2nd/10_Wiki/Topic_Programming/Architecture/Mediator Topology.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-mediator-topology	Mediator Topology	10_Wiki/Topics	verified	self	Event Mediator Topology Mediator-Based Event-Driven Architecture	none	A	0.9	applied	event-driven architecture-topology orchestration		2026-05-10	pending	language framework typescript temporal

Mediator Topology

매 한 줄

"매 central orchestrator 가 매 multi-step event flow 를 직접 지휘". Mark Richards, Software Architecture Patterns — Event-Driven Architecture 의 두 topology 중 하나 (다른 하나는 Broker). Mediator 는 매 sequence + error handling + transactional rollback 이 필요한 매 complex workflow 에 적합. 2026 의 modern 구현: Temporal, AWS Step Functions, Camunda 8, Netflix Conductor.

매 핵심

매 Mediator vs Broker (Richards 의 dichotomy)

• Broker: 매 chain of events, decentralized, 매 service 가 매 다음 step 결정. Pub/sub. 매 simple, fast, hard-to-orchestrate complex workflow.
• Mediator: 매 central event mediator 가 매 process 를 orchestrate. 매 sequence 명시, 매 error handling 중앙화. 매 complex but observable.

매 Mediator components

1. Event Queue: incoming initial events.
2. Event Mediator: 매 process orchestrator (workflow engine).
3. Event Channel: mediator → processor 의 dispatch.
4. Event Processor: 매 single step 실행 (idempotent worker).

매 응용 (when to use mediator)

1. Order fulfillment (validate → pay → reserve → ship → notify).
2. Loan approval (multi-stage with human-in-loop).
3. Subscription onboarding.
4. ETL/data pipeline with retries and DLQ.
5. SAGA pattern compensation orchestration.

매 trade-off vs Broker

• More observable: 매 mediator 의 single source of truth for workflow state.
• More coupled: 매 mediator 의 process knowledge — 매 evolution 시 central change.
• Lower throughput: 매 mediator 의 single point.
• Easier to reason: 매 sequence 가 명시 — debugging.

💻 패턴

1. Temporal workflow (mediator, 2026 standard)

// activities.ts (event processors)
export async function validateOrder(o: Order) { /* ... */ }
export async function chargePayment(o: Order) { /* ... */ }
export async function reserveInventory(o: Order) { /* ... */ }
export async function ship(o: Order) { /* ... */ }
export async function notify(userId: string) { /* ... */ }

// workflow.ts (mediator)
import { proxyActivities } from "@temporalio/workflow";
import type * as activities from "./activities";

const acts = proxyActivities<typeof activities>({
  startToCloseTimeout: "1 minute",
  retry: { maximumAttempts: 5 },
});

export async function orderWorkflow(order: Order) {
  await acts.validateOrder(order);
  try {
    await acts.chargePayment(order);
    await acts.reserveInventory(order);
    await acts.ship(order);
  } catch (err) {
    // SAGA compensation
    await acts.refundPayment(order).catch(() => {});
    await acts.releaseInventory(order).catch(() => {});
    throw err;
  }
  await acts.notify(order.userId);
}

2. AWS Step Functions (managed mediator)

{
  "StartAt": "Validate",
  "States": {
    "Validate": { "Type": "Task", "Resource": "...:validate", "Next": "Charge" },
    "Charge":   { "Type": "Task", "Resource": "...:charge",
                  "Catch": [{ "ErrorEquals": ["PaymentError"], "Next": "Refund" }],
                  "Next": "Reserve" },
    "Reserve":  { "Type": "Task", "Resource": "...:reserve", "Next": "Ship" },
    "Ship":     { "Type": "Task", "Resource": "...:ship",     "Next": "Notify" },
    "Notify":   { "Type": "Task", "Resource": "...:notify",   "End": true },
    "Refund":   { "Type": "Task", "Resource": "...:refund",   "Next": "Fail" },
    "Fail":     { "Type": "Fail" }
  }
}

3. Camunda 8 (BPMN mediator)

<!-- order.bpmn (visual model, executable) -->
<bpmn:process id="orderProcess">
  <bpmn:startEvent id="OrderReceived" />
  <bpmn:serviceTask id="Validate" zeebe:type="validate-order" />
  <bpmn:serviceTask id="Charge"  zeebe:type="charge-payment" />
  <bpmn:exclusiveGateway id="ChargeOk?" />
  <bpmn:serviceTask id="Reserve" zeebe:type="reserve-inventory" />
  <!-- ... compensation events for SAGA -->
</bpmn:process>

