Upgrade & Secure Your Future with DevOps, SRE, DevSecOps, MLOps!
We spend hours scrolling social media and waste money on things we forget, but won’t spend 30 minutes a day earning certifications that can change our lives.
Master in DevOps, SRE, DevSecOps & MLOps by DevOps School!
Learn from Guru Rajesh Kumar and double your salary in just one year.
Every AI operations team I have sat with has a story that starts the same way: somebody pages the on-call engineer about “model_v2_new”, and nobody in the room can say which model that is, where it runs, or who last touched it. The fix that follows is never clever. It is a better name. This list collects eight of those moments, told as they tend to happen, together with the plain naming habit that ended each one.
If you work with machine learning in production, you already know the code itself seldom causes the late nights; the clutter around it does, meaning models, datasets, feature tables, experiment runs, alerts and dashboards. Every one of those needs a label, and the first label is usually chosen in a hurry. Naming AI operations assets clearly is the cheapest reliability work available, because it costs a few minutes once and saves an explanation every time afterwards.
1. The model called “final_final”
The first story is the classic. A forecasting model had been retrained three times, and the registry held entries named “demand_model”, “demand_model_final” and “demand_model_final_final”. When a drift alert fired on a quiet Sunday, the person on call spent the better part of an hour working out which one was actually serving traffic.
The replacement convention was short: purpose, data window, training date. Something like “weekly-demand-forecast_data-to-month-end_trained-on-the-tenth”. It reads like a sentence and it sorts properly. Nobody had to ask which one was live, because the name itself said what it predicted and how fresh it was.
2. A feature called “score”
Features are where vague names breed fastest. One shared table carried a column named “score”, and three teams used it to mean three different things: a fraud likelihood, a customer engagement value and a data quality rating. Two of those teams discovered the clash only when a model quietly started training on the wrong one.
The remedy was to put the subject and the unit into the column name. “engagement_score_0_to_100” is longer, but it cannot be mistaken for anything else. A good feature name answers three questions at once: what is measured, about what, and on what scale.
3. The experiment run named after a mood
Experiment trackers happily accept whatever label you type. Our storyteller’s team had runs called “try-again”, “better-maybe” and “Friday”. Six weeks later, nobody could reproduce the run that had produced the best validation result, because the label said nothing about the settings behind it.
The new habit was to build the run name from the one change that made it different: “learning-rate-halved”, “added-weekend-flag”, “smaller-window”. When two runs sit side by side, the difference is visible in the list before anyone opens a single chart.
4. The alert that said “anomaly detected”
An alert title is the first and often the only thing a tired person reads. “Anomaly detected” tells you nothing about where to look. After one particularly slow response, the team rewrote their alert names so each one carried the service, the symptom and the threshold crossed: “checkout-ranker latency above agreed limit for ten minutes”.
This is not the same job as trimming how many alerts fire. If you want that side of the problem, this site also has a piece about calming a flood of IT events through smarter tech. Naming only decides whether the alerts that do fire can be understood on first read.
5. The dataset folder nobody could find
Training data lived in a bucket with folders called “data”, “data2”, “new_data” and “latest”. The folder called “latest” had not been the latest for months. A new joiner trusted it anyway, and a retrain used stale rows.
The team moved to names that pair the source with the extraction date and never reuse a word like “latest”. A folder called “support-tickets_extracted-early-spring” ages honestly. Anyone can see how old it is without opening it.
6. Branches that described a person, not a purpose
Repository branches such as “work-in-progress” and “my-fixes” hide intent. During a review, the question “what is this branch for?” took longer to answer than the review itself. The convention that stuck was a short verb phrase tied to the change: “add-drift-check-to-batch-scorer”, “rename-feature-columns”.
A reader scanning the branch list now learns the plan from the names alone. Reviews begin with context rather than archaeology.
7. Environment variables with secret meanings
Configuration is the quiet corner where bad names hurt most. A variable called “THRESHOLD” appeared in four services, and each meant a different cut-off. When one of them was changed during a tuning session, the wrong service moved.
Prefixing the variable with the service and describing the behaviour fixed it: “RANKER_MIN_CONFIDENCE_TO_SERVE” cannot be confused with the threshold of an unrelated batch job. It is also easier to search for in a repository, which matters more than people expect.
8. Dashboards named after their authors’ hopes
The last story is about a dashboard titled “Model health”. Which model? Health by what measure? The panel mixed accuracy, latency and data freshness, and viewers kept arguing about whether the line going down was good news.
Splitting it into dashboards named for one question each, such as “serving latency for the ranking service” and “input freshness for the nightly features”, ended the arguments. A dashboard name is a promise about what the viewer will learn, and it should be one the page can keep.
A small naming test you can run today
None of these eight fixes needed new software. Each was a decision, written down once, that a person could follow without asking anyone. That is the quiet strength of naming: it moves knowledge out of one head and into the thing itself, where a stranger on a bad night can read it.
Pick any asset in your pipeline and imagine handing its name, and only its name, to a colleague who joined last week. Could they say what it is, where it lives and roughly how old it is? If not, rename it before the next incident does the teaching.
Outside the engineering world, the same instinct shows up wherever people have to find things quickly. Take GujaratOrbit, a site built around event listings for Gujarat, with venues, a forum and a blog each given a plainly worded section. Visitors never need a guide to guess where to click, and that kind of courtesy scales to any team.
Keeping the habit alive
Conventions decay unless someone writes them down. Keep a one-page naming guide next to your pipeline code, with two or three example names for each asset type, and make it part of review. When a pull request introduces a name that fails the new-joiner test, say so kindly and suggest a better one.
For a softer introduction, a newcomer-friendly article about diagnosing live technical trouble as it happens shows how much quicker diagnosis gets when the evidence carries clear labels. Start with the names, and the rest of the workflow gets a little easier to reason about. Teams that try this often report, informally, that stand-ups get shorter because fewer sentences begin with the words “which one do you mean”.
Finally, resist the urge to rename everything at once. Pick the three assets that caused the most confusion this quarter, rename them with a short note in the changelog, and let the convention spread through new work. A few clear names, applied consistently, are worth more than a perfect scheme nobody follows.