Cloud-native ML Services
Everything in this course — training machines, an experiment tracker, a model registry, serving endpoints, monitoring — can be assembled from open-source tools. Each cloud also sells the whole kit pre-assembled. Choosing between the two is one of the first real MLOps decisions a team makes.
The rule
Managed platforms trade control for less work; count who will run it.
The three main managed ML platforms are Amazon SageMaker (AWS), Vertex AI (Google Cloud) and Azure Machine Learning. Their feature lists line up closely:
- Notebooks hosted in the cloud, near the data.
- Training jobs: you submit a script or container; the platform starts machines, runs it, saves the output to storage and shuts the machines down.
- Pipelines that chain steps like prepare, train and evaluate.
- Experiment tracking and a model registry.
- Endpoints: deploy a registered model behind an HTTPS API with autoscaling, often with built-in traffic splitting for gradual releases.
- Monitoring that compares live inputs with the training data.
What you gain: nobody on the team has to install, patch, secure and scale these services. What you give up:
- Portability. Pipeline code written against one platform's SDK is hard to move to another cloud.
- Control. You work within the platform's options for machines, networking and versions.
- Cost at scale. Managed compute usually costs more per hour than the same raw machine; at large, steady usage the difference adds up.
Many teams mix: open-source MLflow for tracking, a managed training job for compute, and their own Kubernetes for serving.
The shortcut
Ask "Who would run this if we built it ourselves?" If the honest answer is "nobody has time", buy the managed service. If you have a platform team, strict constraints or large steady workloads, owning it can pay off.
Count the people
Two ML engineers, no platform team, no one on call for an MLflow server.
That last step is the habit worth copying: use the platform, but keep your own code free of it where it is cheap to do so.
Drills
Answer fast — trust the rule.
Which is Google Cloud's managed ML platform?
What does a managed "training job" do for you?
Which is a real cost of going all-in on one managed platform?
A company has a ten-person platform team and huge, steady training workloads. Which choice is most defensible?
Remember
- Managed platforms trade control and portability for less operational work.
- Ask who would run it if you built it; "nobody" means buy.
- Trap: tying every line of training code to one platform's SDK when plain code would do.