Azure ML to Databricks Migration: A Practical Guide

 & LatentView Analytics

LatentView Analytics has been helping enterprises make data-driven decisions for nearly 20 years. The company brings deep expertise in data engineering, business analytics, GenAI, and predictive modeling to 50+ Fortune 500 clients across tech, retail, financial services, and CPG. A publicly traded company serving the US, India, Canada, Europe, and Singapore, LatentView is recognized in Forrester's Customer Analytics Service Providers Landscape.

SHARE

Table of Contents

This guide helps data and AI leaders move from Azure Machine Learning to Databricks, from choosing a strategy to shifting data, models and endpoints without disrupting live predictions.

Azure ML to Databricks migration moves data assets, training code, experiments, models and endpoints from Azure Machine Learning onto Databricks, governed by Unity Catalog and tracked with MLflow.

Key Takeaways

  • Azure ML to Databricks migration shifts feature engineering, training, model registry and serving onto Databricks, with Unity Catalog governing data and models together
  • Azure ML retirements, including SDK v1 support and prompt flow, mean many teams are rewriting ML code anyway
  • Both platforms speak MLflow, so experiments and MLflow-format models carry over with modest effort
  • Six phases, repeated in waves of models, move data, training, serving and GenAI workloads in a controlled order
  • Parity checks and parallel endpoints decide when each model can safely switch over

Azure ML Changes That Shape the 2026 Migration Window

Recent Azure ML retirements force many teams to rewrite SDK code, prompt flows and labeling workflows, which makes this a natural time to decide where ML should run long term.

Three changes matter most for planning:

ChangeDateImpact
SDK v1 end of supportJune 30, 2026Code built on azureml-core needs a rewrite
Data labeling retirementSeptember 30, 2026Labeling workloads are deleted
Prompt flow retirementApril 20, 2027GenAI flows in Azure ML and Microsoft Foundry need a new home

A forced rewrite raises an obvious question: rewrite for SDK v2 and stay, or rewrite once where the data already lives? When features are engineered in Azure Databricks but models train in Azure ML, the same data is copied into Azure ML datastores for training and predictions are copied back for dashboards. Every copy adds storage, delay and another permission model, and lineage breaks at each handoff.

The three dates also set the order of work. Microsoft ended support for CLI v1 on September 30, 2025 and for SDK v1 on June 30, 2026, and its upgrade guide warns that existing v1 workflows keep running but could be exposed to security risks or breaking changes. For data labeling, Microsoft’s retirement notice says workloads running labeling are deleted on September 30, 2026 and associated application data is lost, and it directs teams to third-party labeling providers, with labeled data copied into the Azure ML storage account. Prompt flow, retiring April 20, 2027, has the longest runway, which is why GenAI flows sit on their own track later in this guide.

Our Databricks vs Azure Machine Learning comparison covers where each platform fits.

Azure Machine Learning vs Databricks: Mapping Every MLOps Component

Almost every Azure ML asset has a Databricks counterpart, so much of the migration is a mapping exercise, with pipelines, environments and endpoints needing the most rework.

Azure ML AssetDatabricks Counterpart
Datastores and data assetsExternal locations, Delta tables and volumes in Unity Catalog
Compute instances and clustersServerless or classic compute, with serverless GPU through AI Runtime
EnvironmentsDatabricks Runtime for Machine Learning plus libraries
ExperimentsMLflow experiments on Databricks
Model registryModels in Unity Catalog with aliases
Feature storeFeature tables in Unity Catalog with Databricks Feature Engineering
Pipelines and componentsLakeflow Jobs with tasks and task values
Online and batch endpointsModel Serving and batch inference in Lakeflow Jobs
Prompt flowAgents with MLflow tracing and evaluation

Some rows are close to a configuration change. SDK v2 already logs through MLflow, so experiment tracking needs only a new destination, and data in ADLS Gen2 stays where it is.

Others decide the timeline. Pipelines carry Azure ML-specific wiring, and endpoints have applications depending on their URLs and credentials.

Azure ML to Databricks Migration Strategy

The right strategy depends on endpoint dependencies, SDK version, data location and model risk, which together decide the operating model and the technical path for each model.

Full consolidation moves data, training, registry, serving and monitoring in one planned sequence, with the Azure ML workspace closed at the end. It suits teams whose features already live in Databricks and whose models serve internal applications.

