Session

From Training to Production: MLOps for Deep Learning on Databricks

Overview

ExperienceIn Person
TrackArtificial Intelligence & Agents
IndustryEnergy & Utilities, Enterprise Technology, Financial Services
TechnologiesDatabricks Agents
Skill LevelAdvanced
Deep learning breaks most MLOps playbooks. Multi-gigabyte weights, GPU-heavy training, distributed checkpoints, and fine-tuning loops don't fit the lifecycle patterns built for classical ML. Teams end up stitching together fragile pipelines to move models from notebook to production — and paying for it in downtime, drift, and missed retraining cycles.This session shows how to run deep learning MLOps end-to-end on Databricks: distributed training with MLflow tracking, model registry workflows for large models, GPU-aware serving with traffic splitting and autoscaling, and automated monitoring that catches regressions before users do.You'll leave with a reference architecture for deep learning MLOps on Databricks — and the patterns to ship models that stay reliable long after launch.

Session Speakers

Speaker placeholderIMAGE COMING SOON

Michael Shtelma

/Lead Product Specialist - GenAI
Databricks

Puneet Jain

/Lead SSA
Databricks

Full Summary

Operating deep learning models at scale on Databricks

Training is a small part of running deep learning in production. The harder work is building a repeatable loop that prepares governed data, provisions GPUs on demand, runs many trials concurrently, tracks results, evaluates candidates, promotes winners, serves them, and monitors behavior. The session frames that loop on a real object detection task over a tiny dental X-ray dataset, then generalizes the pattern to serious AI and machine learning workloads on the Databricks Platform. Teams looking to understand how these capabilities fit together can explore the Databricks machine learning solutions that underpin this end-to-end approach.

FAQ


Deep learning training on large GPU clusters is expensive, so retraining the same model separately in dev, UAT, and prod is impractical. The Big Book of MLOps permits deploy-model for such cases, and Unity Catalog with MLflow now provides cross-environment lineage, approval gates, and governance to do it safely.

They evaluated DINOv3 and NVIDIA C-RADIO as vision transformer backbones. A pluggable design with a RetinaNet head and feature pyramid lets teams swap in newer backbones without changing the training and deployment pipeline.

AI Runtime provisions GPUs on demand with pre-baked drivers and libraries, so teams pay only for what they use and avoid idle reservations. Multiple trials can run concurrently on separate ephemeral clusters, compressing wall-clock time at comparable per-compute cost.

MLflow logs per-epoch checkpoints, detailed metrics and loss curves, hyperparameters, and GPU utilization. That richer telemetry supports automated or human-guided hypothesis testing and triggers deployment jobs when models are registered in Unity Catalog.

Each registered model runs through an evaluation job that compares performance against the current champion on a held-out dataset. Failing candidates stop there, while winners require an explicit approval before promotion, serving endpoint updates, AI Search embedding refresh, and drift monitoring resets.