hugoduncan Ideal Namespace Structure — Principles
A tool-agnostic statement of what good Clojure namespace organisation looks like. Use to evaluate, design, or critique namespace architecture in any project.
The Core Invariant: Dependencies Point Toward Stability
Every require edge from namespace A to namespace B means A depends on B. For the codebase to be maintainable, A should be less stable than B — it should be acceptable for A to change when B changes, but not the reverse.
This creates a stability gradient from top to bottom:
| Layer | Stability | Character |
|---|---|---|
| Entry points | Unstable | Change freely; depended on by few; depend on many |
| Orchestrators | Moderate | Change deliberately; bounded on both sides |
| Domain logic | Stable | Change rarely; depended on by many; depend on few |
| Foundation | Fixed | Never change; no project dependencies |
Violating this — a stable namespace depending on an unstable one — creates fragility. Every change to the unstable thing risks breaking the stable thing.
Namespace Roles
Every namespace falls into one of four roles based on how many things depend on it (fan-in) and how many things it depends on (reach):
Core — depended on by many, depends on few. The stable foundation of the codebase. Should almost never require other project namespaces.
Peripheral — depends on many, depended on by few. Entry points and leaf nodes. Ideally there is at most one per subsystem — the wiring module that assembles everything else. Changes here don't propagate.
Shared — depended on by many and depends on many. A god-module candidate. When you find one, ask whether it has a single clear responsibility. If not, split it.
Isolated — depended on by few and depends on few. Either intentionally a leaf node or a vestigial namespace that should be removed.
Ideal topology: One peripheral namespace wires everything; all others are core. A change to any single namespace reaches a small fraction of the codebase.
No Cycles — The First-Order Rule
Cyclic dependencies make independent change impossible. Two namespaces in a cycle cannot be changed independently, deployed separately, or understood in isolation.
Fix cycles before anything else. Three strategies:
- Merge — if the two namespaces always change together, they are effectively one thing. Merge them.
- Extract — if both depend on a third concept, extract that concept into a new namespace and have both depend on it.
- Invert — if the cycle is a callback pattern, put a protocol in the more stable namespace and move the implementation out.
Three-Lens Coherence
A dependency between two namespaces can be evaluated against three independent signals:
- Structural: does an explicit
requireedge exist? - Conceptual: do they share domain vocabulary?
- Change: do they co-evolve in git history?
These signals should be consistent. The combinations and what they mean:
| Structural | Conceptual | Change | Meaning | Action |
|---|---|---|---|---|
| Yes | Yes | Yes | Confirmed, intentional coupling | None (or merge if inseparable) |
| Yes | Yes | No | Expected coupling | None |
| Yes | No | No | Vestigial dependency | Remove it |
| No | Yes | Yes | Missing abstraction | Extract — highest priority |
| No | Yes | No | Vocabulary sibling | Monitor |
| No | No | Yes | Implicit data contract | Investigate |
Two invariants follow:
- Every structural dependency should have a conceptual justification. If two namespaces require each other but share no vocabulary and never co-change, the dependency may no longer be needed.
- Every active co-evolution should be acknowledged by a structural dependency. If two namespaces consistently change together but have no explicit relationship, something is being left implicit.
Family Structure
Namespaces sharing a common prefix form a family (e.g. app.user.*). Families should behave like modules:
Façade pattern (permitted): A family can have one namespace that many external callers depend on, while internally it delegates to siblings. This looks like a god-module from the outside but is actually a boundary coordinator. It is informational, not alarming.
Noise vs. signal: Namespaces in the same family naturally share vocabulary from the prefix itself. This shared vocabulary is naming convention, not architectural coupling. Only vocabulary that is independent of the prefix counts as a genuine coupling signal.
Invariants:
- A family should expose one surface to the outside world.
- Internal family namespaces should not be required directly by namespaces outside the family unless there is a specific structural justification.
Test Namespace Invariants
- No production namespace depends on a test namespace. Tests are leaves.
- No source code edges should point into test directories.
- A test that requires only its direct subject is a unit test (expected).
- A test that requires an entry point is an integration test (justified).
- A test whose dependencies grew unexpectedly broad has drifted — isolate it.
- Adding tests should not dramatically increase how much of the codebase a single change can touch.
Health Principles
A change should touch a small fraction of the codebase. Propagation cost is the key top-line metric. When it rises on a previously stable codebase, that is a real signal regardless of absolute value.
No cycles. Always the first thing to fix.
Stable things must not depend on unstable things. Each SDP violation is a fragility risk.
Trend matters more than absolute value. A codebase growing steadily more coupled is more concerning than a large but stable propagation cost.