Ask any machine learning team how many models they have in storage and you will usually get a shrug. Ask which one is serving traffic and you may get an argument. The cause is rarely technical failure; it is naming. When a trained model, a dataset snapshot or an alert rule is called “test_final_v3”, nobody can tell what it does, and every handover becomes archaeology. This piece compares three common ways of naming AI and AIOps artifacts, shows where each one breaks, and ends with a blunt recommendation.
The problem: names are the interface people actually use
Tools store metadata, but humans read names. In a model registry, a pipeline run list or an alert dashboard, the name is the first and often only clue about what an item is. A bad name forces everyone to open the item, read its config, and guess. A good name answers the three questions that come up every time: what is it, what is it for, and which one is current?
The cost of unclear names is paid in small, repeated amounts. A retraining job gets pointed at the wrong dataset. An on-call engineer silences the wrong alert because two rules sound alike. A new colleague spends a week learning which of five folders is real. None of these is dramatic, which is exactly why teams tolerate them for so long.
Approach one: free-form names
This is the default. Someone saves a model as “churn_model_new” or a notebook as “experiment_try2”, and the name reflects whatever was in their head that afternoon.
- Strengths: zero friction, no rules to remember, works fine for a one-person prototype that lives for a week.
- Weaknesses: meaning lives only in the author’s memory. Words like “new”, “latest”, “final” and “fixed” lose meaning as soon as a newer item appears. Searching returns noise, and two people choose different words for the same idea.
- Where it breaks: the day the author goes on leave, or the day a second team wants to reuse the work.
Free-form naming is acceptable only for scratch work with an expiry date. If an item will outlive the sprint, it deserves a better name.
Approach two: descriptive plain-language names
Here the rule is simple: say what the thing is in ordinary words. “Customer churn classifier, retail accounts” or “Disk-full alert, payment database”. No codes, no cleverness.
- Strengths: instantly readable by engineers, analysts and managers alike. Easy to adopt because it needs no training, just a habit. Excellent for dashboards, alert titles and anything a human scans.
- Weaknesses: descriptive names drift. People add a word here, drop a word there, and soon there are three spellings of the same service. They also tend to be long, and they do not sort well, so lists become hard to scan.
- Where it breaks: at scale, when dozens of similar items need to be told apart or sorted by date and version.
Approach three: structured, templated names
The third option is a fixed pattern with slots, for example: task, data scope, algorithm family, version. A model might be “churn-retail-gradient-boosting-v7” and an alert “db-payments-disk-usage-warning”.
- Strengths: sortable, filterable and scriptable. Automation can parse the parts, and a name that breaks the pattern is easy to spot. Great for registries, pipelines and infrastructure.
- Weaknesses: rigid patterns invite abbreviation. If the data-scope slot is two letters, a code appears, and the code needs a lookup table. Patterns can also become bureaucratic if they carry too many slots.
- Where it breaks: when the template is too clever for newcomers, so they invent their own and the pattern erodes.
Side by side
A fair comparison depends on who reads the name. Here is how the three approaches fare for the people who actually touch AI and AIOps artifacts.
- Readability for a new colleague: descriptive wins, structured is close behind, free-form loses.
- Sorting and search: structured wins by a wide margin.
- Effort to adopt: free-form is free, descriptive costs a habit, structured costs a short template and a bit of enforcement.
- Resistance to drift: structured holds up best, provided the slots stay few.
- Good for alerts shown to humans at three in the morning: descriptive, with the service and symptom spelled out.
- Good for automation and registries: structured.
What to name, and how, in an AIOps setup
Different artifacts reward different treatment, so a practical team mixes approaches rather than choosing one forever.
- Models: use the structured form. Purpose, data scope, family and version, in that order. Keep “production” and “staging” as tags or stages in the registry, not part of the model name, because the same model moves between them.
- Datasets and snapshots: say what the data is and the day it was cut. A snapshot called “orders-clean” is a trap; “orders-cleaned-2026-09-30” can be compared against its neighbours.
- Experiments: one sentence in plain words describing what changed. “Added seasonality features” will still make sense six months from now; “run_44” will not.
- Alert rules: descriptive, with the service, the symptom and the severity. Someone half asleep should know what is wrong from the title alone.
- Feature flags and config variables: name the behaviour switched on, not the ticket that introduced it. A ticket number tells you when, not what.
This mix keeps human-facing names plain while letting machine-facing names stay tidy. It also keeps the number of rules small, which is the main factor in whether people follow them.
Evidence from daily work
You can test a naming scheme without a study. Take ten artifacts at random and hand the list to a teammate who did not create them. Ask them to say what each one is and whether it is current. Count the ones they get right without opening anything. The score will tell you more than any style guide.
Readers who want to see how AI supports everyday decisions in a business can look at this piece on how AI helps businesses decide. Every one of those decisions leans on data and models that somebody has to find, so naming discipline sits underneath all of it. Likewise, anyone curious about the mechanics of classification can read an introduction to how image classifiers see and then ask: if I had to hand this model to someone, what would I call it?
Common traps
- Putting the environment in the name. Promotion changes the environment; the name should not need to change with it.
- Naming by author. People move on, and a model named after its creator is orphaned the moment they do.
- Cute project names. A codename is fun until a new engineer has to decode it. If you must use one, pair it with a plain description.
- Renaming without redirects. If dashboards or scripts point at the old name, renaming silently breaks them. Rename, then update references, then retire the old name.
- Skipping the dates. Anything that is a snapshot needs a date in year-month-day form, otherwise sorting scrambles.
A break from the repository
Names matter outside engineering too, which is a handy reminder when you try to sell a scheme to non-technical colleagues. When a team plans an offsite or a weekend out, a Gujarat-focused site such as GujaratOrbit, which pairs event listings and venue pages with a forum and a blog, shows how clear naming decides whether a plan gets shared or ignored. A message that says “the second option” is useless; “the evening comedy show near the hotel” is not.
The recommendation
Do not pick one scheme and force it everywhere. Use plain descriptive names wherever a human decides something quickly, as with alerts, experiment notes and dashboards. Use short structured names wherever a machine or a sorted list does the work, as with registries, pipelines and snapshots. Keep the template to four slots or fewer and write it down on a single page that lives next to the code.
Then make compliance cheap. Add a check in the pipeline that rejects names which break the pattern, and ask reviewers to treat a vague name the same way they treat a vague commit message. In a few weeks the habit becomes invisible, and the payoff arrives in the form of quieter handovers, fewer wrong-dataset mistakes and alerts that explain themselves.