EST 2021

cc @awaelchli @ananthsub @ninginthecloud @rohitgr7 @tchaton @akihironitta

Contributor guide

Research direction
Investigate the MLFlowLogger class to understand how it handles run IDs during checkpoint resumption. Look for ways to reuse the existing run ID when resuming from an HPC checkpoint.
Tech stack
pythonpytorch
Domain
backend
Issue type
Bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
2
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-3 hours
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PythonPyTorch
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
60