robsonkades/agent-skills/gof-abstract-factory

Abstract Factory

Purpose

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. Choose the mechanism for the promised misuse boundary: tests support trusted assembly but do not reject arbitrary caller combinations. A shared generic family parameter distinguishes types, not two runtime sessions of the same type; use a checked owner identity or hide the products when that distinction is required.

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.

Start with ordinary consumer calls, a relevant advanced use (such as a plugin or session-owned family), and likely misuse. Inspect callers, wiring, supported keys and resource ownership before asking about missing constraints; ask only where the answer changes compatibility or selection. Compare the existing composition, a prebuilt bundle and a provider where creation actually varies.

When it is the answer

There are 2+ product types that must agree with each other
        AND creation or selection needs a coherent family boundary
                → consider Abstract Factory or a prebuilt coherent family.
 
The family is selected once per deployment (profile, environment)
                → existing composition root; in Spring, one @Configuration per family.
                  Verify coherent wiring; profiles and qualifiers do not prove compatibility.
 
The family is selected per request / tenant / document / region
                → 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. A constructor, named factory or Supplier may suffice

(java-object-construction). GoF Factory Method applies when an inherited algorithm delegates creation to a subclass hook; not every creation method is that pattern.

• 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 (plain injection, @MockitoBean for singleton beans 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

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.

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

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 prefer existing composition, with or without a container. A factory called once
     still needs assessment of its SPI, compatibility or lifecycle responsibility.
 
IF the family key arrives from a request, a tenant or a document
THEN select a compatible prebuilt family or provider where creation varies, with
     an explicit failure for an unsupported 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 newly required product has no compatible default
THEN providers must supply it before consumers rely on it. Assess source, binary and
     semantic compatibility; a valid default or separate capability can sometimes avoid
     changing every provider. Do not invent a default merely to keep old plugins loading.
 
IF the factory starts caching what it creates
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. Share stateless providers when their products permit it; confine session-owned

families or define synchronization where state is needed. Select one stable family for the whole operation: separately reading a mutable currentFamily for each product can mix families even if each read is thread-safe. Supplier calls are not an atomic multi-resource acquisition.

• 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
• [ ] 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
• [ ] 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
• [ ] Provider/consumer evolution preserves the required contract or has an explicit migration

Report the family invariant, chosen consumer API and why it fits better than the relevant simpler alternative, enforcement/ownership boundary and actual checks. Keep adequate construction or wiring; unresolved lifecycle or compatibility evidence makes the recommendation conditional.

References

• Decision and alternatives — the family-invariant

test and enforcement choices, 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 — a report-export family selected per request,

built first as a classical hierarchy and then as a prebuilt record bundle, with the tenant-scoped variant, the failure path for an unknown format, and what each version costs. Read when implementing.