Baseline: what this repository's rules are compared against
The review is finished only when it names the baseline. This file names three, one per role, and says how each is compared.
The baselines
| Role | Baseline | Why this one |
|---|---|---|
| Method | ADD 3.0 (Cervantes 2016) | a named, repeatable method driven by quality attributes, set out in a practitioner book (publisher page read); it has no controlled test |
| Principle | information hiding (Parnas 1972) | the oldest stated criterion for decomposition in the leads; one Q/A answer still gives it as the reason for private members (SESE 2011k); the criterion has no controlled test, while the broader modularity measures have observational support (review.md 2) |
| Check | ATAM (Kazman 2000) | the evaluation most often cited before building; 31 projects of data exist (Bellomo 2015, abstract read) |
- ATAM as Kazman 2000 prescribes it (read in full): nine steps in four groups. Presentation: the method, the business drivers, the architecture. Investigation: the architect names the approaches, a utility tree of quality attributes is built down to prioritised scenarios, and the approaches are analysed for risks, sensitivity points and tradeoff points. Testing: stakeholders brainstorm and vote on more scenarios, which are run against the approaches again. Reporting: the findings go back to the stakeholders. The report gives one case and no measure of effect.
- For a decision about monoliths and services, the practice baseline is Lewis 2014 with Fowler 2015 (monolith first).
- For a structural check in a build, the practice baseline is a fitness function (ThoughtWorks 2017).
How a rule of this repository is compared
- Task: a person given a requirements brief produces a structure; judges rate it on the attribute scenarios the brief states.
- Conditions: with ADD 3.0 steps, with a plain list of the principles, and with no guide; a crossover design so each person meets each condition.
- Measures: scenarios met out of scenarios stated, hours spent, and a count of dependency cycles.
- Report effect sizes with sample sizes; mark any number from a summary "(secondary)".
- A decomposition rule is checked on a built system: the change cost of a stated change (files touched) before and after.
What a win would be
- The new rule meets as many scenarios as ADD 3.0 in fewer hours, or meets more in the same hours.
- No such study exists yet; this is a proposal, not a result.
Not compared
- Whether a structure is good in the abstract.
- Documentation of the structure: see architecture-descriptions.
- Runtime behaviour across machines: system-design.