MLOps and Deploying AI Practice Test
Practise the production side of AI: versioning, deployment strategies, monitoring, drift, feature stores and keeping LLM applications fast and affordable.
AI Skills
Before you start
- Questions
- 15
- Time
- 18 minutes
- Format
- Single section
- The timer starts when you press Start. When it reaches zero, your answers are submitted automatically.
- Use the numbered navigator to move between questions, and flag any you want to come back to.
- For typed answers, keep within the word limit. Answers over the limit are marked wrong, as in the real test.
- You can submit early. Your score, the correct answers and explanations appear straight away.
MLOps and Deploying AI Practice Test: all 15 questions with answers
Prefer to work through the test on paper? Answer the questions, then open “Show answer” to check.
Choose the correct answer, A, B, C or D.
Q1.What is MLOps?
- A.A programming language for machine learning
- B.A type of neural network
- C.A cloud provider
- D.Practices for reliably deploying, monitoring and maintaining ML systems in production
Show answer
Answer: D. Practices for reliably deploying, monitoring and maintaining ML systems in production
MLOps applies DevOps ideas such as automation, testing and monitoring to machine learning.
Q2.Why should you version datasets and models?
- A.So you can reproduce results and roll back to an earlier version
- B.To make files larger
- C.Because it is required by Python
- D.To slow down training
Show answer
Answer: A. So you can reproduce results and roll back to an earlier version
Knowing exactly which data and code produced a model makes debugging and rollback possible.
Q3.What is data drift?
- A.Data stored in the wrong folder
- B.The input data seen in production changes compared with the training data
- C.A bug in the database
- D.Moving data between clouds
Show answer
Answer: B. The input data seen in production changes compared with the training data
Drift means the model is now seeing different inputs from those it was trained on, which can reduce accuracy.
Q4.In a shadow deployment, a new model
- A.replaces the old model for all users at once
- B.is only tested on old data
- C.receives live traffic, but its predictions are not shown to users
- D.is deleted after training
Show answer
Answer: C. receives live traffic, but its predictions are not shown to users
Shadow mode compares the new model with the current one on real traffic without any risk to users.
Q5.What is a canary release?
- A.Releasing only at night
- B.Releasing without testing
- C.Releasing to every user at once
- D.Releasing a new version to a small share of users first
Show answer
Answer: D. Releasing a new version to a small share of users first
If the small group sees problems, you roll back before most users are affected.
Q6.Why package a model service in a Docker container?
- A.So it runs in the same environment on every machine
- B.To make the model more accurate
- C.To avoid writing any code
- D.To remove the need for a server
Show answer
Answer: A. So it runs in the same environment on every machine
Containers bundle the code, libraries and settings so 'it works on my machine' problems disappear.
Q7.When is batch inference a good fit?
- A.When each prediction must be returned within milliseconds of a user action
- B.When predictions can be computed on a schedule, such as nightly recommendations
- C.When there is no data
- D.When the model has not been trained
Show answer
Answer: B. When predictions can be computed on a schedule, such as nightly recommendations
Batch jobs are cheaper and simpler when results do not need to be instant.
Q8.What do experiment tracking tools such as MLflow record?
- A.Only the final model file
- B.User passwords
- C.Parameters, metrics and artefacts for each training run
- D.Website traffic
Show answer
Answer: C. Parameters, metrics and artefacts for each training run
Tracking makes it easy to compare runs and reproduce the best one.
Q9.What is training-serving skew?
- A.A model that trains too slowly
- B.Too many users at once
- C.Different GPU brands
- D.Features computed differently in training and in production
Show answer
Answer: D. Features computed differently in training and in production
If the production code calculates a feature differently, the model sees inputs it never learned from.
Q10.What problem does a feature store help solve?
- A.Storing images for a website
- B.Selling features to customers
- C.Replacing the model
- D.Keeping feature definitions consistent between training and serving
Show answer
Answer: D. Keeping feature definitions consistent between training and serving
A shared feature store computes features once and serves the same values for training and inference.
Q11.Which set of metrics is most useful to monitor for an LLM-powered endpoint?
- A.Latency, error rate, token cost and quality signals such as user feedback
- B.Only CPU temperature
- C.Only the number of lines of code
- D.Only disk space
Show answer
Answer: A. Latency, error rate, token cost and quality signals such as user feedback
LLM apps can fail on speed, cost and quality, so all three need watching.
Q12.What is a model registry?
- A.A list of employees
- B.A central store of model versions with their metadata and status
- C.A government licence
- D.A type of database index
Show answer
Answer: B. A central store of model versions with their metadata and status
Registries track which model version is in staging or production and how it was produced.
Q13.In CI/CD for machine learning, what should run automatically before a model change is deployed?
- A.Nothing; deploy immediately
- B.A manual email to every user
- C.Tests and an evaluation against a fixed dataset
- D.Deleting the old model
Show answer
Answer: C. Tests and an evaluation against a fixed dataset
Automated tests and evaluations catch regressions before they reach users.
Q14.How can caching help an LLM application?
- A.It makes answers more creative
- B.It trains the model
- C.It removes the need for prompts
- D.It reduces cost and latency for repeated requests
Show answer
Answer: D. It reduces cost and latency for repeated requests
Identical or repeated requests can be answered from the cache instead of calling the model again.
Q15.Why should every deployment have a rollback plan?
- A.So you can quickly return to the previous version if the new one performs worse
- B.To deploy faster without testing
- C.Because models never fail
- D.To save storage space
Show answer
Answer: A. So you can quickly return to the previous version if the new one performs worse
Even well-tested models can behave differently on live traffic.