bookforge-distribution-boundary-designer（Bookforge 发行边界设计师）

Distribution Boundary Designer

Guides you through the decision of whether to distribute a system across processes or machines, where to draw the boundary if distribution is needed, and how to implement the boundary using the Remote Facade and Data Transfer Object patterns. Grounded in Fowler's First Law of Distributed Object Design and the pattern pair that makes distribution workable when it cannot be avoided.

When to Use

Use this skill when your team is considering breaking a monolith into services, extracting a subsystem as a remote service, reviewing whether a microservices architecture is appropriate, or designing the interface at an existing remote boundary. Also use it when an existing remote API shows signs of the distribution-by-class anti-pattern (too many fine-grained calls per operation), or when domain logic has leaked into remote shells.

Prerequisites: a description of the system's purpose, team structure, deployment needs, and any known scaling or security zone requirements. A codebase or architecture diagram is helpful for detecting existing boundary problems.

Context & Input Gathering

Gather these before proceeding:

Required:

• What does the system do? What are its major subsystems?
• What is the stated motivation for distribution? ("scalability," "team independence," "vendor integration," "security zones," etc.)
• Are there currently separate deployment units? If yes, which ones and why?

Helpful:

• How many teams? Do different teams need to deploy independently?
• Are there subsystems with dramatically different scaling needs (e.g., search scales 100x vs catalog)?
• Are there security zone requirements (DMZ vs internal network)?
• Are there vendor package integrations that run in their own process?
• Is the codebase available? Grep for remote stubs, service clients, or proto/schema files to detect existing boundary shape.

Defaults if not provided:

• Assume a single team if not stated — leans toward not distributing
• Assume uniform scaling if not stated — leans toward not distributing
• Assume no security zone requirement if not stated

Sufficiency check: If the only stated motivation is "we want microservices," "clean architecture," or "team size," the filter step will almost certainly return "do not distribute." Be explicit about this.

Process

Step 1 — Apply the First Law Filter

WHY: The single most common architecture mistake is distributing a system that does not need it. Fowler names this explicitly: "Don't distribute your objects." Distribution incurs latency, partial failure, marshaling cost, and interface rigidity — costs that are paid on every call, forever. Every process boundary is a tax.

Check each of Fowler's seven legitimate distribution reasons. The system should distribute ONLY if at least one applies:

IF none apply → Recommend modular monolith: keep the application in one process, divide it into packages/modules with clear interfaces. Name this explicitly. Do not proceed to boundary design.

IF one or more apply → Identify only the subsystem pairs that cross a legitimate boundary. Every other subsystem pair should remain in-process.

Step 2 — Identify Coarse-Grained Boundaries (per subsystem, not per class)

WHY: The distribution-by-class anti-pattern (one remote component per domain class) multiplies call counts and destroys performance. A remote boundary should span a subsystem — a coherent cluster of domain classes that collaborates internally and exposes a small number of use-case-shaped operations externally.

For each legitimate boundary identified in Step 1:

• Name the subsystem on each side (e.g., "order processing" and "payment processor")
• Identify the operations that cross the boundary — framed as user-intent operations ("place order," "authorize payment"), not as CRUD methods on individual entities
• Estimate call frequency: how many times is this boundary crossed per user interaction? If more than ~5, the boundary may be too fine-grained.

Signal that a boundary is too fine-grained: "GetCustomer() then GetOrders() then GetLineItems() then GetProducts()..." — four separate calls to display a single order screen. This should be one call: GetOrderSummary(orderId).

Step 3 — Design Remote Facade Methods

WHY: A Remote Facade is a thin coordination shell — its only job is to translate coarse-grained remote calls into sequences of fine-grained domain object calls. All domain logic lives in the fine-grained domain objects, not in the facade. A facade with domain logic becomes a second domain layer that is harder to test, harder to evolve, and violates the principle that the application should be runnable entirely in-process without the remote shells.

