Reproducible training pipelines in MLOps - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Key principle of reproducible training pipelines

Which of the following is the most important principle to ensure a machine learning training pipeline is reproducible?

A Using fixed random seeds and versioning all data and code

B Running training on the fastest available hardware

C Using the latest deep learning framework version without version control

D Allowing manual changes to data preprocessing steps during training

💡 Hint

Think about what guarantees the exact same results when you run the pipeline multiple times.

Reproducibility requires controlling randomness and tracking versions of data and code so the pipeline can be rerun with the same inputs and produce the same outputs.

💻 Command Output

intermediate

Output of a pipeline run with fixed seed

Given this snippet setting a fixed random seed in Python for training, what will be the output of print(random.randint(1, 100)) if the seed is set to 42 before?

MLOps

import random
random.seed(42)
print(random.randint(1, 100))

A 42

B 15

C 82

D Error: seed must be between 0 and 1

💡 Hint

Python's random.seed sets the starting point for the random number generator.

With seed 42, the first call to random.randint(1, 100) returns 82 consistently.

❓ Configuration

advanced

Correct Dockerfile snippet for reproducible training environment

Which Dockerfile snippet best ensures a reproducible training environment by fixing package versions?

A

FROM python:3.9
RUN pip install tensorflow --upgrade

B

FROM python:3.9
RUN pip install tensorflow

C

FROM python:latest
RUN pip install tensorflow==latest

D

FROM python:3.9
RUN pip install tensorflow==2.12.0

💡 Hint

Fixing package versions avoids unexpected changes in dependencies.

Specifying exact package versions in Dockerfile ensures the environment is consistent across builds.

🔀 Workflow

advanced

Order of steps in a reproducible training pipeline

What is the correct order of these steps in a reproducible ML training pipeline?

A 1,2,3,4

B 1,3,2,4

C 2,1,3,4

D 3,1,2,4

💡 Hint

Think about what must be done before training starts and what should be logged during the process.

First version control code, then set seeds, log data versions, then run training and save outputs.

❓ Troubleshoot

expert

Cause of non-reproducible training results despite fixed seeds

You set fixed random seeds in your training pipeline, but results differ each run. Which is the most likely cause?

A Running training on the same hardware

B Using non-deterministic GPU operations or multi-threading without control

C Saving model checkpoints after training

D Not updating the training data between runs

💡 Hint

Some hardware or library operations can introduce randomness even if seeds are fixed.

GPU operations and parallelism can cause non-deterministic behavior unless explicitly controlled.