Why does reproducibility matter in Data Science?

Imagine you ask a friend for a suggestion for buying a car. At the showroom while buying your friend calls up and says, “oh no, I have changed my recommendation to some other car”. Even worst, the friend tells you after you make the purchase.

You would like your friend to have shared opinions that are fixed, reliable and same every time.

Even to compare if model A did better than model B, the results of model should not change after computing model B metrics.

How would you ensure model reproducibility?