Databricks-ML-Pro ML Ops — 15 Questions

Question 1

A company uses MLflow in Databricks to track experiments. They want to ensure that every experiment run is associated with a specific git commit hash to ensure reproducibility. What is the best way to achieve this?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Use Databricks Repos to automatically log the git hash for notebook runs.. MLflow automatically logs the git commit hash if the notebook is attached to a Databricks Repo. By using Repos, the link between the code version and the MLflow run is established natively, providing full traceability for every experiment. This is essential for auditability and debugging, as it allows engineers to reconstruct the exact environment and code state used during any specific training run.

Answer

Manually add the git hash as a tag to every MLflow run.

Answer

Write a custom script to query the git status and log it as a parameter.

Answer

Upgrade to the Databricks Premium tier to enable git logging.

Question 2

Refer to the exhibit. An engineer attempts to delete a model version but receives this error. What is the most likely cause?

Accepted Answer

The model version must be transitioned to the 'None' stage before deletion.. In MLflow, model versions cannot be directly deleted while they are in a registered stage (such as 'Production' or 'Staging'). To delete a version, it must first be transitioned to the 'None' stage (archived). This safety mechanism prevents accidental deletion of active models, ensuring that production pipelines remain stable and that models are explicitly decommissioned before being removed from the registry.

Answer

The model version does not exist in the registry.

Answer

The user lacks permissions to delete models from the registry.

Answer

The model name is incorrectly specified in the API call.

Question 3

Your team is using MLflow Model Registry to manage production-ready models. You need to ensure that only models reviewed by the lead data scientist are deployed to the production endpoint. Which workflow provides the most secure and scalable approach?

Accepted Answer

Configure an MLflow Webhook to trigger a CI/CD job that executes unit tests and requires manual approval before promoting to production.. Utilizing MLflow Webhooks in conjunction with registry transition requests ensures a formalized, automated approval process. By triggering a CI/CD pipeline upon a 'Pending' transition request, you decouple manual intervention from the deployment automation. This approach enforces governance, maintains audit logs of who approved the transition, and ensures that deployment only occurs after successful validation steps are completed within the CI/CD pipeline runners, significantly reducing manual error risks.

Answer

Grant all data scientists 'Manage' permissions on the Model Registry to allow direct movement to production.

Answer

Manually update the model version alias to 'production' via the Databricks UI without external pipeline triggers.

Answer

Implement a Python script that uses the MLflow client to force move models to the 'production' stage based on time-based schedules.

Question 4

Which TWO actions should be taken when setting up a Feature Store in Databricks to ensure data consistency between offline training and online serving?

Accepted Answer

Define all features in a centralized Feature Store registry using a consistent schema.. Ensuring consistency between offline and online features is critical for avoiding training-serving skew. By using a unified feature table definition, you ensure the same transformation logic is applied. Integrating with Delta Lake for offline storage provides point-in-time correctness for training, while the online store handles low-latency retrieval. These two components work together to provide a consistent feature vector, which is essential for accurate model inference in high-throughput production environments.

Answer

Use separate transformation logic for the online store to optimize for lower latency.

Answer

Exclusively store feature values in memory to maximize retrieval speed for online serving.

Answer

Avoid using Delta tables for the offline store to improve throughput performance.

Question 5

Refer to the exhibit. A MLOps engineer is troubleshooting a deployment pipeline. The model version 14 has reached the registry but is not triggering the automated deployment job. What is the most likely cause?

Accepted Answer

The model version status is 'PENDING_REGISTRATION', preventing it from being visible to transition webhooks.. The exhibit shows the model is in a 'PENDING_REGISTRATION' state, which often occurs if the logging process is interrupted or if the registry transition logic is incorrectly configured. Most automated CI/CD triggers are designed to react to 'READY' or specifically 'STAGING' status transitions. The system ignores models that have not completed the registration handshake, preventing incomplete or malformed artifacts from being mistakenly picked up by production deployment automation.

Answer

The model version is missing a 'production' tag.

Answer

The validation score of 0.88 is too low to trigger the deployment script.

