Prepare notebooks, dependencies and files

Turn an ordinary notebook function into a callable service without relying on its interactive kernel state. The result must run from the generated bundle and its declared runtime inputs.

Before you begin

Use the notebook installation path. Inspection needs the notebook extra, not pandas or scikit-learn. A Run needs the scientific packages in the interpreter running Apizr. A service needs its own transport and selected application dependencies.

Inspect an ordinary notebook first

apizr experiment inspect fraud_detection.ipynb

Use the fraud files for this command. Stored cell outputs and execution counts are not current results. Supported notebooks are nbformat 4 with ordinary Python cells. Magics and shell cells are refused; convert these operations into explicit preparation commands and inspect again.

Keep training out of service startup

Define the prediction function at module level. Put training under the exact if __name__ == "__main__": guard, without an else, and do not rebind __name__. The fraud notebook follows this pattern. A saved model is loaded by trusted prediction code on demand. Static analysis does not deserialize it.

Cells must work in document order from a clean process. A value created only by a previous interactive session is unavailable to Apizr's fresh worker. Verify with an explicit Run in the workload environment before sharing that result.

Keep the service's files explicit

In a project file, [application] resources declares individual regular files, while dependencies declares exact distribution pins. Paths retain their layout under source/. Use paths derived from __file__ so starting from another working directory does not accidentally load a different file.

For a Run-based service, select an output by recorded name (--artifact model), not an arbitrary replacement pathname. The current bytes must match that Run. Resources are limited to 16 MiB each and 32 MiB combined, with at most 128 files. This direct-service workflow is not a model registry or an arbitrary filesystem packager. See application inputs.

Prepare dependencies at the correct time

Inspection does not install dependencies. Download scientific wheels only when preparing to run the notebook. Retain the wheels used by the chosen Run for --dependency-wheelhouse; provider membership remains conditional evidence. The prediction scenario selects joblib and scikit-learn, with pandas remaining a training dependency.

Install both the generated requirements.txt and, when present, application-requirements.txt in the service environment. Application pins alone do not make an unsupported function eligible. Check a real request after startup, then follow safe regeneration for updates.