At the hangar coffee table, the good stories are usually corrections. Somebody swears a certain fix always works, an old hand quietly explains why it does not, and the whole table learns something. MLOps, the craft of running machine learning models in daily service, has its own set of hangar-talk beliefs, and several of them sound as sensible as “just top off the tank, it will be fine”. This article walks through six of them, with an Ops Atlas reference page playing the old hand at the table, and translates each into the world of light-sport flying.
Myth 1: “Getting the model online is the whole job”
The story goes that once a model replies to requests over a network, the work is done. It is the equivalent of believing that a finished airframe is a finished airplane. The Ops Atlas page counters by noting that, in Google’s own breakdown of the lifecycle, deployment is a single process out of seven. The other six cover things like data preparation, evaluation, monitoring and retraining.
Anyone who has owned an aircraft knows the delivery flight is day one of a long relationship. Inspections, logbook entries, small repairs and the occasional surprise fill the years that follow. A model’s life in production is the same: the interface is just where it meets the public.
Myth 2: “A serious team automates everything”
This one flatters ambition. If automation is good, more automation must be better, and the finish line is a model that retrains and promotes itself overnight. The Atlas lists exactly that as an anti-pattern. A loop of that kind can start to learn from the consequences of its own outputs and cause outages, and nobody is watching when it happens.
The honest picture, according to the entry, is that the level of automation is a choice made per model. Google’s maturity ladder has three rungs and Microsoft’s has five, but neither says you must stand on the top one. Google’s own opening rule is blunter still: skip machine learning altogether when the problem does not call for it. A sport pilot would nod. You do not climb to the flight levels to cross the field.
The autopilot test
A neat way to hold the idea: autopilots are splendid, and pilots still decide when to engage them and when to take back control. Keep a human at the promotion step until your checks have earned the right to be trusted.
Myth 3: “Buy the right platform and the problem is solved”
Every airfield has a hangar with an expensive tool nobody has the training to use. The Atlas says you cannot purchase MLOps off a shelf, and it describes a platform lacking ownership and release gates as the starting maturity level with a nicer interface. The shiny dashboard does not decide who is responsible when a model misbehaves on a Friday evening.
The entry’s starter advice is surprisingly modest: a pipeline you can rerun, a registry that remembers the origin of each version and a path back beat an impressive platform that sits unused. If you are weighing tools anyway, our guide on building a sensible set of software tools offers a calm way to compare options before you commit.
Myth 4: “It is all the same family of Ops”
It is easy to hear a string of similar names and assume they are synonyms. The Atlas draws careful lines. Work on prompts, retrieval and safety rails for language models goes to LLMOps, though pipelines that fine-tune a model remain MLOps. Governing every model a company owns, through risk grades and independent review, is ModelOps, and MLOps supplies the registry it relies on. Applying machine learning to IT alerts is AIOps, which is the reverse direction, since MLOps is operations on behalf of machine learning. Upstream ingestion and data contracts are DataOps. And none of these is data science, which designs the model to begin with.
Light-sport pilots know the equivalent confusion between the airframe mechanic, the avionics shop and the examiner. They all touch the same aircraft, and each has a distinct job. A clear boundary prevents the common failure in which everyone assumes someone else is checking.
Myth 5: “Our drift chart will show us the culprit”
A single chart that glows red when something changes is comforting, and it is not enough. The entry distinguishes three situations that look identical on a dashboard. Two pipelines may disagree about how an input is calculated, in which case the pipeline must be fixed. Upstream data may be defective, in which case the source needs repair. Or the world may truly have moved, which may justify retraining, provided labels exist.
The sceptical reader should also know that popular alert levels are not laws of nature. The Atlas observes that the usual cut-offs for one famous index lack academic support and that the number changes with binning and sample size. In its executed lab, the reading for an empty reference bin nearly tripled merely because a tiny smoothing constant was changed, a lot of wobble for something treated as a verdict. A mechanic would say never trust a single gauge that you cannot cross-check.
A related caution is in the same entry: a default in one open-source tool flags dataset drift only when more than half the columns drift, so a single critical column can go bad unnoticed. Think of an engine monitor that stays quiet unless several cylinders overheat at once. For a plain-spoken parallel in hardware terms, see our checklist for looking over landing gear on a light-sport aircraft, where the point of a checklist is that each item gets looked at individually.
Myth 6: “If it goes wrong, just retrain it”
Retraining feels like a universal repair. The entry insists it is a treatment, not a diagnosis, and it is definitely not a rollback. Going back to a known-good model means repointing an alias or moving a traffic split, which takes minutes. A fresh training run takes longer, may produce something different and might repeat the same mistake if the cause was a pipeline disagreement.
The Atlas recommends keeping a plain fallback that does not use machine learning at all and rehearsing the switch every quarter. Pilots call that an emergency procedure. You do not discover it in the air. A related read from our shelf, fuel system mistakes pilots can avoid, makes the same argument about preparing for problems before they appear.
A hangar story that ties it together
Picture a small flying club whose booking tool suggests ideal time slots to members. One spring the suggestions start to feel wrong: mornings that are always crowded keep being recommended. A member who believes myth five stares at the drift chart, sees red and asks for a retrain. A member who has read the Atlas asks instead whether the live tool calculates “daylight remaining” the same way the training job did. It does not; one uses local time and the other uses universal time. The repair is a one-line change shared by both paths. No retraining, no new platform, no automation project. The previous version of the tool was still registered, so reverting while they investigated took a single switch.
That small win is the real point. Most production trouble is dull, local and fixable by someone who stops to ask the right question in the right order.
The story underneath the six myths
Notice what the myths share. Each one promises that a single move solves a whole class of trouble: a deployment, an automation switch, a purchase, a name, a dashboard or a retrain. The Atlas answers each with a small habit instead: record what produced a model, gate what replaces it, release in stages, watch three layers, choose a retraining rhythm with evidence, keep a way back.
If you want one thing to take to the hangar coffee table, take this. Ask any team that talks about models in production a simple question: if it went wrong right now, how would you go back, and when did you last practise? Whatever comes back will reveal more than a polished slide ever could.