A hybrid period moves feature engineering and training first, while selected models keep serving from Azure ML. Because both platforms use MLflow model formats, a model trained on Databricks can still be deployed through an Azure ML endpoint while external applications are repointed. Agree an exit date up front, since two serving platforms double monitoring and on-call effort.

Model-by-model waves suit large estates. Each wave groups a few related models that share data, features or consumers, so feature tables and pipelines are built once for the group.

Each model then gets a technical path: lift and shift for stable models, refactoring into feature tables and Lakeflow Jobs for models retrained often, and rebuilds for Designer pipelines and prompt flows.

Azure ML’s own upgrade guidance points the same way. For pipelines, Microsoft says what runs inside a child job does not need to change, but code specific to Azure Machine Learning should be removed from training scripts and model logging and tracking replaced with MLflow. Scripts that already log through MLflow therefore need the least change on Databricks, while Designer pipelines and prompt flows are platform-specific assets that are rebuilt, not moved.

How to Prioritize Azure ML Models for Migration

Grouping models by how they are trained and served gives a clear migration order, starting with batch models that are easy to validate and finishing with GenAI flows and GPU-heavy workloads.

  1. Batch tabular models: Churn, demand or risk scoring, where predictions land in tables and parity is simple to prove
  2. Real-time models: Fraud, pricing or recommendation endpoints, once a Model Serving pattern and switchover plan exist
  3. Deep learning models: Vision, text or forecasting work that needs GPU sizing and library checks
  4. GenAI applications: Prompt flows and agents, run as a separate track timed against the April 2027 retirement

Ordering within each group comes down to business value, the number of applications calling the model and the amount of SDK v1 code it carries.

Deployment type is a second signal. Microsoft states that deploying models as ACI or AKS web services with SDK or CLI v1 is deprecated, and that managed online endpoints and Kubernetes endpoints are the v2 options, so models still served through v1 web services are strong candidates for an early wave.

Check endpoint traffic before scheduling anything, since models with no requests can simply be retired.

The Six Phases of an Azure ML to Databricks Migration

Azure ML migrations run through six phases, from discovery and architecture to data, training, serving and cutover, with phases three to six repeating for every wave of models.

PhaseMain Output
1. Discovery and assessmentModel inventory, dependencies and migration waves
2. Architecture and landing zoneCatalog, network, identity and compute design
3. Data and feature migrationDelta tables, volumes and feature tables in Unity Catalog
4. Training and registry migrationTraining workflows and models registered in Unity Catalog
5. Endpoint and GenAI migrationModel Serving endpoints, batch inference and agents
6. Validation and cutoverLive models on Databricks and retired Azure ML assets

Phases 1 and 2 happen once. The first wave through phases 3 to 6 is the slowest, because it produces the templates later waves reuse.

Phase 1: Azure ML Migration Discovery and Assessment

Discovery traces every production model back through its endpoint, pipeline, data and environment, and records which applications depend on it before any migration wave is planned.

Azure ML workspaces collect experiments and registered models that never reached production. Start from live endpoints and scheduled pipelines instead, tracing each model backward to its code, data and environment.

For every production model, capture:

  • SDK version and any Azure ML-specific calls such as Run.get_context() or MLClient
  • Data asset types, datastores and where features are engineered
  • Curated or custom images, conda files and private package feeds
  • Compute setup, endpoint traffic targets and calling applications
  • Responsible AI dashboards, approval steps and audit requirements

Searching every repository for azureml and azure.ai.ml imports catches the dependencies reviews miss, usually in helper modules and custom scoring scripts. The output is a wave plan, with agreed parity criteria for each model before any work begins.

Microsoft’s upgrade guide flags two items worth recording during discovery. Database datastores are not supported in v2, and export to object storage, usually Azure Blob, is the recommended path. Workspaces that use a private endpoint have the v1_legacy_mode flag enabled, which blocks use of v2 APIs until network isolation is configured for v2.

Phase 2: Azure ML to Databricks Architecture and Landing Zone

The architecture phase turns the component mapping into a working design, settling catalogs, network, identity, secrets and compute standards that every migration wave will reuse.