4. Custom mediator (lightweight)

type Step<S> = { name: string; run: (s: S) => Promise<S>; compensate?: (s: S) => Promise<void> };

class Mediator<S> {
  constructor(private steps: Step<S>[]) {}

  async execute(initial: S): Promise<S> {
    const completed: Step<S>[] = [];
    let state = initial;
    try {
      for (const step of this.steps) {
        state = await step.run(state);
        completed.push(step);
      }
      return state;
    } catch (err) {
      // SAGA compensate in reverse
      for (const step of completed.reverse()) {
        await step.compensate?.(state).catch(() => {});
      }
      throw err;
    }
  }
}

const orderMediator = new Mediator<Order>([
  { name: "validate", run: validateOrder },
  { name: "charge", run: chargePayment, compensate: refundPayment },
  { name: "reserve", run: reserveInventory, compensate: releaseInventory },
  { name: "ship", run: ship },
]);

5. Broker (contrast — for comparison)

// No central mediator. Each service publishes its own next event.
// order-service: emits "order.validated"
// payment-service: subscribes "order.validated" → publishes "order.paid"
// inventory-service: subscribes "order.paid" → publishes "inventory.reserved"
// shipping-service: subscribes "inventory.reserved" → publishes "order.shipped"
//
// Pros: decoupled, scalable
// Cons: workflow knowledge scattered, hard to add/reorder steps

6. Compensation pattern (SAGA)

// Within mediator, each step has forward + compensating action
const steps = [
  { do: chargePayment,       undo: refundPayment },
  { do: reserveInventory,    undo: releaseInventory },
  { do: scheduleShipment,    undo: cancelShipment },
];

async function saga<S>(state: S) {
  const done: typeof steps = [];
  for (const s of steps) {
    try { await s.do(state); done.push(s); }
    catch {
      for (const d of done.reverse()) await d.undo(state).catch(() => {});
      throw new Error("saga compensated");
    }
  }
}

7. Idempotency (essential for mediator workers)

async function chargePayment(order: Order) {
  const idempotencyKey = `charge:${order.id}`;
  const existing = await db.idempotency.findOne(idempotencyKey);
  if (existing) return existing.result;

  const result = await stripe.charges.create({
    amount: order.amount,
    customer: order.customerId,
    idempotency_key: idempotencyKey,
  });
  await db.idempotency.put(idempotencyKey, result);
  return result;
}
// Mediator may retry → worker MUST be idempotent.

매 결정 기준

상황	Topology
<5 step linear flow, simple notify	Broker (pub/sub).
Multi-step with rollback / SAGA	Mediator.
Long-running (hours/days) workflow	Mediator (Temporal/Step Functions).
Human approval steps	Mediator (BPMN — Camunda).
High-throughput stream (>10k/s)	Broker (Kafka).
Audit + visibility critical	Mediator.
Loose coupling priority	Broker.

기본값: complex business process 는 매 Mediator (Temporal). Simple notification/log 는 매 Broker (Kafka/SNS).

🔗 Graph

• 부모: Event-Driven Architecture
• 변형: Broker Topology
• 응용: Temporal
• Adjacent: Idempotency

🤖 LLM 활용

언제: workflow architecture choice, SAGA design, multi-step business process modeling, retry/compensation strategy. 언제 X: 매 simple fire-and-forget event, 매 high-throughput streaming (broker preferred).

❌ 안티패턴

• God mediator: 매 매 모든 process 가 매 single mediator 에 — 매 monolith of orchestration. 매 domain 별 분리.
• Non-idempotent worker: 매 mediator retry → 매 double-charge / double-ship. 매 idempotency key 필수.
• Mediator knows business rules: 매 if/else 분기 가 매 mediator 에 — workflow engine 의 limit. 매 decision 은 activity 로 push.
• Sync mediator for slow steps: 매 mediator 가 매 slow IO 의 inline wait — 매 timeout. 매 async + signal/callback.
• No timeout/retry policy: 매 hung workflow 의 매 stuck instance.
• Compensation 누락: 매 SAGA 의 매 partial commit 의 매 inconsistent state.

🧪 검증 / 중복

• Verified (Richards, Software Architecture Patterns O'Reilly 2015; Temporal docs; Hohpe & Woolf, Enterprise Integration Patterns — Process Manager).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — Mediator vs Broker + Temporal/Step Functions/Camunda 2026 implementations