For each boundary, design the Remote Facade:

1. Methods are use-case-shaped: GetOrderSummary(orderId), SubmitPaymentAuthorization(authDTO), not GetOrder(), GetCustomer(), UpdateOrderStatus()
2. Each method does one useful unit of work: A single remote call should accomplish a meaningful operation, not just fetch one property
3. Facade has no conditional domain logic: No if (order.status == SHIPPED) applyLateCharge() in the facade. That belongs in domain objects.
4. Facade can have multiple methods corresponding to different client use cases: Different screens may call different facade methods that ultimately touch the same domain objects — this is correct
5. Facade granularity: Prefer fewer, larger facades over many small ones. A moderate application might have one; a large application, half a dozen.
6. Security and transactions are appropriate in the facade: The facade is a natural boundary for access control and transaction demarcation (start transaction → call domain objects → commit). This is not domain logic.

Step 4 — Design DTOs Per Use Case

WHY: A DTO is a data carrier designed for the wire, not for the domain. Auto-deriving DTOs from domain classes couples the wire format to the domain model — any domain refactoring forces a wire format change, which may require client updates. DTOs designed around use cases are stable because use cases change less often than domain internals.

For each Remote Facade method, design the corresponding DTO(s):

1. Shape the DTO around what the client interaction needs — not around the domain class structure. Collapse related objects: embed artist name in AlbumDTO rather than passing a separate ArtistDTO with a separate call.
2. Fields are primitives, strings, dates, or nested DTOs — no domain object references. The DTO must be serializable independently of the domain model.
3. Use a separate Assembler to move data between domain objects and DTOs. The assembler keeps the domain model independent of wire format.
4. Consider separate request and response DTOs if the shapes diverge significantly. Use a single DTO if they are similar.
5. Err on the side of sending too much data rather than requiring a second call: "it's better to err on the side of sending too much data than have to make multiple calls."
6. Modern naming: Fowler's DTO = what the J2EE community called "Value Object." Today: gRPC proto messages, JSON response schemas, GraphQL types, OpenAPI schemas, Avro records.

Step 5 — Audit for Anti-Patterns

WHY: The three most common mistakes at distribution boundaries are predictable and detectable in advance.

Check each boundary against these anti-patterns:

Step 6 — Select Interface Style

WHY: The interface style determines coupling tolerance, latency profile, and async capability. Fowler's 2002 guidance was RPC vs XML/HTTP; the modern decision space is richer.

Fowler's principle still holds: use the simplest mechanism that works. If both sides share a platform, use gRPC or its modern equivalent. Use REST/JSON when interoperability across platforms or external access matters. Use message queues when decoupling and async throughput matter more than latency.

Step 7 — Produce Distribution Design Record

WHY: The decision must be documented so future architects understand WHY the boundary exists and what test should be passed before adding more. Distribution decisions are expensive to reverse.

Produce a distribution design record containing:

• Current topology: monolith / modular monolith / partial services
• Legitimate distribution reasons identified (Step 1) and which subsystem pair each applies to
• Non-distribution alternatives considered (modular monolith, package separation)
• Proposed boundaries: subsystem pair, direction of calls, call frequency estimate
• Remote Facade spec: for each facade — class name, method signatures, WHY each method is shaped this way
• DTO spec: for each DTO — fields, which domain objects it aggregates, assembler location
• Interface style selected with rationale
• Anti-pattern audit result: pass / fail per anti-pattern, with remediation notes if any failed

Inputs

• System description (purpose, subsystems, team structure, deployment needs)
• Stated motivation for distribution
• Optional: codebase, architecture diagrams, existing proto/schema files
• Optional: scaling metrics or capacity constraints

Outputs

Distribution Design Record (distribution-design-record.md) containing:

# Distribution Design Record: [System Name]

## Summary Decision
[Distribute / Do Not Distribute — one sentence]

## First Law Filter Results
| Reason | Applies? | Notes |
|--------|----------|-------|
| ...    | Yes/No   | ...   |

## Recommended Topology
[Modular monolith / specific service boundaries]

## Boundaries (if distributing)
### Boundary: [Subsystem A] ↔ [Subsystem B]
- Legitimate reason: [reason from filter]
- Interface style: [gRPC / REST / queue / GraphQL]
- Remote Facade: [FacadeClass]
  - [method(params) → ReturnDTO] — [WHY this use-case shape]
- DTOs:
  - [DTOName]: fields=[...], aggregates=[domain objects], assembled by=[AssemblerClass]

