Reading notes from 2021; revised November 17, 2026. Reported architecture and scale refer to the source period.

Separate tracking from orchestration

Experiment tracking records parameters, metrics, and artifacts. A pipeline coordinates steps and their dependencies. A registry manages model versions and promotion. Serving handles prediction requests. Comparing only product names can hide those distinctions.

Capability Design question
Tracking Can a run record its data, code, configuration, and metrics?
Orchestration How do steps retry and pass versioned artifacts?
Registry What evidence permits a candidate to be promoted?
Serving Which runtime, scaling, and rollback contract is needed?

SageMaker Pipelines is an AWS-integrated workflow service; MLflow’s tracking/model-management capabilities can be used alongside workflow tools. They are not mutually exclusive alternatives to the entire ML lifecycle.

Cost needs a workload

An illustrative comparison should include compute hours, idle serving time, storage, network transfers, and platform operating effort. Self-managing a tracking server can change the bill while adding backup, upgrades, authentication, and availability work. It is not automatically cheaper in every setting.

Do not infer that SageMaker supports only built-in algorithms from a tracking limitation. Check custom training/container support and tracking APIs for the chosen service/SDK version. The original 2021 opinions are context, not a verified description of current constraints.

Make the handoffs reproducible

Use one small train/evaluate/register workflow and record the same artifact identifiers in both the orchestrator and tracking system. See ML infrastructure tooling for the larger lifecycle.

References