gof-abstract-factory · diff
git:20260903.f199530 to git:20260906.e6ad4bd
25 added, 19 removed. Audit A to A.
---
name: gof-abstract-factory
description: >
Abstract Factory in modern Java: the pattern exists to keep a _family_ of related objects
mutually consistent when the family varies, not to centralise construction. Covers the family
invariant that justifies it, why dependency injection already resolves the deployment-time
- case, when per-request or per-tenant selection makes the factory unavoidable, and how to
+ case, when per-request or per-tenant selection benefits from a family provider, and how to
express it as a record of suppliers or a sealed provider rather than a four-level interface
hierarchy. Use when a factory interface is proposed, when profile-specific object graphs are
being built by hand, when a family of parser/renderer/validator types must never be mixed
across formats, when a plugin SPI must supply several related types at once, or when reviewing
a factory whose products have nothing to do with each other. Does not cover single-product
creation (gof-factory-method), assembling one complex object (gof-builder), copying an
existing instance (gof-prototype), or wiring policy in general (java-dependency-inversion).
---
# Abstract Factory
## Purpose
- Guarantee that objects used together come from the same family. The pattern's product is not
- construction — it is the **impossibility of mixing**: no code path can obtain a Postgres
- repository beside an in-memory unit of work, or a PDF renderer beside an HTML paginator,
- because only the factory hands them out and it only hands out matched sets.
+ Select related products together and preserve their compatibility. Abstract Factory alone does
+ not make mixing impossible: callers can combine products from different factories, and a public
+ aggregate constructor can accept mismatched products. State the enforcement boundary: trusted
+ assembly with contract tests, validated family identities, family-typed APIs, or encapsulated
+ operations that never expose mixable products. Shared family identity may also require the same
+ transaction/session instance, not merely the same vendor or format.
If there is no invariant binding the products to each other, this is not Abstract Factory. It
is a bag of factory methods, and it should be several separate providers or none at all.
## When it is the answer
```text
There are 2+ product types that must agree with each other
- AND the agreement is not checkable by the type system
- AND the family is selected at runtime from data
+ AND creation or selection needs a coherent family boundary
→ Abstract Factory
The family is selected once per deployment (profile, environment)
→ dependency injection: one @Configuration per family.
- The container makes mixing impossible already.
+ Verify coherent wiring; profiles and qualifiers do not prove compatibility.
The family is selected per request / tenant / document / region
- → Abstract Factory, keyed by that value. The container
- cannot decide what varies per call.
+ → select a coherent provider or prebuilt family by key.
+ DI may supply that registry; use factory methods when creation varies.
Third-party code must contribute a whole family
→ Abstract Factory as the SPI shape (ServiceLoader
provider returning the family, not N providers).
```
## When it is not
- **One product type.** That is Factory Method or a `Supplier`; the "abstract" in the name is
precisely the multi-product part.
- **The products are unrelated** — `createRepository`, `createHttpClient`, `createClock`. This
is a service locator with a factory's name, and it re-couples every caller to one type that
knows everything (`gof-pattern-antipatterns`).
- **The family differs only in constants.** Rates, endpoints, limits and timeouts are data. A
class per value is the commonest false Abstract Factory; use configuration instead.
- **Only one family exists, and the second is speculative.** This weakens the case, but does not
decide it: an interface can still be justified as a module or plugin boundary, an ownership
seam, or a stable port. Record that reason; otherwise defer the abstraction until a second
family reveals the real common contract.
- **Testing was the only motivation.** First prefer substituting collaborators at an existing
boundary (`@MockitoBean` in Spring Framework 6.2+, or a test `@Configuration`). A production
family abstraction can still be warranted when the coherent in-memory family is itself a
useful contract, not merely a test hook.
## Modern Java expression
- Prefer a record of factory functions over an interface hierarchy when each product is a single
- constructor call. It carries the same guarantee — one object, one family — with none of the
- class explosion, and the family is a value that can be put in a `Map`:
+ A record of factory functions can package a small family without additional implementation
+ classes. Its constructor and suppliers still need compatibility, null, freshness and ownership
+ contracts; final references do not make captured state or products thread-safe. A record of
+ already-created products is a family bundle, not a factory of fresh products.