Answer

The 'team' tag is not authorized to trigger production deployments.

Question 6

Which THREE criteria are essential for a robust model validation strategy before transitioning a model to the 'Production' stage in MLflow?

Accepted Answer

Validation of the model schema against the expected input format for the production endpoint.. A robust validation strategy must confirm the model's accuracy, compatibility with the production environment, and adherence to safety constraints. By performing automated testing, infrastructure validation, and human review, organizations ensure that the model behaves predictably. These stages prevent the release of models that might fail due to environmental discrepancies or produce biased results, ensuring high system reliability and trust in the ML-driven outputs provided to end users.

Answer

Hard-coding the production credentials directly into the notebook for easy access.

Answer

Deploying the model to production and monitoring it for 24 hours to validate success.

Question 7

Refer to the exhibit. The model serving endpoint is failing during high-latency requests. What is the most appropriate configuration change to resolve this?

Accepted Answer

Increase the SCORING_TIMEOUT setting to allow more time for model inference.. The exhibit shows a 'RequestTimeoutException' occurring at 60 seconds. This indicates the model inference logic is taking longer than the allowed timeout threshold. Increasing the 'SCORING_TIMEOUT' allows the model more time to complete heavy computations or complex data transformations. This is a common requirement for models that handle large input payloads or complex feature engineering, ensuring stability and reducing premature failure rates during peak load.

Answer

Set LOG_LEVEL to 'DEBUG' to get more information about the timeout.

Answer

Disable ENABLE_AUTO_REGISTRATION to prevent the model from being deployed to PROD.

Answer

Change the DEPLOY_ENVIRONMENT to 'STAGING' to bypass production constraints.

Question 8

Your organization wants to prevent models with known vulnerabilities from being deployed. Which Databricks feature should be implemented to scan model dependencies?

Accepted Answer

Integrate an automated dependency scanner into your CI/CD pipeline that validates the model's 'requirements.txt'.. Databricks Model Serving and integrated CI/CD pipelines can be configured to integrate with security scanning tools. By checking the model's environment dependencies (like the Conda or pip requirements) against vulnerability databases, you ensure that only secure models reach production. This security-first approach is crucial for enterprise MLOps, as it prevents the introduction of malicious packages or outdated libraries that could pose significant business risks.

Answer

Use the MLflow UI to manually check the library versions for every model.

Answer

Rely on the default environment provided by Databricks, which is always secure.

Answer

Configure the Model Registry to automatically delete models that use outdated libraries.

Question 9

Your team is deploying a model to Databricks Model Serving. You need to ensure the endpoint remains available during a library update that changes the inference environment. Which strategy minimizes downtime for production clients?

Accepted Answer

Create a new endpoint version and use traffic splitting to migrate clients gradually.. Blue-Green deployment allows you to route traffic to the new version only after verifying its health. This is a standard MLOps practice in Databricks to prevent downtime. By shifting traffic incrementally or fully, you maintain high availability and can instantly roll back if performance metrics degrade, ensuring that critical production pipelines remain unaffected by library dependencies or runtime environment modifications.

Answer

Update the existing endpoint configuration directly and wait for the auto-restart process.

Answer

Delete the endpoint and recreate it with the new configuration during a scheduled maintenance window.

Answer

Deploy the model as a batch inference job instead of a real-time endpoint.

Question 10

You are auditing a Databricks ML workspace to improve governance. Which TWO features in Unity Catalog are essential for tracking the lineage of ML models and their associated data assets?

Accepted Answer

System tables for lineage queries.. Unity Catalog captures automated lineage at the data and model levels, which is crucial for compliance and reproducibility. By linking models to their source tables, data scientists can trace back data quality issues that impact prediction performance. This integration ensures that models are transparent, auditable, and compliant with internal governance policies regarding data access and provenance.

Answer

Delta Sharing for model distribution.

Answer

MLflow Experiment tracking.

Answer

Feature Store workspace-level tags.