Catalog design comes first. Separate catalogs such as ml_dev, ml_staging and ml_prod, with schemas per business domain, let models, feature tables and training data share one permission model. Data scientists write in development, CI/CD tests in staging, and only service principals write to production.

Network and identity design builds on Azure patterns. Azure Databricks supports VNet injection and Azure Private Link, storage access runs through the Access Connector for Azure Databricks, and secret scopes backed by Azure Key Vault read the same secrets Azure ML used. This is also the time to retire account keys and SAS tokens on older datastores.

Compute policies close the phase, covering approved runtimes, instance types, autoscaling limits and cost tags.

Identity is documented on both sides. Azure ML workspaces use a system-assigned or user-assigned managed identity to reach other services, and managed online endpoints can use a managed identity to access Azure resources during inference. Databricks service principals give automated tools and CI/CD API-only access without individual user credentials, and users can be granted permission to run jobs as a service principal, which prevents jobs from failing when a user leaves the organization.

Phase 3: Azure ML Data and Feature Migration to Unity Catalog

Data in Azure storage usually stays in place, while Azure ML datastores, data assets and features are re-registered in Unity Catalog as external locations, Delta tables, volumes and feature tables.

Storage usually stays where it is. Unity Catalog governs the same ADLS Gen2 containers through external locations (the storage account needs a hierarchical namespace), and each data asset type has a natural target:

Azure ML Data AssetDatabricks Target
mltableDelta tables, adding versioning, time travel and lineage
uri_folderUnity Catalog volumes for images, audio and documents
uri_fileVolume files, or Delta tables when tabular

Features deserve far more attention than files. When the same customer attributes are calculated in several scripts, each version can differ slightly. Consolidating them as feature tables in Unity Catalog, which are Delta tables with primary keys, gives every model one definition. Time-series feature tables add point-in-time lookups, joining each label to feature values as they were at that moment, which stops future data from leaking into training.

Existing v1 assets carry over in part. Microsoft states that a v2 project can reuse existing v1 resources such as workspaces and compute and existing assets such as models and environments, and that object storage datastores created in v1 remain usable in v2.

Dataset types map automatically when a v1 dataset is consumed in a v2 job: a FileDataset becomes a folder (uri_folder) and a TabularDataset becomes a table (mltable). Forward compatibility is not provided, so v2 data assets cannot be used in v1 code.

Phase 4: Azure ML Training Code, Experiment and Model Registry Migration

Training migration rewrites Azure ML SDK calls into MLflow and Spark, rebuilds environments and pipelines, carries over the experiment history that matters and registers models in Unity Catalog.

Model logic rarely changes. What changes is how code logs, reads data and registers results:

# Azure ML SDK v1
from azureml.core import Run
Run.get_context().log(“rmse”, rmse)

# Databricks
import mlflow
with mlflow.start_run():
    mlflow.log_metric(“rmse”, rmse)

Azure ML Input references become spark.table() reads or volume paths, and custom score.py logic moves inside an mlflow.pyfunc model. Environments shrink considerably, since Databricks Runtime for Machine Learning already includes PyTorch, TensorFlow and XGBoost, although the DBR 18.0 ML release notes say TensorFlow is no longer top-tier support. Images with system dependencies run through Databricks Container Services.

Pipelines are where most effort goes. Each step becomes a Lakeflow Jobs task, outputs pass through Unity Catalog tables or task values, and compute is chosen per task, so GPU training and CPU scoring sit in one workflow. Other assets need separate handling:

  • Parallel jobs that train one model per store become applyInPandas over grouped data.
  • Designer drag-and-drop pipelines are rebuilt as notebooks and job tasks.
  • Databricks AutoML is no longer built into Databricks Runtime 18.0 ML and above, so check any AutoML dependency before moving it.

Registered models move to three-level Unity Catalog names, with aliases such as @champion replacing stages. Only the owner can reassign an alias, so ownership should sit with a small release group. Often it is enough to move only the experiment runs that produced production models, since Azure ML supports just part of the MLflow API.

Registration itself is a small code change. The sequence is short:

  1. Point the MLflow registry at Unity Catalog.
  2. Log the model with a three-level name, which registers it in one step.
  3. Grant EXECUTE on the registered model plus USE CATALOG and USE SCHEMA on its catalog and schema, which loading or deploying requires.
  4. Load a version by alias in notebooks and scheduled jobs, so promoting a new version needs no code change in the workloads that consume it.

