A research lab shares an expensive instrument, and three colleagues each propose a different way of deciding who may book it. Maya wants a list of job titles. Jonas wants rules that check facts such as training status and the hour. Priya wants the system to follow relationships: whoever supervises a project may book on its behalf. All three are right in a particular setting, and the argument between them is the heart of this comparison. It sits inside IAMOps, an informal label for treating authorisation as policy you write, test and refine, as set out in the Ops Atlas.
The Atlas explains the three models and their standards in its entry on IAMOps, while noting that the word is not an established discipline. Quantum and emerging-tech teams, with their pricey instruments and mixed staff of academics, vendors and automated schedulers, are an excellent place to see the contrast.
Maya’s way: role-based access
Role-based access control groups permissions into roles such as technician, researcher and administrator, then assigns people to roles. It has a formal pedigree: the Atlas cites an American national standard from 2004. The benefits are plain. Auditors understand it, onboarding means adding a person to a role, and the list of roles can be printed on one page.
The weakness appears as the lab grows. A role called “researcher” soon needs variants: researcher on project A, researcher after training, researcher on weekends. Roles multiply until nobody can describe them all, a pattern engineers call role explosion. Maya’s list worked beautifully for twelve people and began to creak at sixty.
Jonas’s way: attribute-based access
Attribute-based control decides from facts about the principal, the resource and the environment. A rule might say a booking is allowed if the person’s safety training is current, the instrument is not in maintenance and the time falls in working hours. The Atlas points to a 2014 US government guide as the standard reference.
This model is expressive. It handles context naturally and cuts the number of roles. The price is that rules turn harder to audit, because the answer to “who can use the instrument?” is no longer a list but a computation. There is also a subtle trap the Atlas highlights: a rule worded “allow unless the training tag equals expired” gives a surprising answer if the tag is absent altogether, because negated tests can match an absent value. Jonas’s scheme needs careful tests for missing data.
Priya’s way: relationship-based access
Relationship-based systems store a graph of links, such as “Priya supervises project Orion” and “project Orion owns the booking calendar”, and answer questions by walking the graph. The idea became widely known through a 2019 paper describing Google’s authorisation system, and open projects such as OpenFGA follow its lead. The Atlas notes the original reported very low response times for most requests, which matters when every page view triggers a check.
This model fits sharing and hierarchy particularly well. Documents, folders, teams and projects relate to one another naturally. Its weakness is operational: you now maintain a graph, and mistakes in a relationship can quietly grant access to everyone beneath a node.
Choosing for a lab like this one
A few questions can steer the decision. How many distinct job families exist, and do they change often? How much does context matter, such as maintenance windows, certification expiry or the sensitivity of a sample? Do people share work in nested groups, so that ownership flows downward? If the first answer is “few and stable”, start with roles. If the second is “a lot”, add attributes. If the third is “constantly”, look at relationships. Write the answers down before shopping for software, because the shape of the problem should drive the tool and not the other way round.
A side-by-side summary
- Easiest to explain: roles.
- Best for context such as time and device: attributes.
- Best for sharing and nested ownership: relationships.
- Hardest to audit as a simple list: attributes.
- Most operational overhead: relationships, because the graph must be kept accurate.
Rarely does a mature lab choose just one. A common arrangement uses roles for broad job families, attributes for context checks and relationships for project-level sharing. The point of IAMOps is not to pick the winning model; it is to write the combination down as code, test it and watch how it behaves.
What stays true whichever model you choose
Several rules apply across all three. Every decision has the same four inputs: who, what, which object and what circumstances. A grant gives and a ceiling only limits. A written deny wins. And least privilege remains a loop of setting, checking and trimming, with the caveat that rarely used duties hide from short windows of data. The lab’s quarterly calibration run, for example, would vanish from a thirty-day review.
Teams building shared tooling around such systems often find the planning advice in a beginner roadmap to platform engineering a useful way to decide who owns the policy repository and who reviews changes.
Back to the lab
In our story the three colleagues did not declare a winner. They kept Maya’s roles for job families, added Jonas’s training check as a condition and let Priya’s graph handle project calendars. They wrote ten tests: five requests that must succeed, five that must fail. The first month produced two surprises, both caused by missing attributes, and both were fixed in the repository instead of in a hurry on the day of an experiment.
The trade-off thinking will feel familiar to readers who enjoy weighing options in the style of the way quantum algorithms attack difficult optimisation tasks: no single answer is best everywhere, and the right design depends on the constraints you actually face.