+ The examples target Java 17 without preview features (records and sealed types); pattern switches
+ over sealed hierarchies require Java 21 to avoid preview. Inspect the project's actual release and
+ dependencies; the pattern also works with ordinary classes on older Java without upgrades.
+
```text
Classical Modern
───────────────────────────────── ────────────────────────────────────
interface ReportFactory record ReportFamily(
Renderer newRenderer() Supplier<Renderer> renderer,
Paginator newPaginator() Supplier<Paginator> paginator,
StyleSheet newStyleSheet() Supplier<StyleSheet> styles)
class PdfReportFactory implements static ReportFamily pdf()
class HtmlReportFactory implements static ReportFamily html()
selection: if/else or a Map Map<Format, ReportFamily>, or a
sealed Format with exhaustive switch
```
Keep the interface when a product needs more than construction from the family — shared
configuration, a `supports()` predicate, a lifecycle to close — or when third parties implement
it, since an interface is a stabler SPI contract than a record's component list.
## Decision rules
```text
IF the products can be used in any combination without breaking
THEN there is no family. Inject each product independently.
IF the family is fixed at startup by profile or property
THEN dependency injection. Do not add a factory the container calls once.
IF the family key arrives from a request, a tenant or a document
THEN Abstract Factory keyed by that value, with an explicit failure for
an unknown key — never a silent default family.
IF the key comes from outside the process
THEN validate it against the supported registry before selection. Never turn an
untrusted class name into reflective loading; an extensible plugin key need not
be a compile-time closed enum, but it still needs authorization and failure policy.
IF a new product is added to the family
THEN every implementation must change. If that is unacceptable, the
family is not stable enough for this pattern — reconsider.
IF the factory starts caching what it creates
- THEN that is Flyweight or a singleton scope arriving unannounced;
- make the lifetime explicit (gof-flyweight, gof-singleton).
+ THEN define sharing, eviction, closure and thread safety. A cache alone is neither Flyweight
+ nor Singleton; select those patterns only if their separate intent fits.
```
## Cross-cutting checks
- **Concurrency.** A factory is normally a stateless immutable value shared by all threads —
keep it that way. The moment it holds mutable state (a cache, a counter, a "current family"
field) it needs a memory model argument, and a mutable `currentFamily` field is a race that
hands out mixed families under load.
- **Distribution.** The pattern is process-local. A "remote factory" that returns handles to
objects living elsewhere is a Proxy problem with the failure semantics that implies
(`gof-proxy`). Where families correspond to protocol or schema versions, the selection is
capability negotiation and needs an explicit unsupported-version path.
- **Performance.** The pattern does not imply allocation: a family may return cached, pooled or
newly constructed products. Dispatch may inline at stable call sites and may become
megamorphic with many implementations. Neither effect is a design-level reason to adopt or
reject the pattern; profile the actual construction and call sites
(`jit-inlining-and-escape-analysis`).
- **Testing.** The legitimate testing benefit is a whole coherent in-memory family, which makes
integration-style tests fast without mocks. The illegitimate one is a factory added so that a
single collaborator can be stubbed — inject that collaborator instead.
## Review checklist
- [ ] There are two or more products, and a stated invariant binds them
- - [ ] Mixing families is impossible by construction, not by convention
- - [ ] The selection key is named, closed, and validated when externally supplied
+ - [ ] The actual compatibility enforcement boundary is explicit and tested
+ - [ ] The selection key is named and validated against the authorized supported registry
- [ ] An unknown key fails loudly rather than falling back to a default family
- - [ ] At least two families exist today
- - [ ] The factory holds no mutable state
+ - [ ] Multiple families exist, or a concrete SPI/module boundary justifies the abstraction
+ - [ ] Mutable factory, supplier and product state has an explicit concurrency/lifetime contract
- [ ] The products differ in behaviour, not only in configuration values
- [ ] Adding a product to the family is an acceptable change to every implementation
## References
- [Decision and alternatives](references/decision-and-alternatives.md) — the family-invariant
test, Abstract Factory against dependency injection, `Map<Key, Supplier>`, `ServiceLoader` and
configuration, and how it differs from Factory Method and Builder. Read before introducing or
removing a factory interface.
- [Worked example](references/worked-example.md) — a report-export family selected per request,
built first as a classical hierarchy and then as a record of suppliers, with the tenant-scoped
variant, the failure path for an unknown format, and what each version costs. Read when
implementing.