ML Data Scientist Databricks-Machine-Learning-Professional Syllabus Exam Questions Answers

Streaming with Spark as a model deployment strategy means applying a machine learning model to data streams that are processed incrementally and continuously by Spark Structured Streaming. Spark Structured Streaming is a scalable and fault-tolerant stream processing engine that enables complex analytics on live data streams using the Dataset/DataFrame API1. Spark Structured Streaming supports various sources and sinks for streaming data, such as Kafka, Kinesis, TCP sockets, Delta tables, etc2. Spark Structured Streaming also supports various types of operations on streaming data, such as aggregations, windowing, joins, and stateful transformations3. To deploy a machine learning model on streaming data, you can use the MLflow model registry to managethe model lifecycle and versioning4. You can also use the MLflow model serving feature to serve the model as a REST API endpoint that can be invoked by Spark Structured Streaming5. Alternatively, you can use the UDF (user-defined function) feature to apply the model to streaming data within Spark Structured Streaming6.

The inference of incrementally processed records as soon as trigger is hit describes the streaming with Spark as a model deployment strategy. A trigger defines when the results of a streaming query should be written to the output sink. A trigger can be based on a processing time interval, a data size limit, or a continuous mode that writes the results as soon as they are available. The trigger ensures that the streaming query is executed incrementally and continuously, and the model inference is applied to the latest available data. The other options are incorrect because:

• Option A: The inference of batch processed records as soon as a trigger is hit does not describe streaming with Spark, but rather batch processing with Spark. Batch processing means applying a machine learning model to a finite set of data that is processed as a single job. Batch processing does not require a trigger, as the results are written to the output sink when the job is completed.
• Option B: The inference of all types of records in real-time does not describe streaming with Spark, but rather a generic definition of real-time processing. Real-time processing means applying a machine learning model to data streams that are processed as soon as they arrive, with minimal latency. Real-time processing does not necessarily use Spark Structured Streaming, as there are other frameworks and tools that can support it, such as Apache Flink, Apache Storm, etc.
• Option C: The inference of batch processed records as soon as a Spark job is run does not describe streaming with Spark, but rather batch processing with Spark. Batch processing means applying a machine learning model to a finite set of data that is processed as a single job. Batch processing does not depend on a Spark job, as the model inference can be done outside of Spark, such as using a REST API endpoint, a command-line tool, etc.
• Option E: The inference of incrementally processed records as soon as a Spark job is run does not describe streaming with Spark, but rather a contradiction. Incrementally processed records imply streaming processing, while a Spark job implies batch processing. Streaming processing and batch processing are different paradigms of data processing, and cannot be mixed in this way. References: Structured Streaming Programming Guide, Input Sources and Output Sinks, Operations on streaming DataFrames/Datasets, MLflow Model Registry, MLflow Model Serving, Apply machine learning models, [Triggers], [Trigger Types], [Batch Processing], [Real-time Processing], [Real-time Data Processing Frameworks], [Deploy machine learning models], [Batch vs Streaming Processing]

Drift is the change over time in the statistical properties of the data that was used to train a machine learning model. This can cause the model to become less accurate or perform differently than it was designed to1. Drift can be detected by monitoring the statistics of the input and output data over time and comparing them with the baseline statistics from the training data2. Depending on the type and severity of the drift, different responses may be appropriate. Some possible responses are:

• Retraining and deploying a model on more recent data: This can help the model adapt to the changes in the data and improve its performance. However, this may require frequent retraining and deployment cycles, which can be costly and time-consuming. Also, retraining may not be sufficient if the drift is caused by a change in the underlying concept or relationship between the input and output variables3.
• Rebuilding the machine learning application with a new label variable: This can help the model capture the new concept or relationship that has emerged in the data. However, this may require a significant redesign of the application and the data pipeline, as well as collecting and labeling new data. Also, rebuilding may not be feasible if the concept or relationship is constantly changing or unknown3.
• Sunsetting the machine learning application: This can help avoid the risks and costs of maintaining a model that is no longer reliable or useful. However, this may mean losing the benefits and value of the application and the data. Also, sunsetting may not be an option if the application is critical or mandatory for the business or the users3.

Therefore, all of these responses are probable, depending on the situation and the trade-offs involved. References:

• Databricks Machine Learning Professional Exam Guide, Section 4: Solution and Data Monitoring, p. 5
• Databricks Machine Learning Documentation, Monitoring ML Models, Data Drift Detection, p. 2-3
• A Gentle Introduction to Concept Drift in Machine Learning, Types of Concept Drift, p. 3-4
• Understanding Data Drift and Model Drift: Drift Detection in Python, Types of Drift, p. 2-3

MLflow Model flavors are a key concept that makes MLflow Models powerful: they are a convention that deployment tools can use to understand the model, which makes it possible to write tools that work with models from any ML library without having to integrate each tool with each library. MLflow defines a set of built-in model flavors for popular ML frameworks, such as scikit-learn, TensorFlow, PyTorch, etc. Each flavor defines a standard format for model data and a common interface for model inference. For example, the python_function flavor allows any model to be loaded as a Pythonfunction that can be applied to input data. MLflow also supports custom model flavors for user-defined models or community-developed flavors. MLflow Model flavors enable model portability and interoperability across different platforms and environments12 References:

• MLflow Models — MLflow 2.9.1 documentation
• Models, Flavors, and PyFuncs in MLflow

To compute predictions for spark_df when the missing feature values can be found in the Feature Store by searching for features by customer_id, you can use the following code block:

Python

# Get the missing features from the Feature Store using the model URI and the customer_id column

df = fs.get_missing_features(spark_df, model_uri, lookup_key="customer_id")

# Score the DataFrame using the model URI and the Feature Store Client

fs.score_batch(model_uri, df)

AI-generated code. Review and use carefully. More info on FAQ.

The fs.get_missing_features method takes a Spark DataFrame, a model URI, and a lookup key as arguments. It returns a new Spark DataFrame that contains the originalcolumns plus the missing features that are required by the model. The missing features are retrieved from the Feature Store by joining the DataFrame with the feature tables using the lookup key. The lookup key must match the primary key of the feature tables. The model URI must point to a registered model that was trained using features from the Feature Store1.

The fs.score_batch method takes a model URI and a Spark DataFrame as arguments. It applies the model to the DataFrame and returns a new Spark DataFrame that contains the original columns plus a prediction column. The model URI must point to a registered model that was trained using features from the Feature Store2.

The other options are incorrect because:

• Option A: fs.score_model is not a valid method name, as it is missing an underscore. The correct method name is fs.score_batch2.
• Option B: fs.score_model without getting the missing features will not work, as the model expects the DataFrame to have all the features that were used for training. The correct way is to use fs.get_missing_features before fs.score_batch12.
• Option D: fs.score_batch without getting the missing features will not work, as the model expects the DataFrame to have all the features that were used for training. The correct way is to use fs.get_missing_features before fs.score_batch12.
• Option E: fs.score_batch without specifying the lookup key will not work, as the fs.get_missing_features method requires a lookup key to join the DataFrame with the feature tables. The correct way is to use fs.get_missing_features with the lookup key “customer_id” before fs.score_batch12. References: Get missing features, Score batch