adapt-inference

# Adapt Inference (Batch + Realtime) Target files: - `notebooks/3_inference/batch_inference.ipynb` - `notebooks/3_inference/realtime_inference.ipynb` Target job: `resources/3_batch_inference_job.yml` Both notebooks always load the model via the `@champion` alias — never hardcode a version. ## When to use Use this skill whenever the user wants to: - Replace the sample input with their own new/unseen data. - Change the inference output table or its schema. - Enable Lakehouse Monitoring on inference results. - Call the realtime serving endpoint with a custom payload. ## Batch inference Steps for `batch_inference.ipynb`: 1. **Replace the input data source.** Swap the sample data for the customer's actual new/unseen data. Same options as data ingestion: Delta table, cloud storage, JDBC, API. 2. **Match the input schema** to what the model expects. Use `mlflow.models.get_model_info(...).signature` to verify, or call `loaded_model.metadata.get_input_schema()`. Cast / reorder columns to match. 3. **Update prediction post-processing.** Examples: - Classification: map class indices back to label strings. - Regression: clip predictions to a valid range. - Forecasting: attach forecast horizon timestamps. 4. **Rename the inference table.** Replace `iris_inferences` with a meaningful name. Keep the three-level reference `{catalog_name}.{schema_name}.<table_name>`. 5. **Keep Change Data Feed (CDF) enabled** at the end of the notebook: ```python spark.sql(f"ALTER TABLE {table_name} SET TBLPROPERTIES (delta.enableChangeDataFeed = true)") ``` This enables [Lakehouse Monitoring](https://docs.databricks.com/aws/en/lakehouse-monitoring/) on the inference table. Drop it only if you have no monitoring plan. 6. **Preserve idempotency.** Existing logic creates the table on first run and appends afterwards. Keep that pattern. 7. **Sync the job YAML.** Update `resources/3_batch_inference_job.yml`: notebook path, `base_parameters`, and `email_notifications.on_failure`. ## Realtime inference Steps for `realtime_inference.ipynb`: 1. **Update `sample_input`** to match the model's input schema (same as batch). 2. **Derive the endpoint name** from the convention `{catalog_name}-{schema_name}-{model}-endpoint` (dots replaced with dashes). The deployment notebook creates the endpoint with this exact name. 3. **Set authentication** via the SDK or REST. The notebook uses the workspace context by default; for external callers, document how to obtain a token (service principal preferred). ## Parameterization contract Both notebooks read `catalog_name` / `schema_name` widgets, and the batch job passes them through `base_parameters`. Keep the three-level UC reference convention from `mlops-quickstart-overview` for all tables and models. ## Edge cases - **Schema mismatch at scoring time**: catch with the MLflow signature check up front and emit a clear error before scoring. Don't silently reorder columns. - **Streaming inference**: split the batch notebook into a `readStream` + `foreachBatch` writer; the model load and post-processing logic stay the same. - **Cold-start endpoint**: if `scale_to_zero_enabled` is true, the first request can take seconds to warm up. Document this in the realtime notebook so consumers know what to expect. - **PII in inferences**: apply column masks via Unity Catalog on the inference table rather than masking in the notebook. - **Model not yet promoted**: if `@champion` does not exist yet (first deployment hasn't run), the notebook should fail with a clear message pointing the user at the deployment job.