Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/salesforce-pack/skills/salesforce-reliability-patterns/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill salesforce-reliability-patterns. Copyright stays with the author (MIT).

Salesforce Reliability Patterns

Overview

Production-grade reliability patterns for Salesforce integrations: circuit breakers for API outages, idempotent operations using External IDs, graceful degradation with cached data, and dead letter queues for failed operations.

Prerequisites

• jsforce connection configured
• Understanding of Salesforce error codes (see salesforce-common-errors)
• Redis or database for state management (optional)
• opossum or similar circuit breaker library

Instructions

Step 1: Circuit Breaker for Salesforce API

import CircuitBreaker from 'opossum';
import { getConnection } from './salesforce/connection';

// Circuit breaker wraps all Salesforce calls
const sfBreaker = new CircuitBreaker(
  async (fn: () => Promise<any>) => fn(),
  {
    timeout: 30000,                // SF calls can be slow — 30s timeout
    errorThresholdPercentage: 50,  // Open circuit at 50% error rate
    resetTimeout: 60000,           // Try again after 1 minute
    volumeThreshold: 10,           // Need 10 calls before evaluating
    errorFilter: (error: any) => {
      // Don't count client errors as circuit-breaking failures
      const nonCircuitErrors = ['INVALID_FIELD', 'MALFORMED_QUERY', 'REQUIRED_FIELD_MISSING'];
      return nonCircuitErrors.includes(error.errorCode);
    },
  }
);

sfBreaker.on('open', () => {
  console.error('CIRCUIT OPEN: Salesforce API failing — requests will fail fast');
  // Alert ops team
});

sfBreaker.on('halfOpen', () => {
  console.info('CIRCUIT HALF-OPEN: Testing Salesforce recovery...');
});

sfBreaker.on('close', () => {
  console.info('CIRCUIT CLOSED: Salesforce API recovered');
});

// Usage — all SF calls go through the breaker
async function safeSfQuery<T>(soql: string): Promise<T[]> {
  return sfBreaker.fire(async () => {
    const conn = await getConnection();
    const result = await conn.query<T>(soql);
    return result.records;
  });
}

Step 2: Idempotent Operations with External IDs

// Salesforce's upsert with External ID is naturally idempotent
// Same data sent twice = same result (no duplicates)

async function idempotentSync(
  objectType: string,
  records: Record<string, any>[],
  externalIdField: string = 'External_ID__c'
): Promise<{ success: number; failed: number; errors: any[] }> {
  const conn = await getConnection();
  let success = 0;
  let failed = 0;
  const errors: any[] = [];

  // Process in batches of 200 (sObject Collections limit)
  for (let i = 0; i < records.length; i += 200) {
    const batch = records.slice(i, i + 200);

    const results = await conn.sobject(objectType).upsert(batch, externalIdField);

    for (const result of Array.isArray(results) ? results : [results]) {
      if (result.success) {
        success++;
      } else {
        failed++;
        errors.push(result.errors);
      }
    }
  }

  return { success, failed, errors };
}

// Safe to retry — same External_ID__c values will update, not duplicate
await idempotentSync('Account', [
  { External_ID__c: 'EXT-001', Name: 'Acme', Industry: 'Tech' },
  { External_ID__c: 'EXT-002', Name: 'Globex', Industry: 'Manufacturing' },
], 'External_ID__c');

Step 3: Retry with Salesforce-Specific Error Classification

const SF_RETRYABLE_ERRORS = [
  'REQUEST_LIMIT_EXCEEDED',    // API limit — wait and retry
  'SERVER_UNAVAILABLE',        // SF is down temporarily
  'UNABLE_TO_LOCK_ROW',        // Record contention
  'INVALID_SESSION_ID',        // Token expired — re-auth and retry
];

const SF_FATAL_ERRORS = [
  'INVALID_FIELD',             // Code bug — won't fix itself
  'MALFORMED_QUERY',           // Code bug
  'REQUIRED_FIELD_MISSING',    // Data issue
  'INVALID_TYPE',              // Wrong sObject name
  'INSUFFICIENT_ACCESS_OR_READONLY', // Permission issue
];

async function retryableSfCall<T>(
  operation: () => Promise<T>,
  maxRetries = 3
): Promise<T> {
  for (let attempt = 1; attempt <= maxRetries; attempt++) {
    try {
      return await operation();
    } catch (error: any) {
      const errorCode = error.errorCode || error.name;

      if (SF_FATAL_ERRORS.includes(errorCode)) {
        throw error; // Don't retry — it won't help
      }

      if (errorCode === 'INVALID_SESSION_ID') {
        // Re-authenticate, then retry
        await getConnection(); // Forces re-login
        continue;
      }

      if (attempt === maxRetries) throw error;

      const delay = 2000 * Math.pow(2, attempt - 1); // 2s, 4s, 8s
      console.warn(`Retryable SF error ${errorCode}, attempt ${attempt}/${maxRetries}`);
      await new Promise(r => setTimeout(r, delay));
    }
  }
  throw new Error('Unreachable');
}

