Hangar talk is full of confident half-truths. “You never need to check the oil on a short hop.” “Newer avionics fix everything.” “If it flew fine yesterday it will fly fine today.” Every flying community has them, and the technology world has its own set. When people discuss how to run the systems that watch software, a handful of myths turn up again and again. The Ops Atlas, which uses the name ObservabilityOps for that work, takes aim at several of them. Here are six, busted in the friendly spirit of a Saturday fly-in conversation.
Myth 1: “ObservabilityOps is a product you can buy”
It sounds like one. Vendors love a neat label. But according to the Atlas, the term is not a standard, a certification or a vendor programme; its own searches in October 2026 found nothing of the kind. The reading it gives is straightforward: the work of keeping the observability machinery running, in contrast to leaning on observability while you hunt a fault. There is a commercial tool with a similar name, and the entry specifically warns against confusing the two.
The reality: it is a job description, not a SKU. The practice does exist under other titles, for example an in-house team at a code-hosting firm and a platform group inside a British ministry. When you evaluate anything with this word on the box, ask which parts of the work it actually covers.
Myth 2: “Logs, metrics and traces are three pillars, so we need three tools”
This one sounds like arithmetic, and it has launched many a budget request. The CNCF observability group’s 2023 whitepaper deliberately moved away from “pillars” toward “primary signals”, leaving space for newer signals such as profiles and making it harder to argue that each type demands its own separate stack. Some organisations do choose separate back ends; the point is that the slogan does not force the decision.
The reality: choose back ends for the questions you need to answer, the people you have and the money you can spend. It is like choosing an aircraft: a light sport machine and a turboprop are both airplanes, and the question is the mission.
Myth 3: “Collect everything now and decide later”
It feels prudent, like carrying every spare part in the baggage compartment. The trouble is that storage and indexing are priced by volume, and some forms of volume grow multiplicatively. Every unique combination of label values becomes another series, so a single label tied to something like customer numbers can turn one metric into tens of thousands. The Atlas works through a case where one such label produced 1.8 million series and lifted one team’s overall count to 1,946,880 when its allowance was only 200,000.
The reality: deciding later usually means paying now. Better to set caps at the gateway, keep unbounded values out of metric labels and tier retention so that recent data stays quick to reach and older data moves to cheaper storage. If you enjoy the money angle, our post about taking charge of where your spending goes each month applies the same instinct to household budgets.
Myth 4: “Self-hosting is free”
The licence may cost nothing. The people, the disks, the upgrades and the nights on call do not. “Self-hosting is free” appears on the Atlas’s list of named anti-patterns, and it links the cost to team shape. Once several teams rely on stateful stores you run yourself, a platform team tends to be warranted; if the back end is bought, a slim standards group can be enough. A very small company might get by with one named person on call, mostly where regulation or self-hosting applies.
The reality: price the whole package. A hosted service has a bill; a self-run one has a bill hidden in salaries. Neither is automatically cheaper, and the entry makes no claim that one is. If you are building your own skills before deciding, this guide to developing tech skills through a training school may be a handy starting point.
Myth 5: “Once the SDK is installed, tracing just works”
Installing the library is the easy half. The harder half is making sure each hop passes the identifier to the next. The Atlas lists where it fails: message queues, where explicit span links are needed, proxies that drop the standard header, and systems that mix two propagation formats. In the default setup, OpenTelemetry uses the W3C trace context plus a companion mechanism for small bits of extra data.
The reality: tracing is a property of the whole path, and it needs testing end to end. Also keep the safety angle in mind. The extra data can travel to outside services, so allow only named keys and strip them at egress. It is the equivalent of making sure your logbook does not leave the airfield in someone else’s bag.
Myth 6: “AI will find the problem for us”
Pitches that begin with “AI” tend to skip the tidying step. The sheet’s blunt line is that claims about AI observability and observability pipelines remain marketing until proven on your own data. Machine-learning correlation of events is also a separate topic, outside the scope the Atlas draws for this work. Correlation tools need consistent labels, sensible limits and an owner, which are precisely the basics that the discipline supplies.
The reality: garbage in, confident garbage out. Run a trial on your own records, set the success criteria beforehand and compare against what a careful human with clean data can do. That includes the unglamorous work of reviewing alerts. In the Atlas’s example, only 15 of 42 pages led to action, or 35.7 percent, and no clever layer fixes a pager that cries wolf.
Bonus: “Sampling is only for giants”
Not quite, but the opposite error is common too. The OpenTelemetry guidance that the Atlas leans on suggests thinking about sampling once you approach roughly a thousand traces a second. Below that, keeping everything is often simpler and cheaper than building the machinery to choose. Above it, the choice between deciding at the start of a trace and deciding at the end matters: the first is inexpensive but cannot promise to keep errors, the second keeps them but must send every span of a trace to one Collector, which the guidance calls hard to run.
The reality: start with a measurement of your own trace rate, then decide. A club with six aircraft does not need an airline’s operations centre, and it does not need to ignore scheduling either.
Another bonus: “Monitoring is the platform team’s problem alone”
The Atlas cites the DORA research view that monitoring should not sit with a single person or team, and it counts a central team owning everyone’s alerts as an anti-pattern. The platform owner supplies standards, templates and a healthy pipeline. Service teams still decide which signals matter for their own services. Think of the airport: the operator keeps the runway lit, but each pilot still reads the weather.
Notice what the six main myths have in common: each swaps a decision for a slogan. “Pillars”, “everything”, “free” and “just works” are all ways of avoiding a conversation about who does what and at what cost. The remedy is rarely a tool. It is a short meeting with a whiteboard, an agreed owner and a number or two, such as series count, unit cost or the proportion of pages that mattered. Pilots call it a briefing, and nobody ever regretted having one.
What to do instead
- Name an owner and write down the inventory.
- Add a heartbeat alert, plus one probe from outside.
- Fix your three highest-variety labels.
- Limit what a release can add to your series count.
- Review last quarter’s pages and delete the dead ones.
If you want to read the original and judge it for yourself, the Ops Atlas page covering ObservabilityOps and its anti-patterns sets all of this out, along with a five-level maturity scale of the Atlas’s own devising. Read it as a reference, not gospel; the entry itself makes that clear.
A parting thought from the ramp
Myths last because each carries a grain of truth wrapped in a lot of convenience. “More data is safer” is true until the invoice arrives. “The tool will do it” is true when someone has done the groundwork. Next time one of these statements comes up in a meeting, do what a good flight instructor does with a hangar legend: smile, ask for the evidence and offer a checklist. Our roadmap for corporate IT infrastructure training is one place a team can go to build that evidence habit together.

