AWS to Azure Migration (ML Models)

Question

AWS to Azure Migration (ML Models)

wfhDataEngineer 0

Hi All,

I have a requirement to migrate my AWS Python code from Sagemaker to Azure cloud.

For doing this I have two options :

Option 1 - Azure ML

Option 2 - Azure + Databricks ML

My Requirements :

Cost Effective
Les Complicated
Minimal Code Changes(Panada, ML Libraries)
Scalable solution

Can you suggest which one is the best option and why ??

Thanks,

2

Answer accepted by question author

1 additional answer

Your answer

Answer 1

Sridhar M 2,840 Microsoft External Staff Moderator

Hi wfhDataEngineer

Welcome to Microsoft Q&A and Thanks for sharing the details.

For a straightforward migration with minimal changes, Azure Machine Learning might be the better choice .

Azure Machine Learning (Azure ML)

Benefits:
- It's designed for easy integration and offers a suite of tools tailored for machine learning workflows.
- You can leverage Azure's automated machine learning features, which can handle many tasks with minimal manual coding adjustments.
- Azure ML provides scalability options and better manageability with models at various stages through its Model Registry.
Considerations: While Azure ML has many features and a user-friendly interface, depending on your current workloads and specific use cases, the initial learning curve might require some time investment.

Azure + Databricks ML

Benefits:
- Databricks offers a collaborative platform built around Apache Spark, suitable for big data processing and machine learning, which could be beneficial if your models are data-intensive.
- If you're familiar with Spark and want to harness its distributed processing capabilities, this might be the best option.
- Depending on your current setup and the amount of data processed, Databricks could provide a cost-effective solution due to its pay-per-use pricing model.
Considerations: If your team is not familiar with Databricks or Spark, it might involve a steeper learning curve. Additionally, migrating specific ML objects and ensuring compatibility with MLFlow can be complex.

Conclusion

For a straightforward migration with minimal changes, Azure Machine Learning might be the better choice as it aligns more closely with your needs for simplicity and reduced code changes. If you're looking for scalability and your workloads involve significant data, Databricks may also be a strong candidate.

Recommendations:

Review your current workload requirements against both platforms. Consider a small test migration with both options to measure performance and integration ease before fully committing.

I hope this helps you make a decision! Feel free to reach out if you have more questions or need further clarifications.

Please "Accept the answer and vote up" to help other community members seeking a similar solution.

wfhDataEngineer 0 Reputation points

2025-12-05T09:10:31.2833333+00:00

thanks for sharing this. Is there any standard Architecture for ML Flow that I can follow ? Documentation is quite confusing.

Is this stable or have you faced any scaling issues ??
Sridhar M 2,840 Reputation points Microsoft External Staff Moderator

2025-12-05T16:28:42.58+00:00
Hi wfhDataEngineer

MLflow is flexible, most teams converge on three common reference architectures depending on scale and maturity.

Small Team / Local-First Architecture:

Use when: 1–5 users, experiments mostly on laptops or a single server.

Components:

MLflow Tracking Server: hosted on a VM or container

Backend Store: SQLite or PostgreSQL

Artifact Store: local filesystem or S3/GCS/Azure Blob

Model Registry: built into MLflow

Pros: Simple, minimal ops Cons: Not ideal for high concurrency or large artifacts\

Mid-Scale / Production Team Architecture (Most Common):

Use when: 5–50 users, CI/CD integration, model promotion workflows

Components:

MLflow Tracking Server (with authentication + Nginx or API gateway)

Backend Store: PostgreSQL (strongly recommended)

Artifact Store: Cloud Storage (S3/GCS/Azure Blob)

Model Registry: MLflow built-in

Optional:

MLflow Model Serving (basic use-cases)

Databricks / SageMaker / K8s for production serving

Architecture diagram (conceptually):

Users / Notebooks / CI | (HTTPS) | +-----------------------+ | MLflow Tracking Server| +-----------------------+ | | v v PostgreSQL S3/GCS/Azure (metadata) (artifacts)

Pros:

Stable, widely used

Easy to scale

Supports multi-user workflows

Cons:

Needs some infra & DevOps discipline

Enterprise / High-Scale Architecture:

Use when: 50+ users, heavy job scheduling, many experiments

Components:

MLflow Tracking Server deployed behind Load Balancer + API Gateway

Highly available PostgreSQL/MySQL backend (clustered)

Artifact Store: S3/GCS with lifecycle policies

Kubernetes-based MLflow deployments

Feature Store + Model Deployment System (Feast / Vertex AI / SageMaker / Databricks)

Pros:

Fully scalable

Highly reliable

Cons:

Requires platform engineering

MLflow is quite stable — but scaling depends on your backend choices.

Where people run into issues:

SQLite backend → breaks under concurrency

Common issue when multiple users write simultaneously

Fix: always use PostgreSQL or MySQL in production

Artifact storage on local disk → slow + not scalable

Large models or images quickly overwhelm local storage

Fix: use S3 / GCS / Azure Tracking server under load:

MLflow is not CPU-heavy, but high QPS from many training jobs can overwhelm it

Fix: deploy behind nginx or load balancer Model registry with huge metadata:

Use a proper DB config (connection pooling, indexes) MLflow model serving:

OK for small loads

Not recommended for production-level high traffic

Better alternatives: KServe, Sagemaker, Vertex AI, BentoMLMLflow is flexible, most teams converge on three common reference architectures depending on scale and maturity
Sridhar M 2,840 Reputation points Microsoft External Staff Moderator

2025-12-08T17:18:00.3533333+00:00

Hi wfhDataEngineer

Following up to see if the below answer was helpful. If this answers your query, do click Accept Answer and Yes for was this answer helpful
Sridhar M 2,840 Reputation points Microsoft External Staff Moderator

2025-12-09T16:25:42.6333333+00:00

Hi wfhDataEngineer

Following up to see if the below answer was helpful. If this answers your query, do click Accept Answer and Yes for was this answer helpful

Answer 2

Hi ,

Thanks for reaching out to Microsoft Q&A.

Choose Azure ML. It is simpler, cheaper, and requires fewer code changes than Azure + Databricks ML while remaining fully scalable. If your priorities are cost control, low complexity, minimal code changes, and scalability, the better option is Azure ML.

Azure ML gives you a managed ML environment that is closer to how SageMaker works. Your existing Python, Pandas, and ML libraries will run with almost no changes. You get built-in compute, managed endpoints, environment/version control, and MLOps without needing to stitch multiple services together. It is simpler and usually cheaper for small to mid-scale workloads.

Databricks ML is powerful but adds extra cost and operational complexity. It shines when you already have large-scale Spark workloads or need collaborative notebooks across big engineering teams. For straightforward Python ML migration, it is overkill.

Please 'Upvote'(Thumbs-up) and 'Accept' as answer if the reply was helpful. This will be benefitting other community members who face the same issue.

Share via

AWS to Azure Migration (ML Models)

1 additional answer

Your answer