ExaMLOps compared with other MLOps options for HPC¶
Short answer: many MLOps platforms assume Kubernetes or a cloud. ExaMLOps is built for the case where models train on a batch-scheduled HPC cluster (Slurm or Flux), and it combines tools you may already know — MLflow as the registry, Ray Serve for serving, Prefect for pipelines — with the parts that are usually missing around them: job submission to the scheduler, gated promotion, drift-triggered retraining, cost and carbon accounting, and an audit trail. If your organisation is Kubernetes-native, or you only need experiment tracking, one of the alternatives below is likely the better fit.
Claims about other projects are taken from their own documentation, checked on 2026-09-10 and linked in Sources. Projects change; if something here is out of date, please use the edit button on this page.
At a glance¶
| Runs training on Slurm / Flux | Needs Kubernetes | Model registry | Gated promotion + serving | Notes | |
|---|---|---|---|---|---|
| ExaMLOps | Yes — Slurm and Flux, local or over SSH, plus a mock mode |
No (Docker Compose or Helm for the services) | MLflow | Metric and evaluation gates, approval queue for model changes, Ray Serve with traffic splits | Open source (Apache-2.0); young project, one maintainer |
| Kubeflow | Not its model — described as "the Kubernetes-native stack for data & AI workloads" | Yes | Kubeflow Hub (formerly Model Registry) | Pipelines, Trainer and other components on Kubernetes | Large community; the natural choice if you already run Kubernetes |
| MLflow on its own | Not one of the components it describes | No | Yes — "centralized model versioning, stage management, and model lineage tracking" | Deployment to REST APIs, cloud platforms and edge devices; you build the gates | ExaMLOps uses MLflow as its registry |
| Ray on Slurm (do it yourself) | Yes, with care — Ray's own docs call Slurm usage "a little bit unintuitive" | No | No (bring your own) | Ray Serve for serving; everything else is yours to build | ExaMLOps uses Ray Serve for serving |
| ClearML | Yes — the Slurm Glue maps ClearML queues to Slurm jobs | Not for the Slurm path | Not checked here — see ClearML's docs | clearml-serving for "model deployment and orchestration" |
The Slurm Glue "is available under the ClearML Enterprise plan" |
| Metaflow | Via the metaflow-slurm extension (by Outerbounds; 0.0.4, Dec 2024) |
Not for the Slurm path | Runs and results are tracked through its metadata service | Not checked here | The extension runs individual steps on Slurm over SSH |
| Valohai | Yes — "now supports Slurm" (announced April 2024) | Not checked here | Not checked here | Not checked here | Commercial platform (pricing page, free trial) |
| DKube | Yes — a "hub and spoke model to integrate HPC/Slurm" | Yes, for the hub ("the associated Kubernetes workloads") | Built "on top of" Kubeflow and MLflow | Not checked here | "A commercial MLOps offering" |
"Not its model" and "not one of the components it describes" mean the project's own documentation does not present that capability — not that it is impossible to build. "Not checked here" means we did not verify it; the project may well offer it.
When ExaMLOps is a good fit¶
- Your models train on an HPC cluster run by Slurm or Flux, and you want training runs, versions and serving to be tracked and governed rather than held together by batch scripts.
- You need promotion to be gated and recorded — metric conditions, evaluation gates, an approval queue for model changes, and a hash-chained audit trail — see Who decides.
- A cluster must be admitted by an operator before any job can land on it (HPC fleet).
- You want cost and carbon accounting per model and project, with formulas you can swap (FinOps providers).
- You want AI agents to operate the platform through MCP, read-only unless you allow writes (All interfaces).
When to choose something else¶
- You already run Kubernetes for ML. Kubeflow and KServe are built for that environment and have much larger communities. ExaMLOps can generate Kubernetes serving manifests, but its centre of gravity is the HPC scheduler.
- You only need experiment tracking and a registry. A plain MLflow server is simpler.
- You want a packaged commercial product with Slurm integration. ClearML's Enterprise plan includes a Slurm Glue, Valohai supports Slurm, and DKube integrates Slurm from a Kubernetes-based hub. ExaMLOps is community software; for hands-on help with a deployment, the maintainer takes engagements.
- You need a mature project with many maintainers. ExaMLOps is young and has one
maintainer. Its reference deployment schedules on Flux; the Slurm path is covered by tests and
mockmode and would benefit from validation at more Slurm sites — reports are very welcome.
How ExaMLOps relates to the tools it uses¶
ExaMLOps does not replace MLflow, Ray Serve or Prefect — it runs them and adds the platform around them:
| Tool | Role in ExaMLOps |
|---|---|
| MLflow | Model registry and experiment tracking; the Staging / Canary / Production aliases |
| Ray Serve | Multi-model serving with traffic splits between versions |
| Prefect | Auto-discovery training pipelines and deployments |
| Slurm / Flux | Where training jobs run, through one scheduler abstraction |
| Prometheus, Grafana, Loki, Tempo | Metrics, dashboards, logs and traces |
The system map shows how these pieces connect.
Sources¶
- Kubeflow — Introduction: "Kubeflow is composed of modular, open source projects that form the Kubernetes-native stack for data & AI workloads." · Components: "Kubeflow Hub (formerly Model Registry)".
- MLflow — Documentation: tracking, model registry ("centralized model versioning, stage management, and model lineage tracking"), deployment and evaluation.
- Ray — Deploying on Slurm: "Slurm usage with Ray can be a little bit unintuitive."
- ClearML — Slurm (Native): "ClearML Agent can run tasks on Linux clusters managed by Slurm." · "Slurm Glue is available under the ClearML Enterprise plan." · ClearML Serving: "
clearml-servingis a command line utility for model deployment and orchestration." - Valohai — Introducing Slurm Support (2024-04-02): "Valohai, the leading MLOps platform, now supports Slurm".
- DKube — MLOps on HPC/Slurm with Kubeflow: "DKube implements a hub and spoke model to integrate HPC/Slurm. The hub runs the MLOps management and control plane and the associated Kubernetes workloads." · "DKube is a commercial MLOps offering that is built on top of best-of-breed open-source AI/ML platforms such as Kubeflow & MLflow."
- Metaflow —
metaflow-slurmon PyPI: "Slurm extension for Metaflow", by Outerbounds, version 0.0.4 (2024-12-10) · Client API: "The Client API consults the metadata service to gather results".