SageMaker v3 drops the framework estimators, and your training code is the migration

The mechanical change worth reading twice is where your code lives when the job starts. The 2021 script mode pitch was that managed framework containers from AWS let you run custom training and inference code without building or maintaining Docker images ^[4]. In v3, the SDK syncs a local `source_dir` into the container at launch through a `SourceCode` object, and the code is never baked into the image ^[7]. That relocates the versioning problem rather than removing it. AWS describes its own example training container as intentionally minimal, carrying only the runtime and framework libraries and no training code, so the same image can be reused across scikit-learn models ^[13].

The rewrite is not a find and replace. Three import paths collapse into one training class ^[5][14], but the argument surface is different: training entry is a `command` string inside `SourceCode`, while inference entry is an `entry_script` in the same object ^[8]. Deployment moves at the same time, because `ModelBuilder` takes over from the v2 Model and Predictor pattern [c2b], and the tabular example in the post serves the model through DJL Serving ^[11]. A team with a working estimator and a hand-written inference handler is touching both halves and then re-testing a live endpoint.

Then there is the prerequisite list, which reads oddly against the original selling point. It asks for `sagemaker>=3.0`, an S3 bucket, an IAM execution role, and a training container image pushed to Amazon ECR ^[12]. The post supplies Docker build and push shell scripts and notes that Docker must be installed on the environment doing the build ^[1]. An AWS Deep Learning Container or a third-party image is permitted ^[9], so the strict requirement is to choose an image, not to author one. The walkthrough still opens with a Dockerfile ^[1].

So the accurate description for a bring-your-own-model shop is a faster loop on script changes ^[3], bought with an image you now own, tag and keep. The SDK makes no assumption about what is inside it [c5b], which is freedom in the first week and a patching queue in the second year.