MLRun

Serverless Functions and Jobs

Wrap Python or containerized code as MLRun functions that scale on Kubernetes. Schedule batch jobs, launch ad-hoc experiments, and reuse steps as building blocks across projects. Dependency specs and images are versioned so runs remain reproducible, and logs, metrics, and artifacts store with each execution for audit. Reusable blueprints standardize logging and resource limits, and labeled runs simplify audits. Promote the same function from dev to prod without rewrites.

Pipelines and Automation

Compose multi-step workflows for data prep, training, evaluation, and deployment, with parameters flowing between steps and conditional logic handling retraining or rollback. Git integration pins code at commit, and approvals protect promotions so governance is met before models meet traffic. Schedules trigger retraining on drift or data arrival, and human-in-the-loop gates capture sign-offs in the pipeline view. Chat notifications summarize outcomes and link metrics and artifacts for quick checks.

Model Serving and Real-Time Features

Serve models with auto-scaling endpoints; add feature transforms; and route canary, shadow, or A/B traffic to de-risk changes. Latency, errors, and drift are monitored continuously so regressions are caught early. Transforms run near the model so payloads stay lean, while schema validation rejects malformed requests before they cause errors. Blue/green patterns keep capacity ready during upgrades, and health probes confirm readiness before shifting real user traffic safely.

Experiment Tracking and Artifacts

Record parameters, metrics, plots, and data versions for each run, then compare candidates without manual spreadsheets. Artifacts capture datasets, models, and reports so results can be audited and reproduced later. Rich diffs compare metrics and confusion matrices, and shared views keep evaluation criteria consistent. Dataset and model fingerprints prevent accidental reuse of stale assets, while promotions require checks to pass, preserving confidence when moving to production.

Security, Roles, and Integrations

Integrate with SSO, secrets, and registries; control access by project or function; and connect to data lakes, queues, and CI systems through APIs and SDKs. Quotas and policies prevent noisy neighbors, and templates standardize GPU and data access. Secret mounts keep credentials out of logs, and policy scaffolds define which registries or buckets are allowed. Resource quotas curb overuse, and cost labels attribute compute across teams to improve accountability and budgeting.