Question 11

Refer to the exhibit. A data scientist attempts to promote a model to 'Production' in the Unity Catalog Model Registry and receives this error. What is the most likely cause?

Accepted Answer

The user is missing the required permissions to modify the model object.. The error explicitly states that the user lacks the required permission level. In Unity Catalog, transitioning a model version status typically requires 'CAN_MANAGE' or 'APPLY_TAG' permissions depending on the specific governance settings. This highlights the importance of Role-Based Access Control (RBAC) in ensuring that only authorized personnel can promote models to sensitive production environments, preventing accidental or malicious changes.

Answer

The model version is already in the 'Archived' state.

Answer

The model depends on an external library that is not installed.

Answer

The Unity Catalog metastore is currently offline.

Question 12

Which THREE criteria are required for an MLflow model to be eligible for 'Model Serving' on Databricks?

Accepted Answer

A defined input signature.. Model serving requires a well-defined format to ensure the endpoint can load and execute the model. The model must be logged with a signature, have the correct dependency environment defined, and use a compatible framework. These requirements guarantee that the serving environment can resolve all necessary libraries and understand the input/output schema for successful request processing.

Answer

The model must be stored as a raw CSV file.

Answer

The model must be in the 'Archived' registry state.

Question 13

Your team needs to implement a 'Shadow Deployment' strategy in Databricks. What is the correct approach to achieve this?

Accepted Answer

Configure the serving endpoint to log shadow model predictions.. Shadow deployment involves routing the same real-time traffic to both the current production model and the new candidate model. The production model returns the actual response, while the shadow model's prediction is logged for comparison. This allows you to evaluate the new model's performance on live production data without exposing it to customers, providing a risk-free way to validate accuracy and latency before a full rollout.

Answer

Deploy two separate endpoints and use a load balancer to split traffic.

Answer

Run two inference jobs in a single notebook periodically.

Answer

Manually copy request data from the production logs for re-testing.

Question 14

An ML engineer needs to deploy a custom MLflow PyTorch model to a Databricks Model Serving endpoint with maximum throughput and minimum latency. The model requires specialized GPU acceleration and depends on custom inference preprocessing logic. Which architectural approach best fulfills this requirement in Databricks Model Serving?

Accepted Answer

Define a custom MLflow pyfunc model wrapper containing the preprocessing logic and provision the endpoint using GPU-accelerated instance types.. Deploying a custom PyTorch model with preprocessing requires leveraging MLflow model pyfunc custom flavor definitions packaged inside the model artifact. Databricks Model Serving natively supports GPU instance types such as GPU_SMALL for accelerated deep learning inference workloads. Using native endpoints ensures auto-scaling, managed infrastructure, and zero-downtime rolling updates aligned with enterprise MLOps best practices.

Answer

Package the model with a custom pyfunc signature and deploy it to a CPU-based Serving endpoint using standard instance configurations.

Answer

Export the model weights to an external S3 bucket and invoke them via an external API gateway bypassing Databricks serving infrastructure.

Answer

Run an interactive notebook continuously in production that reads requests from a Delta table and writes predictions back periodically.

Question 15

Your team is using MLflow Model Registry to manage the lifecycle of a forecasting model. You need to transition a model version from 'Staging' to 'Production' only after it passes an automated integration test suite in the CI/CD pipeline. Which approach is the most scalable way to implement this gate?

Accepted Answer

Implement a Python script that uses the MLflow Client API to transition versions triggered by a webhook.. Integrating MLflow Webhooks with a CI/CD orchestration tool like GitHub Actions or Jenkins ensures that model transitions are governed by external quality gates. This pattern decouples environment promotion from manual intervention, reducing human error. It is critical in MLOps to enforce automated compliance, ensuring only models that satisfy performance thresholds move to production, thereby maintaining system reliability and auditability in professional Databricks environments.

Answer

Manually update the model version status in the MLflow UI after each successful test run.

Answer

Add a tag to the model version using the MLflow SDK and check for the tag in a daily scheduled job.

Answer

Hardcode the transition logic inside the training notebook using the transition_model_version_stage method.