Phase 5: Azure ML Endpoint Migration to Databricks Model Serving

Endpoint migration moves Azure ML online endpoints to Model Serving, turns batch endpoints into scheduled batch inference, and rebuilds prompt flows as governed agents before any client switches.

Azure ML Online EndpointDatabricks Model Serving
Several deployments with traffic splitSeveral served models with traffic split
Key or Entra ID authenticationDatabricks OAuth tokens or service principals
Application Insights loggingInference tables in Unity Catalog

Deploying the endpoint is the quick part. The slow part is the applications: each needs a new URL, new credentials and sometimes a different payload, and each belongs to a team with its own release calendar, so agree switch dates with every owner early.

Databricks documents the main serving settings. Compute can be CPU or GPU, and its production guide advises starting on CPU and moving to GPU only when latency is high or the model needs specialized compute such as deep learning. Provisioned concurrency sets how many simultaneous requests an endpoint can process, sized as target queries per second multiplied by average latency in seconds. Scale to zero takes effect after 30 minutes of inactivity and adds cold-start latency to the next request, so it is not recommended for production. Inference tables log each request and response from the endpoint to a Unity Catalog Delta table.

Batch endpoints get simpler. A scheduled Lakeflow Jobs workflow loads the @champion model, scores tables with Spark and writes predictions straight to the Delta tables that reports already read, removing the export step.

Prompt flows follow their own route:

  • Finished agents deploy through Model Serving or Databricks Apps
  • MLflow Tracing captures each step of an agent run
  • Retrieval through Vector Search, with Delta Sync indexes kept current from Delta tables
  • MLflow evaluation scorers that rate every prompt or model change against an evaluation dataset

One shared evaluation set compares quality, latency and cost before users switch. Microsoft’s own migration guidance points prompt flow users to Microsoft Agent Framework, which is the alternative for teams that stay on Azure ML and Foundry.

Phase 6: Azure ML Model Validation, Cutover and Decommissioning

The final phase proves each migrated model matches Azure ML on data, predictions, quality and latency, switches clients in a controlled way, and retires Azure ML assets once results hold.

Each model follows a short runbook:

  1. Pin library versions, since a new scikit-learn or XGBoost release can shift predictions
  2. Compare feature values row by row on a sample
  3. Score a fixed holdout set on both platforms against an agreed tolerance
  4. Replay recorded production requests and compare latency and error rates
  5. Move applications one at a time, least critical first
  6. Keep the Azure ML endpoint as a fallback until the observation period ends
  7. Delete Azure ML endpoints and compute, and file a parity report with owner sign-off

Tolerances depend on the model. Simple linear models should match closely, while retrained tree ensembles and neural networks need agreed limits on metric drift. After cutover, data profiling in Unity Catalog tracks drift, and the parity report becomes the evidence risk teams ask for first.

Databricks recommends running old and new setups in parallel and moving traffic across gradually when workloads shift to Unity Catalog models, and the same approach suits cutover. A single Model Serving endpoint can serve multiple models or model versions with a controlled traffic split between them, which gives the same gradual shift between versions after cutover.

Retirement does not switch Azure ML workloads off immediately. Microsoft states that existing SDK v1 and CLI v1 workflows continue to operate after their end-of-support dates but could be exposed to security risks or breaking changes, so the observation period should not assume v1 code will stay unchanged.

Azure ML to Databricks Migration Best Practices

Strong migrations govern models like data, standardize promotion, keep compute costs visible and help data scientists become productive on Databricks from the first wave.

Treat Models as Governed Data Assets

Register every production model in Unity Catalog next to the tables it was trained on, and grant EXECUTE to groups and service principals, never to individuals. Tag each model with its owner, business domain and risk tier. When an auditor asks which high-risk models read a particular customer table, lineage and tags give a starting point in minutes, instead of a search across several workspaces and registries.

Automate Promotion With Aliases and CI/CD

Aliases only help when every team uses them the same way. Promotion follows a fixed routine:

  1. Agree a convention, such as @champion for live traffic and @challenger for a candidate under test.
  2. Define what a version must pass before it takes the champion alias: parity tolerance, quality thresholds and, for regulated use cases, fairness checks.
  3. Run those checks in CI/CD with Declarative Automation Bundles, so promotion becomes a tested pipeline step instead of a manual approval.