## Anti-Pattern Audit
| Anti-Pattern | Result | Notes |
|--------------|--------|-------|
| Distribution by class | Pass/Fail | ... |
| Domain logic in facade | Pass/Fail | ... |
| Chatty interface | Pass/Fail | ... |
| DTO-domain coupling | Pass/Fail | ... |
| Distribution without reason | Pass/Fail | ... |

## Non-Distribution Alternatives Considered
[What modular monolith structure was evaluated and why distribution was preferred]

Key Principles

1. The First Law is a default, not a suggestion. Distribute only when you have a concrete, operational reason from Fowler's list. "We want microservices" is not a reason. "The payment processor must be PCI-compliant in a separate security zone" is a reason. The default is always in-process.

WHY: Remote calls are orders of magnitude slower than in-process calls. Every process boundary is a permanent tax on every call that crosses it. The cumulative cost of unjustified distribution dwarfs the cost of an in-process module boundary.

2. Boundaries are per subsystem, not per class. A Remote Facade corresponds to a subsystem (a cluster of collaborating domain objects), not to an individual class. If every domain class has a remote interface, the boundary design is broken.

WHY: Each fine-grained call to a remote class pays the full inter-process latency cost. A subsystem-level facade batches those calls into one network round-trip per use case.

3. Facades contain no domain logic — none. A Remote Facade is a translation layer: coarse-grained interface → fine-grained domain object calls. Any conditional logic, business rules, or workflow coordination in the facade must be moved to domain objects or a non-remote Service Layer.

WHY: A facade with domain logic becomes a hidden second domain layer. The application can no longer be tested or run in-process without the remote shell, because the logic lives there. This creates a testing and evolution trap.

4. DTOs are designed for the interaction, not the domain. Each DTO is shaped around a specific client use case — what data that client needs to display or send in a single interaction. DTOs are not auto-generated mirrors of domain classes.

WHY: Domain classes change frequently as business rules evolve. DTOs change when the client interaction changes. Coupling these two change rates together means every domain refactoring potentially breaks clients across the wire.

5. The modular monolith is always on the table. When a team says "we need microservices for separation of concerns," the correct counter-offer is: separate your packages/modules/bounded contexts within a single process. You get team ownership, clear interfaces, and independent evolution without network cost.

WHY: The costs of distribution (latency, partial failure, operational complexity, debugging difficulty) are real and ongoing. The benefits of a modular monolith (same testability, same team ownership, same interface discipline) are available without those costs.

6. Send more data per call rather than fewer. When designing DTOs and facade methods, err toward aggregating more data in one call rather than planning a second call. Over-fetching slightly is far cheaper than a second round-trip.

WHY: Remote call latency is fixed overhead per call. Bandwidth is cheap. If a client might need the order's line items after getting the order header, include the line items in the initial DTO.

Examples

Example 1 — Reject Distribution for a Two-Team LOB App

Scenario: A 12-person engineering team across two squads is building an internal CRM. Engineering lead proposes microservices "for clean boundaries between the sales squad (contacts, deals) and the support squad (tickets, SLAs)."

Trigger: "Should we use microservices for our two-squad CRM?"

Process:

1. Apply First Law Filter:

- Client-server divide: Yes (browser client + server). But this is the standard web-app boundary, not an argument for internal microservices. - Independent scaling: No — contacts/deals and tickets/SLAs have similar load profiles. - Different security zones: No. - Vendor package: No. - Independent deployment: "Nice to have" but not operationally required. - Result: No legitimate reason for internal service distribution.

1. Recommend modular monolith: define a contacts-deals package and a tickets-slas package with explicit module boundaries, shared domain types, and clear internal APIs. Each squad owns their package.
2. Note: the existing web-app boundary (browser ↔ app server) is the natural Remote Facade. A single CRMService facade handles screen-level operations for both squads.
3. Anti-pattern audit: distribution-by-class risk = not applicable (not distributing). Domain logic in facade = not applicable.

Output: Distribution Design Record stating "Do not distribute internally." Modular monolith with squad-owned packages. One Remote Facade for the browser-server boundary.

