🐾Does AWS retire SageMaker Experiments and Model Registry?🐾
🤓 After the launch of MLflow in SageMaker Studio many of us started asking: is it a retirement of SageMaker native functionality for models and experiments tracking? The functionality provided by both of them is quite similar, so this question is reasonable. Moreover in AWS documentation you can find recommendation to use MLflow, so let’s dive deeper into this topic.
TL;DR
IMHO, SageMaker Experiments eventually will be deprecated and replaced by MLflow. As Model Registry approval flow is a foundation for MLOps, I doubt that it will be deprecated before new mechanism is released. Also I am concerned about the price, because it’s much higher than Classic Experiments (you pay only for S3).
SageMaker Classic Ecosystem vs MLflow
Experiment tracking using the SageMaker Experiments Python SDK is only available in Studio Classic. From my perspective, it looks like a first step for further deprivation of Experiments Classic experience, but there are no official statements from AWS team. You pay only for the underlying compute and storage resources, so pricing is based on number of artefacts stored on S3.
You can deploy MLflow Models to SageMaker Inference without building custom MLflow containers. You can use all MLflow capabilities, such as MLflow Tracking, MLflow Evaluations, and MLflow Model Registry. But pricing for small MLflow server in Frankfurt region is $0.886 per hour, so you will pay ~$650 per month only for server.
As MLflow models are automatically promoted to Model Registry, looks like it will continue working in the future. Moreover, judging from available EventBridge events there are no events like “Champion” and “Challenger” alias updates, so currently MLOps still should be based on SageMaker Model Registry.
🎁 Serverless MLflow
For the ones who are still interested in MLflow, but are not willing to pay at least 500$ per month for that, recently I came across the serverless solution by InfinStor. It uses AWS Lambda, S3 and DynamoDB to setup MLflow server and Cognito for authentication.
Thank you for reading, let’s chat 💬
💬 What do you think about MLflow integration with SageMaker?
💬 Have you encountered any issues while using MLflow server?
💬 Do you want me to create a detailed post about MLflow MLOps flow?
I love hearing from readers 🫶🏻 Please feel free to drop comments, questions, and opinions below👇🏻
