How to Handle Concurrency Issues with Parallel LLM Calls in n8n

1// Workflow structure for sequential LLM processing:
2//
3// [Trigger] → [Get Documents] → [SplitInBatches (size=1)]
4//    ↓                                    ↑ (loop back)
5// [LLM Node] → [Save Result] → [SplitInBatches]
6//    ↓ (done)
7// [Merge Results] → [Output]
8
9// In the SplitInBatches node:
10// Batch Size: 1 (one item at a time)
11// Options → Reset: false (to accumulate results)

1// Code node: Rate limit detection and backoff calculation
2const item = $input.item;
3const json = item.json;
4
5// Detect rate limit errors
6const isRateLimit = json.error && (
7  json.error.statusCode === 429 ||
8  json.error.message?.includes('rate limit') ||
9  json.error.message?.includes('Rate limit') ||
10  json.error.message?.includes('Too Many Requests')
11);
12
13if (isRateLimit) {
14  const retryCount = json._retry_count || 0;
15  const maxRetries = 5;
16
17  if (retryCount >= maxRetries) {
18    return [{
19      json: {
20        ...json,
21        status: 'rate_limit_exhausted',
22        error: 'Exceeded max retries due to rate limiting'
23      }
24    }];
25  }
26
27  // Exponential backoff: 2s, 4s, 8s, 16s, 32s
28  const backoffMs = Math.pow(2, retryCount + 1) * 1000;
29
30  // Check for Retry-After header
31  const retryAfter = json.error.headers?.['retry-after'];
32  const waitMs = retryAfter ? parseInt(retryAfter) * 1000 : backoffMs;
33
34  return [{
35    json: {
36      ...json,
37      _retry_count: retryCount + 1,
38      _wait_ms: waitMs,
39      _should_retry: true,
40      status: 'rate_limited'
41    }
42  }];
43}
44
45// Not rate limited — pass through
46return [{
47  json: {
48    ...json,
49    _should_retry: false,
50    status: json.error ? 'error' : 'success'
51  }
52}];

1# Docker Compose for n8n queue mode
2services:
3  n8n-main:
4    image: n8nio/n8n:latest
5    environment:
6      - EXECUTIONS_MODE=queue
7      - QUEUE_BULL_REDIS_HOST=redis
8      - QUEUE_BULL_REDIS_PORT=6379
9      - N8N_CONCURRENCY_PRODUCTION_LIMIT=5
10    command: start
11
12  n8n-worker:
13    image: n8nio/n8n:latest
14    environment:
15      - EXECUTIONS_MODE=queue
16      - QUEUE_BULL_REDIS_HOST=redis
17      - QUEUE_BULL_REDIS_PORT=6379
18      - N8N_CONCURRENCY_PRODUCTION_LIMIT=3
19    command: worker
20    deploy:
21      replicas: 2
22
23  redis:
24    image: redis:7-alpine
25    volumes:
26      - redis_data:/data
27
28volumes:
29  redis_data:

1// Code node: Simple request throttle using static data
2const staticData = $getWorkflowStaticData('global');
3const now = Date.now();
4const MIN_DELAY_MS = 1000; // Minimum 1 second between LLM calls
5
6// Check last request timestamp
7const lastRequest = staticData.lastLLMRequest || 0;
8const elapsed = now - lastRequest;
9
10if (elapsed < MIN_DELAY_MS) {
11  // Need to wait — pass the delay to a subsequent Wait node
12  const waitTime = MIN_DELAY_MS - elapsed;
13  staticData.lastLLMRequest = now + waitTime;
14  return [{
15    json: {
16      ...$json,
17      _throttle_wait_ms: waitTime,
18      _throttled: true
19    }
20  }];
21}
22
23// No throttling needed
24staticData.lastLLMRequest = now;
25return [{
26  json: {
27    ...$json,
28    _throttle_wait_ms: 0,
29    _throttled: false
30  }
31}];