Saga Pattern

The Saga pattern manages data consistency across microservices in distributed transaction scenarios without two-phase commit (2PC). A Saga coordinates a sequence of local transactions, executing compensating transactions backward if any intermediate step fails.

When to Use

Distributed Transactions Across Microservices: Coordinating multi-step business transactions without two-phase commit (2PC) locks.
E-Commerce Checkout & Fulfillment: Managing sequences spanning inventory reservation, payment charging, order placement, and delivery dispatch.
Long-Running Business Processes: Orchestrating processes that span seconds, minutes, or hours with explicit failure compensation.
Failure Recovery & State Reversal: Rolling back partially completed distributed steps using compensating transactions.

Quick Start

// Minimal Orchestrator-based Saga Definition
interface SagaStep<TContext> {
  name: string;
  execute: (ctx: TContext) => Promise<void>;
  compensate: (ctx: TContext) => Promise<void>;
}

async function runSaga<T>(context: T, steps: SagaStep<T>[]) {
  const executedSteps: SagaStep<T>[] = [];
  try {
    for (const step of steps) {
      console.log(`Executing step: ${step.name}`);
      await step.execute(context);
      executedSteps.push(step);
    }
    console.log("Saga completed successfully.");
  } catch (err) {
    console.error(
      `Saga failed. Rolling back compensating transactions...`,
      err,
    );
    for (const step of executedSteps.reverse()) {
      try {
        await step.compensate(context);
      } catch (compErr) {
        console.error(`Compensation failed for ${step.name}:`, compErr);
      }
    }
    throw err;
  }
}

Core Concepts

Orchestration vs Choreography Sagas

Sagas coordinate multi-service transactions using either a centralized orchestrator or decentralized events:

[ Orchestrator ] ──1. Reserve Stock──→ [ Inventory Service ]
                 ──2. Charge Card───→ [ Payment Service ] (FAILS!)
                 ──3. Compensate: Release Stock ──→ [ Inventory Service ]

Compensating Transactions (Semantic Rollback)

Because local transactions commit at each step, failures must be undone semantically through compensating transactions:

Orchestrator State Machine Definition

Tracks workflow progression and triggers compensation on error:

// sagas/order-saga-orchestrator.ts
export class OrderSagaOrchestrator {
  async execute(order: OrderPayload) {
    let stockReserved = false;
    try {
      await inventoryService.reserve(order.id, order.items);
      stockReserved = true;

      await paymentService.charge(order.id, order.total);
      await shippingService.schedule(order.id);
    } catch (err) {
      console.error("Saga failed! Initiating compensation...", err);
      if (stockReserved) {
        await inventoryService.compensateRelease(order.id, order.items);
      }
      throw new Error("Order processing failed; changes compensated.");
    }
  }
}

Common Patterns

Idempotent Step Execution

Problem: Network retries during step execution can result in duplicate orders or double charges.

Solution: Pass an idempotency key with every saga step command:

async function processPayment(orderId: string, idempotencyKey: string) {
  const existing = await paymentDb.findByKey(idempotencyKey);
  if (existing) {
    return existing.result; // Return previous outcome without recharging
  }
  const result = await paymentGateway.charge(orderId);
  await paymentDb.save({ idempotencyKey, orderId, result });
  return result;
}

Best Practices

Do:

Prefer Orchestration for Complex Multi-Step Sagas: Use workflow orchestrators (Temporal, AWS Step Functions) for complex multi-branch sagas.
Make Compensating Transactions Idempotent: Ensure compensation actions can be safely retried multiple times without side effects.
Record Saga State Persistently: Store the current step and execution state in a durable outbox/database before calling external services.
Design for Forward Recovery: When possible, retry or alert operators to fix transient issues rather than compensating immediately.

Don't:

Assume compensating transactions can never fail: Implement alerts and manual intervention queues for failed compensations.
Use distributed locks across sagas: Avoid blocking local resources while waiting for distant service responses.
Use choreography for sagas with > 4 services: Event-driven choreography becomes impossible to trace and reason about at scale.

Troubleshooting

| Error | Cause | Solution | | :------------------------- | :--------------------------------------------------- | :-------------------------------------------------------------------------------------------- | | Saga stuck in PENDING | Orchestrator crashed or message dropped in queue. | Use persistent state machines (e.g., Temporal/Step Functions) with automated timeout retries. | | Compensation Failed | Compensating action encountered non-retryable error. | Flag transaction for operator review and write to dead-letter queue (DLQ). | | Duplicate Step Execution | Retry occurred without idempotency check. | Enforce unique idempotency keys per saga step execution. |