Plan Cost Before the First Wave

The pricing models differ. Azure ML adds no charge of its own beyond the compute and Azure services it uses, while Databricks bills Databricks Units, plus Azure VM costs on classic compute. Estimate the first wave’s spend from its real training hours and endpoint traffic, then keep it in check:

  • Job compute or serverless for training, with all-purpose compute reserved for development
  • Scale-to-zero on non-production endpoints only, since it is not recommended for production
  • Cost tags by team and model, reviewed in billing system tables after every wave

Build a Shared Feature Layer

Each wave should add to one shared set of feature tables, not create private copies. For each feature table:

  • Assign an owner.
  • Document its definition and refresh schedule.
  • Check new feature requests against what already exists.

As waves build up, more new models find their inputs already governed, which shortens both development and parity testing.

Invest in the People Moving With the Models

Most Azure ML users already know notebooks and MLflow, so the learning curve sits in Unity Catalog permissions, Lakeflow Jobs and Model Serving. Name a champion in each data science team and run hands-on sessions built around that team’s own models. Azure ML users already working with MLflow tracking, as Microsoft recommends for v2 code, can keep that habit on Databricks.

Our Unity Catalog migration guide and Databricks cost optimization guide cover governance and spend controls in more depth.

How LatentView Helps With Azure ML to Databricks Migration

As a Databricks Elite C&SI partner, we bring a Databricks Center of Excellence of 400 data engineers, with 150+ certified specialists. For a leading US tech company, we moved legacy ETL processes and data from AWS-hosted, Postgres-backed Python pipelines to Databricks in under six months, with Unity Catalog governance, as this case study shows.

Turn your Azure ML estate into governed, production-ready ML on Databricks with a team backed by a Fortune 500 client base. We combine data engineering, MLOps and data science expertise to move features, models and endpoints wave by wave, and UCMate, our Unity Catalog migration solution, supports the governance setup. Plan your first model migration with a Data Engineering Expert.

Frequently Asked Questions About Azure ML to Databricks Migration

1. Can MLflow Experiments Move From Azure ML to Databricks?

Yes. Azure ML supports MLflow tracking, so key runs and artifacts can be read with MLflow APIs and re-logged in Databricks.

2. What Replaces Azure ML Managed Online Endpoints on Databricks?

Databricks Model Serving replaces managed online endpoints, serving Unity Catalog models as autoscaling REST endpoints with traffic splitting and request logging.

3. Does Data Need to Move Out of Azure Data Lake Storage?

No. Data in ADLS Gen2 usually stays in place and is governed by Unity Catalog through external locations, Delta tables and volumes.

4. What Happens to Azure ML Prompt Flow Workloads?

Prompt flow retires on April 20, 2027. On Databricks, flows are rebuilt as agents, with MLflow tracing and evaluation, and deployed through Model Serving or Databricks Apps.

5. How Long Does an Azure ML to Databricks Migration Take?

Timelines vary with model count, endpoint traffic and client applications. The Phase 1 wave plan gives the first realistic estimate, and the first wave takes longest because it builds the templates later waves reuse.

LatentView Analytics has been helping enterprises make data-driven decisions for nearly 20 years. The company brings deep expertise in data engineering, business analytics, GenAI, and predictive modeling to 50+ Fortune 500 clients across tech, retail, financial services, and CPG. A publicly traded company serving the US, India, Canada, Europe, and Singapore, LatentView is recognized in Forrester's Customer Analytics Service Providers Landscape.

CATEGORY

LatentView Analytics has been helping enterprises make data-driven decisions for nearly 20 years. The company brings deep expertise in data engineering, business analytics, GenAI, and predictive modeling to 50+ Fortune 500 clients across tech, retail, financial services, and CPG. A publicly traded company serving the US, India, Canada, Europe, and Singapore, LatentView is recognized in Forrester's Customer Analytics Service Providers Landscape.

Take to the Next Step

"*" indicates required fields

consent*

Related Blogs

This guide helps data leaders plan a move from Amazon EMR to Databricks, from choosing a…

This guide helps data leaders plan a move from Palantir Foundry to Databricks, decide which workloads…

This guide helps data leaders move pipelines from Azure Data Factory to Databricks one use case…

Scroll to Top