# Pattern Selection And Composition > Choosing enterprise patterns from forces rather than familiarity, and combining them into an architecture whose parts reinforce rather than fight each other: the selection criteria that discriminate, the compositions that work, the pairs that conflict, and the relationship graph. Use when a design is starting and the patterns are about to be chosen by habit, when a pattern name is proposed before the problem is stated, when two chosen patterns produce friction, when a reference architecture is being copied wholesale, when someone asks which enterprise patterns a new module should use, or when an architecture must be explained as a set of decisions. Does not cover the individual patterns' guidance, whether the framework provides one (patterns-and-modern-frameworks), or detecting overuse (enterprise-architecture-smells). - Skill: `robsonkades/pattern-selection-and-composition` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add robsonkades/pattern-selection-and-composition` - Raw SKILL.md: https://api.skillmd.com/api/skills/robsonkades/pattern-selection-and-composition/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: robsonkades (https://skillmd.com/u/robsonkades) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/robsonkades/pattern-selection-and-composition --- # Pattern Selection and Composition ## Purpose Derive an architecture from forces, and check that its parts fit. Patterns are the _output_ of reasoning about a problem, never the input — "we will use a domain model with repositories and DTOs" stated before the forces are known is a preference, and it will be defended rather than tested. The second half matters as much: patterns interact. Some pairs reinforce each other (Domain Model with Unit of Work and Identity Map); some fight (Domain Model with a repository that returns rows on its invariant-enforcing write path; a facade that leaves remote calls chatty). A design of individually defensible choices can still be incoherent. ## The selection inputs ```text Business complexity do the rules interact? how many are conditional on other rules? Data complexity does the object model diverge from the schema? who owns the schema? Work shape per-instance decisions, or set-shaped work? Concurrency conflicts within a transaction, or across a user's thinking time? Transaction scope one write, several writes, or across a boundary? Distribution one process, or several? whose driver? Performance round-trip budget; read/write asymmetry Team and lifespan who maintains it, for how long, at what size Operational constraints deploy cadence, ownership, regulatory limits ``` Everything below is derived from these. If a pattern cannot be traced back to one of them, it has no justification yet. ## Workflow 1. **Inspect evidence for the relevant inputs** for the module — not just the system. Reuse accepted constraints and existing code. Record consequential unknowns and investigate the smallest missing basis; do not require nine new answers for a small decision. 2. **Start with logic organisation, then iterate with fixed schema, transaction, concurrency and delivery constraints.** It constrains many downstream choices but is not a one-way dependency (`domain-logic-organization`). 3. **Choose the data-source pattern** consistent with it (`data-source-patterns`). 4. **Add only the patterns a named force requires.** Each addition must trace to an input. 5. **Check the composition** against the conflicts list in the references. Fix the friction by removing unjustified machinery or correcting boundaries. An adapter is justified when independently owned contracts must coexist; measure its cost instead of forbidding it. 6. **Write down the choices with their forces**, so the next person can re-open a decision on evidence (`architecture-decision-making`). ## Selection rules These are candidates to compare with the existing design, not mandatory stacks. Schema ownership does not alone require duplicate object models; use explicit mapping when the semantic/API independence gained justifies conversion and synchronization costs. ```text Rules do not interact; work is per-transaction → Transaction Script + Table Data Gateway. Add a Service Layer when transaction, authorization, orchestration or a stable use-case boundary earns it—not from a write-count threshold. Rules interact; invariants span objects; schema is yours → Domain Model (entities with behaviour) + Data Mapper (JPA) + Repository per aggregate + Service Layer for the transaction. Rules interact; schema is owned elsewhere or must diverge → Domain Model + separate persistence model + explicit Mapper. Pay the mapping; you are buying independence. Mostly CRUD; entity ≈ table; you own the schema → consider Active Record where the record owns persistence, or a simple mapper/repository-backed CRUD design already supported by the stack. Add service/DTO boundaries only for a named contract or policy. Work is set-shaped (bulk