Become Rustic. MLOps without Fancy Stacks and with Zero Cost

Cover image source: Source

When you start exploring MLOps, you’ll often run into fancy tools like SageMaker, S3, Google AutoML, AzureML, and others. But you don’t need to get FOMO and jump into them immediately because MLOps is not a platform. You can begin with a simple and free setup that works.

Note: This tutorial is not only designed for entry-level practitioners. I know lot of people doing rustic architecture at scales, no shame, including myself because why not as long as it works and costs less money. You might need more complexity as your system or business scales, but you can always learn on your way.

Okay, enough talking — let’s start! Here is what we will use

MLFlow

1. Beginning#

What you need:

Google Colab
Google Drive
PyNgrok
MLflow
A bit of creativity

We’ll build and track a simple linear regression model.

2. Set Up Environment and Train ML Model#

Step 1: Prepare Google Drive#

Create a folder in your Google Drive. This will store MLOps artifacts like models, metrics, and parameter logs.

Step 2: Mount Drive and Install Dependencies#

Open a new Colab notebook and run:

1
!pip install mlflow --quiet
2
!pip install pyngrok --quiet
3

4
from google.colab import drive
5
drive.mount('/content/drive')
6

7
%cd /content/drive/MyDrive/mlflow/linear

3. Launch MLflow and Ngrok Tunnel#

Start the MLflow UI on port 80:

1
import os
2
import warnings
3
import sys
4
import pandas as pd
5
import numpy as np
6
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
7
from sklearn.model_selection import train_test_split
8
from sklearn.linear_model import ElasticNet
9
from urllib.parse import urlparse
10
import mlflow
11
import mlflow.sklearn
12
import logging
13

14
logging.basicConfig(level=logging.WARN)
15
logger = logging.getLogger(__name__)
16

17
get_ipython().system_raw("mlflow ui --port 80 &")

Set up Ngrok#

Get your token from ngrok.com and save it in a text file:

1
from pyngrok import ngrok
2

3
# Terminate open tunnels if exist
4
ngrok.kill()
5

6
# Read token securely from a file
7
file1 = open('token.txt', 'r')
8
NGROK_AUTH_TOKEN = file1.readlines()
9
ngrok.set_auth_token(NGROK_AUTH_TOKEN[0])
10

11
# Open a secure tunnel on port 80
12
public_url = ngrok.connect(port="80")
13
print("MLflow Tracking UI:", public_url)

Now you have a public link to monitor your ML experiments.

4. Define Evaluation Metrics#

1
def eval_metrics(actual, pred):
2
    rmse = np.sqrt(mean_squared_error(actual, pred))
3
    mae = mean_absolute_error(actual, pred)
4
    r2 = r2_score(actual, pred)
5
    return rmse, mae, r2

5. Train the Model and Track with MLflow#

1
warnings.filterwarnings("ignore")
2
np.random.seed(40)
3

4
# Load wine-quality dataset
5
csv_url = ("http://archive.ics.uci.edu/ml/machine-learning-databases/wine-quality/winequality-red.csv")
6

7
try:
8
    data = pd.read_csv(csv_url, sep=";")
9
except Exception as e:
10
    logger.exception("Unable to download training & test CSV. Error: %s", e)
11

12
# Train-test split
13
train, test = train_test_split(data)
14
train_x = train.drop(["quality"], axis=1)
15
test_x = test.drop(["quality"], axis=1)
16
train_y = train[["quality"]]
17
test_y = test[["quality"]]
18

19
# Hyperparameter tuning
20
alphas = [0.2, 0.3, 0.4, 0.5, 0.6]
21
l1_ratios = [0.2, 0.3, 0.4, 0.5, 0.6]
22

23
for alpha in alphas:
24
  for l1_ratio in l1_ratios:
25
    with mlflow.start_run():
26
        lr = ElasticNet(alpha=alpha, l1_ratio=l1_ratio, random_state=42)
27
        lr.fit(train_x, train_y)
28

29
        predicted_qualities = lr.predict(test_x)
30
        (rmse, mae, r2) = eval_metrics(test_y, predicted_qualities)
31

32
        print(f"Elasticnet model (alpha={alpha}, l1_ratio={l1_ratio}):")
33
        print(f"  RMSE: {rmse}")
34
        print(f"  MAE: {mae}")
35
        print(f"  R2: {r2}")
36

37
        # Log params and metrics
38
        mlflow.log_param("alpha", alpha)
39
        mlflow.log_param("l1_ratio", l1_ratio)
40
        mlflow.log_metric("rmse", rmse)
41
        mlflow.log_metric("r2", r2)
42
        mlflow.log_metric("mae", mae)
43

44
        tracking_url_type_store = urlparse(mlflow.get_tracking_uri()).scheme
45
        if tracking_url_type_store != "file":
46
            mlflow.sklearn.log_model(lr, "model", registered_model_name="ElasticnetWineModel")
47
        else:
48
            mlflow.sklearn.log_model(lr, "model")

This loop trains models for various hyperparameters and logs each run in MLflow.

6. Monitor Your Experiments#

While training, you can use the Ngrok URL to monitor and compare experiment runs via the MLflow web UI.

MLFlow Web UI

7. Locate the Artifacts#

After training, your results will be stored under the mlruns directory inside your Google Drive folder.

You can check the experiment ID in the UI (e.g., 33e9f65687374b0b89c87624a8bef70f) and use it to find logs, parameters, metrics, and saved models — all ready for reuse or deployment.

That’s All!#

This is a simple but functional end-to-end MLOps practice using nothing but free tools — perfect for beginners or anyone needing a lightweight workflow without cloud overhead.

You don’t always need enterprise tools to do serious work. Just some free services, smart use of notebooks, and curiosity to explore.

References#

Epilogue#

Starting out in MLOps doesn’t have to be expensive or complicated. As this tutorial shows, you can build a functional, end-to-end pipeline using free and accessible tools like Google Colab, Drive, MLflow, and Ngrok.

If costly cloud platforms or complex setups feel overwhelming, remember: you can always find your own way as long as you have craftmanship mindset. Begin with what you have, keep learning, and gradually scale as your needs grow.

The path to mastery often starts with humble, practical steps, and every expert was once a beginner.

Keep exploring and keep building!