Ontology vs Schema vs Topology: One Idea, Many Dialects
Five words drift through architecture meetings — ontology, schema, topology, data policy, traversal policy — and each speaker bends the meaning a little, until nobody is sure they are describing the same thing. One engineer says the knowledge graph needs an ontology. Another says the schema already handles it. Someone mutters that the topology won't support it, and the room splits between people picturing network layouts and one person quietly picturing open sets.
The conversation feels circular because the words genuinely overlap — then pull apart at the exact point most teams never name.
Ontology, schema, topology, data policy, and traversal policy all name one commitment — a stance about what kinds of things exist, how they relate, and how you may move through or touch them. They agree on what you are describing. They differ in the contract that description carries once you hand it to a machine and ask it to do work: inference, enforcement, invariants, or permissions.
They are not interchangeable.
TL;DR
- Ontology, schema, topology, data policy, and traversal policy share one commitment: what kinds of things exist, how they relate, and how you may move through or touch them.
- Each adds a different contract: ontology adds inference, schema adds enforcement, mathematical topology adds invariants, and data or traversal policy adds permissions and procedure.
- The split shows up in one question: what does absence mean here — false, unknown, or forbidden? Teams that argue in circles are usually not disagreeing about the data; each has silently picked a different answer to that question. Name it, and the argument turns into a five-minute design decision.
Ontology vs schema: the roll-call
The philosophical room starts with being. The older tradition asks about "being qua being" (Stanford Encyclopedia of Philosophy); the word ontology itself arrived many centuries later, so it should not be placed in Aristotle's mouth. Quine gave the compact modern test for what a theory says exists: "To be is to be the value of a variable" (Quine, "On What There Is," 1948). A theory is committed to exactly the entities its variables must range over for its statements to be true.
The knowledge-engineering room makes that explicit. Gruber's canonical formulation: "An ontology is an explicit specification of a conceptualization" (Gruber, 1993). The conceptualization is the objects, concepts, and relations presumed to exist in a domain; the ontology is a shared, formal vocabulary — classes, relations, functions — for them. OWL carries the idea into the Semantic Web, supplying classes, properties, individuals, and data values, then defining terms through their relationships to other terms (W3C OWL 2 Overview, 11 December 2012).
The database room prefers schema. Codd's relational model represents data as relations grounded in set theory and first-order predicate logic (Codd, 1970); later database practice turned that structural commitment into named tables, fields, and constraints. JSON Schema states the enforcement role plainly: describe what the data should look like, then let a validator reject anything else (JSON Schema overview).
And psychology had the word long before any database did. Back in 1932 — four decades before Codd — Bartlett described a schema as an active organisation of past experience that shapes a well-adapted response (Bartlett, Remembering, 1932). The same commitment about how things relate, in a field with no databases in sight: the word did not spread outward from computing; every room reached for it on its own.
The overlap is not careless speech. Steven Dam of the Lifecycle Modeling Organization puts it bluntly: "For the systems engineer, these terms seem to be effectively synonyms" (Dam / LMO). And the flagship artifact of the web's structured data proves the point: schema.org calls itself a schema and a vocabulary, yet its types inherit from parent types and its properties connect entities. In common use, it behaves ontology-like.
One commitment, many meeting rooms.
What each dialect actually adds
The useful question is not which noun wins. Ask which contract is in force.
| Dialect | Contract it adds | Operational question |
|---|---|---|
| Ontology | Formal semantics and inference | What else follows from these axioms? |
| Schema | Constraints and enforcement | Does this data pass or fail? |
| Mathematical topology | Invariants | What survives continuous deformation? |
| Data policy | Authority and permissions | Who may read, write, or retain this data? |
| Traversal policy | Search procedure | Which relation comes next, and when should the walk stop? |
Ontology adds formal semantics and inference. An ontology corresponds to a logical theory — a set of axioms — from which a reasoner derives new, entailed facts. A database schema defines structure but derives nothing; inference capability, not vocabulary, is the real dividing line (A Knowledge Engineering Primer, 2023). That capability drags a world-assumption along with it: OWL ontologies default to the open-world assumption, where the absence of a fact means unknown, while a relational schema defaults to the closed-world assumption, where absence means false (OWL world assumptions).
