Docker for ML reproducibility in MLOps - Step-by-Step Execution

Process Flow - Docker for ML reproducibility

Write Dockerfile

↓

Build Docker Image

↓

Run Container with ML Code

↓

ML Code Executes in Container

↓

Results Reproducible Anywhere

↓

Share Image or Dockerfile

This flow shows how you create a Docker image from a Dockerfile, run your ML code inside a container, and get reproducible results anywhere.

Execution Sample

MLOps

FROM python:3.12-slim
WORKDIR /app
COPY requirements.txt ./
RUN pip install -r requirements.txt
COPY . ./
CMD ["python", "train.py"]

This Dockerfile sets up a Python 3.12 environment, installs dependencies, copies ML code, and runs training.

Process Table

Step	Action	Details	Result
1	Read Dockerfile	FROM python:3.12-slim	Base image set to Python 3.12 slim
2	Set working directory	WORKDIR /app	Working directory inside container is /app
3	Copy requirements.txt	COPY requirements.txt ./	requirements.txt copied to /app
4	Install dependencies	RUN pip install -r requirements.txt	Python packages installed
5	Copy ML code	COPY . ./	All local files copied to /app
6	Set command	CMD ["python", "train.py"]	Container will run train.py on start
7	Build image	docker build -t ml-model .	Docker image 'ml-model' created
8	Run container	docker run ml-model	ML training starts inside container
9	Training completes	train.py finishes	Model trained reproducibly inside container

💡 Training completes and container stops; environment is consistent everywhere

Status Tracker

Variable	Start	After Step 3	After Step 4	After Step 5	After Step 8	Final
Docker Image	None	Base python:3.12-slim	With dependencies installed	With ML code copied	Image built and ready	Image used to run container
Container State	None	Not started	Not started	Not started	Running ML training	Training complete

Key Moments - 2 Insights

Why do we copy requirements.txt and run pip install separately before copying all files?

Because installing dependencies first uses Docker cache efficiently. If only code changes but dependencies don't, Docker skips reinstalling packages (see execution_table steps 3 and 4).

How does running ML code inside a container help with reproducibility?

The container has the exact same environment everywhere, so the ML code runs with the same Python version and packages (see execution_table steps 8 and 9).

Visual Quiz - 3 Questions

Look at the execution table, at which step is the Python environment prepared with required packages?

A Step 6

B Step 3

C Step 4

D Step 8

💡 Hint

Check the 'Action' and 'Details' columns for pip install command in step 4

At which step does the container start running the ML training code?

A Step 8

B Step 7

C Step 6

D Step 9

💡 Hint

Look for 'Run container' action in the execution table

If you change only train.py code, which step can Docker reuse from cache to save time?

A Step 5 (copy ML code)

B Step 4 (install dependencies)

C Step 7 (build image)

D Step 8 (run container)

💡 Hint

Refer to key moment about caching and see steps 3 and 4 in execution_table

Concept Snapshot

Docker for ML reproducibility:
- Write Dockerfile with base image, dependencies, and code
- Build image to create consistent environment
- Run container to execute ML code
- Results are reproducible anywhere with same image
- Use caching by separating dependency install from code copy

Full Transcript

Docker helps make machine learning reproducible by packaging code and environment together. First, you write a Dockerfile starting from a Python base image. Then you copy your requirements.txt and install dependencies. Next, you copy your ML code and set the command to run training. Building the image creates a snapshot of this environment. Running a container from this image executes your ML training inside a consistent setup. This means your results will be the same on any machine. Docker caching speeds up rebuilds by reusing steps if dependencies don't change. This step-by-step process ensures your ML work is easy to share and reproduce.