recalculation, indexation, reporting) → SQL in a gateway, beside whatever the write side uses. Not a domain model looping over objects. Reads are slow because they go through the write model → add a read model (projections/query objects). This is the most common missing pattern in otherwise good designs. Concurrent edits across a user's thinking time → Optimistic Offline Lock, coarse-grained at the aggregate. Pessimistic only when the lost work is expensive. A remote boundary exists → define coarse-enough operations and an explicit wire contract. A dedicated DTO is usual; a stable schema-generated or immutable boundary type can already satisfy that role. A remote boundary is being CONSIDERED → module boundary first; distribute only for a named driver (distribution-boundaries). Multi-request conversation state → place it per item (session-state-strategies), not as a session by default. ``` ## Composition rules - **ORM-backed Domain Models commonly use Unit of Work and Identity Map.** A domain model alone implies neither; explicit SQL, immutable models and event persistence can differ. Inspect actual persistence semantics rather than inventing ORM behavior (`orm-behavioral-patterns`). - **Repository provides collection-like access to domain objects.** When adopting DDD aggregates, organize domain repositories around aggregate roots and protect their mutation boundaries; absence of aggregates alone does not make a Repository a DAO (`repository-pattern`). - **Remote Facade implies DTOs.** A coarse operation returning domain objects re-couples the caller to the model, and the facade's whole purpose was one round trip with a stable payload. - **Optimistic Offline Lock implies an identity field, a version and a conflict experience.** The first two are mechanical; the third is where implementations fail. - **Service Layer defines an application operation boundary.** Transaction ownership, authorization, orchestration and multiple callers can justify it; read-only or remote operations need not start a database transaction (`service-layer-design`). - **Set-based operations need explicit invariant and concurrency handling.** Table Module can own business rules over tabular data; it does not inherently bypass every invariant. - Reads and writes may use different patterns when query and invariant forces diverge. This is the most under-applied composition in enterprise architecture and the one that resolves most performance–purity arguments (`architecture-and-performance`). ## Conflicts to check for ```text Domain Model + repositories returning rows/DTOs → check write paths for bypassed invariants; separate read projections are compatible. DDD aggregate + independent write repositories for its child tables → inspect whether child writes bypass root invariants; table-oriented infrastructure alone does not prove the aggregate boundary is broken. Active Record + a Repository abstraction + DTOs everywhere → paying a Data Mapper's price for Active Record's coupling. Choose one position. Transaction Script + a rich domain model half-built → two homes for every rule; the rule will be in the wrong one. Remote Facade + fine-grained service methods behind it → compatible for local calls: aggregating them is the facade's purpose. If calls remain remote, inspect the remaining network hops and latency. Aggregate + bulk updates that skip the version → optimistic locking is silently defeated (offline-concurrency-control). Lazy Load + entities crossing a boundary → LazyInitializationException, or Open Session In View and its costs. Coarse-Grained Lock + a large aggregate → measure contention and invariant scope before resizing; splitting an atomic invariant introduces a coordination problem, not a free optimization. Distribution + a shared database → coupled schema, availability and ownership. This may be a deliberate transitional or jointly owned architecture; it does not by itself prove one service, but it weakens independent evolution. ``` ## References Return the selected/rejected composition, the relevant evidence and constraints, its main trade-off, and a concrete validation or reconsideration condition. Do not infer a performance improvement from pattern names; inspect versions and measure the target implementation. - [Selection criteria](references/selection-criteria.md) — the nine inputs turned into questions with observable answers, worked selections for four common system shapes, and the decisions that should differ per module rather than per system. Read when starting a module or reviewing a proposed set of patterns. - [Reference architectures and the relationship graph](references/reference-architectures.md) — four complete compositions traced end to end (rich domain, transaction script, remote API, read/write split), the pattern relationship graph showing which pattern implies which, and the consequences of each composition stated concretely. Read when assembling an architecture or explaining an existing one.