Comparisons
Pairs that get conflated in real conversations and in real pull requests — coupling and cohesion, abstraction and indirection, refactoring and rewriting, debt and mess. Neither column wins; what decides is the requirement. Each record leads with the confusion, because the confusion is the reason the record exists.
Mocks vs Fakes
These get used interchangeably and they fail differently, which is why the distinction is worth keeping. A mock asserts on *how* your code did something, so it couples the test to the implementation: rename the method, reorder two calls, add a caching layer, and a test fails even though nothing observable changed. A suite built this way stops being a safety net for refactoring and becomes an obstacle to it, which is exactly backwards, and it is the concrete meaning of the advice against mocking everything. A fake asserts on *what* happened, so it survives refactoring, but it is real code you have to write and keep faithful — a fake repository that ignores a uniqueness constraint the real database enforces will happily green-light a bug. The second confusion is the assumption that one of them is the sophisticated choice. Neither is: the design question underneath is how many collaborators the unit has at all, and a test that needs six of anything is reporting a coupling problem no test-double strategy will fix. A useful default is fakes for anything you own and will call many times, mocks reserved for the few genuinely interaction-shaped assertions, and a contract test making sure the fake and the real thing still agree.
When the interaction itself is the behaviour under test and there is no observable result otherwise: an email was sent, an event was published, a retry happened exactly twice.
When you need the collaborator to actually work so the test can assert on outcomes: an in-memory repository, a fake clock, a stub payment processor that approves some cards and declines others.
| Dimension | Mock — a substitute that records and asserts on the calls made to it | Fake — a working, simplified implementation of the same interface |
|---|---|---|
| Asserts on | The calls that were made | The resulting state or returned value |
| Coupled to | The implementation's call sequence | The interface only |
| Survives refactoring | Often not — that is the main complaint | Yes, as long as the interface holds |
| Cost to create | Low, usually generated by a library | Real code, maintained like any other |
| Reusable across tests | Configured per test | Written once, used by the whole suite |
| Failure mode | A green suite that proves only that the code calls itself | Drift: the fake permits something the real thing rejects |
| Guard against that | Assert on outcomes wherever an outcome exists | A contract test both the fake and the real implementation must pass |
| What both are telling you | How many collaborators the unit needs before it can answer one question | The same thing — and that number is a design finding |
The same question, five structures
Layered, hexagonal, clean, vertical slice and modular monolith — compared without naming a winner, and with the block that says where the comparison stops being true.
A tidy table implies an equivalence that does not exist. These are not five points on one axis: layered, hexagonal and clean are statements about dependency direction; vertical slice is a statement about directory grouping; modular monolith is a statement about deployment and module visibility. Most real systems combine several. The where this comparison misleads block on every row is the part worth reading, and it is the reason this page names no winner — none of these is mandatory, and a team that adopts one because a diagram was pretty has skipped the only question that decides it.
These are not five points on one axis. Layered, hexagonal and clean are all statements about *dependency direction*; vertical slice is a statement about *directory grouping*; and modular monolith is a statement about *deployment and module visibility*. You can — and most real systems do — combine several of them: a modular monolith whose modules are vertical slices, each with a hexagonal boundary at its edges. Comparing them as alternatives is the single most common way this table is misread.
The word complexity is doing two jobs here. Layered and vertical slice are cheap to *set up* and can be expensive to *live in* once the codebase is large; clean and hexagonal are expensive up front and their cost is roughly flat afterwards. Any comparison made at week one inverts the ranking you would get at year three, and neither reading is dishonest — they are answering different questions.
Locality is a property of whether the boundaries match the change history, not of the style name. A vertical slice cut along the wrong capability lines has terrible locality, and a layered codebase with only one real feature has perfect locality. The only honest way to compare these columns is to open the last thirty merged changes in your own repository and count the directories each one touched.
Every column here is a claim about *fast tests without infrastructure*, and any of the five achieves that as soon as dependencies are injected rather than constructed — which is a separate decision none of these styles owns. What differs is the default test boundary each one nudges you toward, and that matters more than the theoretical maximum: layered nudges toward class-level tests with mocks, vertical slice toward feature-level tests, and the difference shows up in how much your suite has to change during a refactor.
The columns are answering to different pressures: hexagonal responds to *external* volatility, clean to *domain* richness, vertical slice to *feature count*, and modular monolith to *team count*. A system can score high on one pressure and low on the rest, which is why picking a style from a comparison table rather than from your own pressures is how teams end up with four rings around a CRUD application.
Team fit is not a tiebreaker, it is often the deciding factor, and it is the one this table cannot capture. A structurally superior design that the team will not maintain under deadline degrades into the worst version of itself — half-applied clean architecture, with some code respecting the ring rule and some not, is harder to work in than consistent layering. The right question is which of these your team will still be following in eighteen months.
This row compares familiarity, not intrinsic difficulty, and familiarity is a property of the industry at a moment in time rather than of the design. Layered wins here largely because it is what most people have seen, which is an argument for it and also the reason it is over-applied. It is also worth separating cost-to-read from cost-to-contribute-correctly: vertical slice inverts on those two, and the table's single number hides it.
Ceremony is only waste when the feature did not need it, and every column here is right for some features and wrong for others in the same codebase. That is the actual finding of this row: a uniform ceremony level applied to every feature guarantees you are overpaying on the simple ones or underpaying on the complex ones. Allowing different features to carry different amounts of structure is more valuable than choosing which column to standardise on.
shared/ directory that grows back under a new name, or slices that each reimplement infrastructure slightly differently.Every one of these degradations is the style's own strength taken past the point where it repays — which is why none of them can be called wrong, and why §138 forbids teaching any as mandatory. What makes a codebase bad is not the column it started in but the absence of anyone asking whether the structure still matches the changes arriving. The right comparison to make is between your current structure and your last thirty changes, not between two names on a page.