Step 4: Graceful Degradation with Stale Data

import { Redis } from 'ioredis';
const redis = new Redis(process.env.REDIS_URL);

async function queryWithFallback<T>(
  soql: string,
  cacheKey: string,
  cacheTtlSeconds = 300
): Promise<{ data: T[]; stale: boolean }> {
  try {
    // Try live Salesforce query
    const records = await safeSfQuery<T>(soql);

    // Update cache for fallback
    await redis.set(cacheKey, JSON.stringify(records), 'EX', cacheTtlSeconds * 10);

    return { data: records, stale: false };
  } catch (error) {
    // Salesforce unavailable — serve cached data
    const cached = await redis.get(cacheKey);
    if (cached) {
      console.warn(`SF unavailable, serving stale data for ${cacheKey}`);
      return { data: JSON.parse(cached), stale: true };
    }

    throw new Error(`Salesforce unavailable and no cached data for ${cacheKey}`);
  }
}

// Usage
const { data: accounts, stale } = await queryWithFallback<Account>(
  "SELECT Id, Name, Industry FROM Account WHERE Industry = 'Technology' LIMIT 50",
  'sf:accounts:tech'
);
if (stale) {
  // Show warning to user: "Data may be outdated"
}

Step 5: Dead Letter Queue for Failed Operations

interface SfDeadLetter {
  id: string;
  operation: string;
  objectType: string;
  payload: Record<string, any>;
  errorCode: string;
  errorMessage: string;
  attempts: number;
  firstFailure: Date;
  lastAttempt: Date;
}

class SalesforceDeadLetterQueue {
  async enqueue(entry: Omit<SfDeadLetter, 'id' | 'firstFailure' | 'lastAttempt' | 'attempts'>): Promise<void> {
    const dlq: SfDeadLetter = {
      ...entry,
      id: crypto.randomUUID(),
      attempts: 1,
      firstFailure: new Date(),
      lastAttempt: new Date(),
    };
    await redis.lpush('sf:dlq', JSON.stringify(dlq));
    console.error(`DLQ: ${entry.operation} on ${entry.objectType} failed: ${entry.errorCode}`);
  }

  async reprocess(): Promise<{ processed: number; failed: number }> {
    let processed = 0, failed = 0;
    let entry: string | null;

    while ((entry = await redis.rpop('sf:dlq')) !== null) {
      const dlq: SfDeadLetter = JSON.parse(entry);
      try {
        const conn = await getConnection();
        await conn.sobject(dlq.objectType)dlq.operation;
        processed++;
      } catch (error: any) {
        dlq.attempts++;
        dlq.lastAttempt = new Date();
        if (dlq.attempts < 5) {
          await redis.lpush('sf:dlq', JSON.stringify(dlq));
        } else {
          console.error(`DLQ: Giving up on ${dlq.id} after 5 attempts`);
          // Move to permanent failure store
        }
        failed++;
      }
    }
    return { processed, failed };
  }
}

Output

• Circuit breaker preventing cascading failures
• Idempotent upserts using External IDs
• Error classification (retryable vs fatal)
• Graceful degradation with stale cache data
• Dead letter queue for failed operations

Error Handling

Examples

Make an External-ID upsert resilient to duplicate delivery

Require a stable source event ID, map it to an External ID on the target object, and write through an idempotent upsert instead of a create call. Classify Salesforce responses so permission and schema errors go to a reviewed dead-letter queue while short outages receive bounded retries. Reprocess a synthetic duplicate in a sandbox to prove only one record exists, and alert before the dead-letter backlog threatens the agreed recovery window.

Resources

• Circuit Breaker Pattern
• Opossum Documentation
• External ID Fields
• Salesforce Error Codes

Next Steps

For policy enforcement, see salesforce-policy-guardrails.