Schema adds enforcement. The same commitment becomes a gate — constraints that a validator (JSON Schema, SHACL shapes, SQL DDL) checks and rejects on. SHACL and OWL can run over the same RDF graph and do opposite jobs: SHACL validates under closed-world (absence → violation), OWL reasons under open-world (absence → unknown, no violation) (SHACL in the presence of ontologies, 2025).
Mathematical topology adds invariants. A topology on a set is a collection of open sets containing the empty set and the whole set, closed under arbitrary unions and finite intersections — the point-set definition carried by texts such as Munkres's Topology (open equivalent). It studies what survives continuous deformation: connected components, holes and voids. Plain adjacency never names these; persistent homology captures higher-order structure that pairwise edges miss (Persistent Homology of Networks).
Data and traversal policy add permissions and procedure. Not "what exists" but "what you may do, and in what order." Data governance, in the DAMA-DMBOK framing, is the exercise of authority and control over the management of data assets (DAMA International, DAMA-DMBOK, 2nd ed., 2017). A traversal policy pushes the same idea onto the graph: from a seed node, which relation to follow, how wide to branch, how deep to go, and when to stop. LLM-guided graph search such as Think-on-Graph formalizes that as a beam search over relations and entities (Think-on-Graph) — proof that traversal is a genuine contract, described in full in navigating knowledge graphs.
Same commitment. Different contracts.
The Alice test: what does absence mean?
Take one record: Alice has no recorded parents.
Under a bare closed-world assumption, the absence simply reads as false: Alice has no parents (OWL world assumptions). Add a SHACL cardinality rule that requires a parent, and the same absence becomes a violation — the shape says every person must have one, and the validator fails the record (SHACL in the presence of ontologies, 2025). Under an OWL reasoner working open-world, the identical fact is a harmless unknown; maybe nobody recorded her parents, and the reasoner concludes nothing about whether they exist.
Same Alice. Same graph. Opposite verdicts.
And a third room never reaches the truth question at all. Under a data policy that marks parentage as restricted personal data, the absence is a permission decision before it is ever a truth value — you may not read it, whether or not it exists. That is forbidden, the third meaning of missing.
That is why "it's literally all the same thing" over-claims. The shared commitment says that parents and people can exist and relate; the contract decides what the system may conclude when a relation is missing. The question every architecture meeting needs to settle is:
What does absence mean here—false, unknown, or forbidden?
Answer that, and the vocabulary arms-race becomes a design choice.
Graph topology: concede the overlap
On topology, everyday usage nearly collapses into structure. "Network topology is the arrangement of the elements (links, nodes, etc.) of a communication network." It is also, in the same source, "an application of graph theory" (network topology). In that colloquial sense, topology just means the graph's arrangement or shape, and calling it another structural dialect is fair.
The mathematical sense is the genuine exception. It asks what stays invariant under continuous deformation and can represent higher-order features beyond pairwise edges (persistent homology of networks). Outside specialist work, almost nobody means that.
A similar truce settles the open-world / closed-world split — a practitioner's rule of thumb rather than a standard (Volodymyr Pavlyshyn): use an open-world stance when modelling a domain that extends beyond any dataset, where absence is merely missing evidence; use a closed-world stance when modelling operations — a pipeline needs closure to be correct, and missing data can be a real failure. Both sides are right; they are signing different contracts for different jobs.
Common questions
Is schema.org an ontology?
schema.org calls itself a schema and vocabulary, but its type inheritance and entity-linking properties make it behave ontology-like in common use. Source: https://schema.org/docs/datamodel.html
What is the difference between an ontology and a schema?
An ontology supplies formal semantics from which a reasoner can derive entailed facts. A schema supplies constraints that a validator can enforce. Source: https://arxiv.org/pdf/2305.17196
What is the difference between graph topology and topology in mathematics?
Graph topology usually means the arrangement of nodes and links. Mathematical topology studies structures defined through open sets and properties preserved under continuous deformation. Sources: https://en.wikipedia.org/wiki/Network_topology and https://www.math.ucla.edu/~sharifi/topology.pdf
What does open-world vs closed-world mean?
Under an open-world assumption, an absent fact remains unknown. Under a closed-world assumption, absence can count as false or trigger a validation failure. Sources: https://en.wikipedia.org/wiki/Web_Ontology_Language and https://arxiv.org/abs/2507.12286
Related
- Knowledge graph memory for agents
- Navigating knowledge graphs
- Hypergraph vs hyperbolic graph
- AI agent memory
Edward Izgorodin · Mnemoverse · last updated 2026-07-11