Example 2 — Extract Payments as a Distributed Boundary

Scenario: E-commerce platform. Payments team requires PCI-DSS compliance in an isolated network zone. The search subsystem handles 50x the load of the product catalog at peak.

Trigger: "We need to separate payments for PCI compliance and search for scaling. How do we design those boundaries?"

Process:

1. Apply First Law Filter:

- Security zone (PCI): Yes — payments must live in an isolated environment. Legitimate. - Independent scaling (search): Yes — 50x load differential is a real scaling reason. - All other subsystems: No legitimate reason.

1. Two boundaries identified: Catalog ↔ Payments, Catalog ↔ Search. Everything else (orders, inventory, user) stays in the monolith.
2. Design Payments Remote Facade (PaymentService):

- AuthorizePayment(orderId, paymentMethodDTO) → PaymentResultDTO - CapturePayment(authorizationId) → CaptureResultDTO - RefundPayment(paymentId, amount) → RefundResultDTO - NOT: GetCard(), UpdateCard(), GetAuthorization() — these are fine-grained domain calls

1. Design DTOs: PaymentMethodDTO (cardToken, billingAddress, amount), PaymentResultDTO (authorizationId, status, errorCode). Assembled from domain card/billing objects. Not coupled to domain class structure.
2. Design Search Remote Facade (SearchService):

- SearchProducts(query, filters, pagination) → ProductSearchResultDTO - SuggestProducts(partialQuery) → SuggestionDTO[]

1. Interface style: Payments = gRPC (internal, same platform, high reliability needed). Search = REST/JSON (browser clients query search directly; must be cross-platform).
2. Anti-pattern audit: no distribution-by-class, facades are thin, DTOs are use-case-shaped.

Output: Distribution Design Record with two boundaries (PCI, scaling), facade specs, DTO specs, interface styles.

Example 3 — Fix Legacy Chatty Distributed Objects

Scenario: Legacy J2EE system from 2005. Every entity bean (Customer, Order, OrderLine, Product, Address) is a separate remote EJB. Displaying an order summary screen makes 15-20 remote calls. Performance is unacceptable.

Trigger: "Our EJB app is painfully slow. Each screen makes dozens of remote calls. How do we fix this?"

Process:

1. First Law Filter: Application server is separate from DB (legitimate). The internal EJB-to-EJB calls are the problem — distribution by class applied to the domain model.
2. Diagnosis: Classic distribution-by-class anti-pattern. The EJBs mirror the domain class structure. Every property getter is a remote call.
3. Remediation:

- Keep domain objects as plain in-process Java objects (POJOs) — remove remote interfaces from Customer, Order, OrderLine, Product. - Introduce a single OrderService Remote Facade with use-case methods: GetOrderSummary(orderId) → OrderSummaryDTO, UpdateOrderStatus(orderId, statusDTO). - OrderSummaryDTO aggregates order header + customer name + line items + product names in one call. - Assembler populates the DTO from the fine-grained domain objects.

1. Interface style: Keep as EJB session bean Remote Facade (same platform) or migrate to gRPC for modern replacement.
2. Expected result: 15-20 calls → 1-2 calls per screen interaction.

Output: Refactoring plan with identified chatty calls, new facade spec, DTO design, migration sequence.

References

• references/distribution-reasons-checklist.md — Fowler's seven legitimate distribution reasons as a quick-reference filter
• references/remote-facade-design-guide.md — detailed guidance on Remote Facade granularity, method design, session facade vs remote facade distinction
• references/dto-design-guide.md — DTO design patterns, assembler pattern, serialization format trade-offs, modern parallels (proto, JSON schema, GraphQL types)
• references/interface-style-selector.md — gRPC vs REST vs message queue vs GraphQL selection criteria

License

This skill is licensed under CC-BY-SA-4.0. Source: BookForge — Patterns of Enterprise Application Architecture by Martin Fowler, David Rice, Matthew Foemmel, Edward Hieatt, Robert Mee, Randy Stafford.

Related BookForge Skills

Install related skills from ClawhHub:

• clawhub install bookforge-enterprise-architecture-pattern-stack-selector
• clawhub install bookforge-domain-logic-pattern-selector

Or install the full book set from GitHub: bookforge-skills