Machine learning architecture: from data to a deployed model
A machine learning project is a loop: data in, a model trained and tested, predictions served, results watched, and new data fed back. Here is how the pieces fit, which service plays each part on AWS, Azure and Google Cloud, and what to settle first.
Quick answer
A machine learning project is a loop: raw data goes into object storage and is cleaned into a shared dataset, a model is trained, tracked and tested on a managed ML platform, then served as an endpoint or batch job and monitored, so it can be retrained when its quality drops.
| Best for | Predictions from your own historical data, such as scores, forecasts and classifications |
|---|---|
| Less suited when | A simple rule or report gives the answer, or there is too little data to learn from |
| Start with | A baseline model with AutoML on data you already hold |
| Main services on AWS | Amazon SageMaker Canvas, Amazon EC2 GPU instances, Amazon CloudWatch, Amazon S3, Amazon S3 + AWS Lake Formation, Amazon SageMaker AI |
| Main services on Azure | Azure Machine Learning automated ML, Azure N-series GPU VMs, Azure Monitor, Azure Blob Storage, Microsoft Fabric OneLake, Azure Machine Learning |
| Main services on Google Cloud | BigQuery ML, Compute Engine GPUs, Cloud Monitoring, Cloud Storage, BigQuery lakehouse, Gemini Enterprise Agent Platform |
The idea in plain words
A machine learning model learns patterns from past data and uses them to predict something new: a price, a risk score, a category. Building one is mostly data work, with training in the middle and a long tail of deployment and monitoring.
Each cloud has a managed machine learning platform that covers the repeatable parts: notebooks, training jobs, experiment tracking, a model registry and hosted endpoints. Around it you need storage for the data and, for deep learning, GPU capacity.
Many business problems need only a rule, a report or a simple forecast. Agree what decision the model supports, and how you will know it works, before you collect any data.
The architecture
Boxes show the job each part does. Pick a cloud to see the service for each job.
Step by step
- Collect raw dataLand the data you will learn from (files, exports, logs) in object storage, with access limited to the team that needs it.
- Prepare a shared datasetClean the data, handle missing values and add labels where needed. Keep it in a data lake with versions, so any result can be repeated.
- Train and compare modelsUse notebooks and managed training jobs to try approaches. Record each experiment and keep the best versions in a model registry. Use GPU servers for heavy deep-learning jobs, or AutoML for a quick first baseline.
- EvaluateTest on data the model has not seen, check the cases that matter most to your business, and look for unfair differences between groups. Agree a pass mark before deployment.
- DeployPublish the chosen model as a live endpoint for instant predictions, or run it as a scheduled batch job for large volumes.
- Use the predictionsYour app sends data in and gets predictions back. Keep a way to switch back to the previous version quickly.
- Monitor and retrainWatch speed, errors and the quality of predictions as real data changes (drift), and retrain when quality drops. Feed new data back to step 1.
Services side by side
The main service for each job on each cloud, with a second option where one exists. Names checked on 23 October 2026.
| Job | AWS | Azure | Google Cloud |
|---|---|---|---|
| No-code and low-code machine learning (AutoML) |
|
|
|
| GPU servers for AI and graphics |
|
|
|
| Monitoring, metrics and alerts |
|
|
|
| Object storage (files, images, backups) |
|
|
|
| Data lake and lakehouse |
|
|
|
| Build, train and deploy machine learning models |
|
|
|
Common variations
- No data science team yet. Start with AutoML or a built-in model on data you already hold, to learn whether the data carries a useful signal before investing more.
- Deep learning or very large models. Add GPU or specialised-chip capacity for training, and size serving separately. Stop servers the moment a job finishes.
- Predictions needed in real time. Use a live endpoint with autoscaling, and keep the features it needs close to it. For overnight reports, a batch job is much cheaper.
- Data is already in a warehouse. Some warehouses can train and run simple models in place, which saves moving data. Check what the one you use offers.
Managed ML platform or build it on your own servers?
The managed platforms in the table above bundle notebooks, training, a registry and endpoints. The alternative is to assemble the same steps yourself on virtual machines or containers.
Managed ML platform
Good
- Notebooks, training, registry and endpoints in one place
- Scaling, security and logging largely handled
- Faster to a working pipeline
Watch out
- Platform-specific tools can make moving harder
- Idle notebooks and endpoints keep costing money
- A learning curve for each platform's way of working
Build it on servers or containers
Good
- Any framework and full control
- Easier to keep portable across clouds
- Can reuse your existing platform tools
Watch out
- You build and run pipelines, scaling and monitoring
- You own security and cost control for every part
- Slower to start and more to maintain
Keep in mind
- Start with the question. Agree which decision the model supports and how success will be measured, before you collect data. Many problems need a rule or a report, not a model.
- Data quality and rights. Results are only as good as the data. Check that you may use it for this purpose, including consent where laws such as the GDPR or India's DPDP Act require it, and remove or protect personal data.
- Reproducibility. Version the data, the code and the model together, so you can explain and repeat any result later.
- Cost. GPU training and always-on endpoints are the big costs. Stop idle notebooks and endpoints, use batch instead of live endpoints when you can, and use discounted interruptible capacity for training where the provider offers it.
- Fairness and explanations. Check how the model treats different groups, and be able to explain important decisions, especially in lending, hiring, insurance and health.
- Drift. Models decay as the world changes. Monitor inputs and outcomes, and schedule regular retraining.
- Security. Protect the training data, the model files and the endpoint with access control and network rules. Models can reveal parts of the data they were trained on.
Words used on this page
Services in this pattern that have changed
Cloud providers rename and retire services often. These affect the parts above. Checked on 23 October 2026.
- Google Cloud Renamed Vertex AI is now Gemini Enterprise Agent Platform (22 April 2026). Vertex AI Search is now Agent Search, and Vertex AI Agent Engine is now Agent Runtime. Google says the Vertex AI documentation is no longer updated. Details
Sources and further reading
Check details such as features, limits, prices and regions in the providers' own documentation.
- AWS documentation (opens in a new tab)
- Microsoft Azure documentation (opens in a new tab)
- Google Cloud documentation (opens in a new tab)
Announcements about the changed services above
Frequently asked questions
Do I need GPUs?
Only for deep learning on large data, such as images, speech or very large models. Many business models, such as forecasts and scores on tabular data, train well on ordinary servers.
Should I use AutoML or build a custom model?
AutoML is a good first baseline and needs little code. Build a custom model when the baseline is not good enough, when you need special control, or when you must explain the model in detail.
Which cloud's machine learning platform is best?
AWS, Azure and Google Cloud each cover the same steps. Choose the cloud your data and team already use, and compare on a small pilot with your own data.
How much data do I need?
It depends on the problem. Start small to learn whether the data has a useful signal, then add data where the results show it will help.
When is machine learning the wrong tool?
When a simple rule, a report or a lookup gives the answer, when you have too little data, or when you cannot say how you would measure success.
Want this reviewed for your case?
CloudBrum architects can adapt this pattern to your documents, users, rules and budget, and estimate the cost. Nothing is sent until you choose.
Building something else?
Describe it in plain words and get matching services on all three clouds.
Service suggester → Cloud consulting →