MLOps in Plain Language: A Bakery, an Apprentice and a Model

Rajesh Kumar

Rajesh Kumar is a leading expert in DevOps, SRE, DevSecOps, and MLOps, providing comprehensive services through his platform, www.rajeshkumar.xyz. With a proven track record in consulting, training, freelancing, and enterprise support, he empowers organizations to adopt modern operational practices and achieve scalable, secure, and efficient IT infrastructures. Rajesh is renowned for his ability to deliver tailored solutions and hands-on expertise across these critical domains.

Imagine a neighbourhood bakery that learns, over years, which loaves sell on which mornings. One day the owner hires a very fast apprentice who studies the old sales sheets and starts deciding how many loaves to bake. The apprentice is a machine-learning model. The shop, its ovens, its delivery van and its morning rush are the real world the model has to live in. MLOps is simply the set of habits that keep that apprentice useful once the shop doors are open. No special background needed; if you can picture a bakery, you can follow this article.

The ideas below are taken from the plain-reference page on MLOps in the Ops Atlas, and I have translated them into everyday language. For readers who like to look up unfamiliar words as they go, the glossary on this site, a beginner-friendly word list covering fifty-odd DevOps terms, is a good companion.

What MLOps means, in a single breath

Put simply, it covers getting a trained model out of the lab and into daily use, and then keeping its answers correct as time passes. The Atlas lists what that involves: recording exactly which data, code, settings and software produced the model, automating the way it is trained and released, watching for signs that it has gone off, and retraining in a controlled way. Notice that it says nothing about being clever. It is about being dependable.

Why a model is more like bread than like a building

A building, once finished, stays finished. Bread goes stale. A model is closer to bread, because its quality depends on the ingredients it was made from, and ingredients change. The Atlas puts it this way: the model is something you build, and what you build it from includes data. If you cannot say which flour went into Tuesday’s loaf, you cannot explain why it tasted different from Monday’s.

This is why the first habit is keeping a recipe card for every training run. The card lists the code version, a fingerprint of the data, the settings and the environment. A friend should be able to follow the card and produce a loaf that tastes about the same. The word “about” matters. Even good bakers do not get identical loaves, and training has some randomness too, so the practice asks for results within a reasonable margin rather than perfect copies.

Practising the way you will perform

Suppose the apprentice trained on neat sales sheets typed up by the owner, but in the shop the numbers arrive scribbled on receipts with different date formats. The apprentice would make odd choices, not because it is foolish but because the sheets it studied and the receipts it sees do not match. The Atlas calls this skew, and the fix is boringly practical: prepare information the same way in both places, keep a record of what the model actually saw, and test that both routes give the same answer for the same input. When the two agree, the mismatch count is zero.

Trying a new apprentice without risking the shop

Say a newer apprentice looks better on paper. Do you hand over the whole bakery on day one? Of course not. The Atlas describes a careful sequence:

  1. Same exam for both. Give old and new the same fixed set of past cases, and compare not just the overall score but how each does on special days, such as holidays or the first week of a new shop opening.
  2. Shadow. Let the new apprentice make guesses in the back room while only the old one’s orders are used.
  3. A small slice. Let the newcomer handle a small share of real orders, a canary in the coal mine.
  4. A real experiment. Measure whether the shop actually earns more, not just whether the guesses look neater. A travel-booking company reported that better model scores did not guarantee a business benefit.

Watching the shop while it runs

After launch, the Atlas suggests watching on three levels. First, is the shop open and quick? Second, are customers and orders looking like the ones the apprentice studied? Third, are its decisions actually good? The third is awkward, because the true result of a decision often arrives late. A baker learns instantly how many loaves were left over, but a company predicting who might cancel a subscription may wait a month to find out. In the meantime you watch stand-ins, such as whether the apprentice is suddenly ordering far more loaves than usual.

Three kinds of “something changed”

When the numbers look strange, the temptation is to say “retrain”. The reference suggests asking which of three things happened:

  • The sheets and the receipts disagree (the skew above). Fix the paperwork.
  • A supplier started sending broken data, such as empty fields. Fix it at the source.
  • The world truly changed. Maybe a new office opened next door and the morning crowd is different. Only then does retraining on fresh examples make sense, and only if you can find out what really sold.

It is worth saying again: a single “alarm chart” cannot tell these apart. A person has to look.

Going back when a change goes wrong

Sometimes a new apprentice causes trouble anyway. The calm response is to put the old one back at the counter by flipping a switch, not by training a replacement under pressure. A sensible bakery keeps a simple fallback too, such as “bake last week’s amounts”, in case no apprentice is trustworthy that day. The Atlas recommends practising this switch on a schedule, roughly four times a year, so that nobody is learning it during an emergency.

A day in the life, bakery edition

To make the habits concrete, here is a made-up morning. At five o’clock the shop’s system compares today’s early orders with the usual pattern and finds nothing unusual. At six, the apprentice’s proposed bake list goes out, and a small slice of it comes from the newcomer rather than the veteran. At seven the owner glances at one simple page: orders arriving, guesses made, leftovers so far. At nine, a delivery sheet comes in with the date written in a different order, and the check that compares sheets with receipts raises a quiet flag. Nobody panics. A person looks, spots the format change at the source, and corrects it before the afternoon batch. At closing time the day’s record is saved, so tomorrow’s review starts from facts rather than memories.

Nothing in that morning needed a grand platform. It needed a record, a comparison, a switch and a person who owns the page. That is the quiet truth behind most successful MLOps stories.

Where MLOps ends

It helps to know what is not included. MLOps is not the same as inventing the apprentice, which is data science. It is not the work of keeping suppliers’ deliveries tidy, which belongs to data operations. It is not the rulebook for risk across all your apprentices, a separate discipline called ModelOps. And nobody sells it in a box: the Atlas says a platform without ownership or gates is just the beginner stage with nicer screens. If you are curious how the field’s story unfolded in Japan and beyond, our overview of the road ahead for DevOps practice at home and abroad sets the wider scene.

Words you will hear, in bakery terms

  • Registry: the shelf where each finished apprentice is kept, with a label showing where it came from.
  • Alias: the sticky note that says which apprentice is on the counter right now. Moving the note is how you switch.
  • Lineage: the trail from a loaf back to its flour, so you can answer “what went into this?”
  • Slice: one group of customers or days that you check on its own, because an average can hide a bad corner.
  • Retire: to take an apprentice off duty deliberately, after telling everyone who relied on it.

A small starter plan

You do not need to do everything at once. The reference’s own suggestion is modest: one repeatable pipeline, a place where finished models are stored with a clear history, and a way back. Do those three things for a single model and you will have passed a good number of teams who skipped them. Later, add watching, a written rule for when to retrain, and an honest plan for retiring a model when it is no longer needed. Like a bakery, it is mostly small, steady habits done every day.

Leave a Reply