n8n workflows ready to run — AI, SEO & DevOps

Default retries in most n8n nodes are 3 attempts, 5 seconds apart. That's fine for brief network blips but wasteful for rate-limited APIs and useless for permanent failures like a 401.

Three patterns that actually work:

• Exponential backoff — use 1s, 4s, 16s intervals. Most rate-limit windows clear within that range without hammering the endpoint.
• Dead-letter queue — when retries are exhausted, route the failed item to a Google Sheet or Slack channel labeled "Needs review." You can batch-reprocess them later.
• Conditional retry — use an IF node before retrying. HTTP 429/503 → retry. HTTP 400/401 → no retry, those won't fix themselves.

Wrap risky flows in an Error Trigger workflow. It catches failures across all executions without bloating individual flows.

Getting 429 errors? Don't crank up retries — that makes it worse. The API is literally telling you to slow down.

• Respect the Retry-After header. Most APIs return it. Use a Wait node with a dynamic delay: {{ $json.headers['retry-after'] * 1000 }}
• SplitInBatches + pacing. A 2-second Wait between batches of 10 keeps you under most rate limits.
• Cache aggressively. If you're fetching the same data multiple times per run, store it once in a Set node and reference it downstream. Most rate-limit issues are self-inflicted.

For spiky traffic (webhooks firing all at once), use Redis as a rate limiter: each incoming webhook increments a counter with TTL; over threshold → queue instead of drop.

A prompt that works once can fail 20% of the time at scale. Three rules prevent that:

• Force structured output. Always request JSON with a schema, include a sample. OpenAI supports response_format: { type: "json_object" } natively — use it.
• Low temperature in production. 0.1-0.3 for extraction, classification, or anything deterministic. Save 0.7+ for creative writing only.
• Parse defensively. Wrap JSON.parse in try/catch in a Code node. On failure, log to a review sheet and use safe defaults — never crash the whole flow.

Before shipping, test your prompt against 20 edge cases. The 5% of inputs that break it are the ones your users will hit first.

Hardcoded API keys in Code nodes are a disaster waiting to happen. Three must-dos:

• Set N8N_ENCRYPTION_KEY. Without it, credentials sit in plaintext in your database. Set on first install, back it up — if you lose it, you lose all saved credentials.
• Use the Credentials UI, not env vars. n8n's credential store is encrypted at rest and supports OAuth refresh. Env vars expose secrets to every code node and log file.
• Rotate every 90 days. Calendar reminder. Boring, but when something leaks you'll thank yourself.

For Docker: mount /home/node/.n8n as a persistent volume and back it up daily. That's where your credentials database lives.

Most n8n failures in production are things that worked on your machine. Four habits catch them:

• Execute Node button, not Execute Workflow. Click each node individually. Lets you inspect intermediate output and catch schema mismatches early.
• Write to a test sheet first. Point at a test tab until you're sure. Creating 500 duplicate rows while debugging is painful.
• Check empty and malformed inputs. What happens if the API returns an empty array? A null field? A string instead of a number? n8n is loose about types.
• Real data only. Don't test with hand-crafted clean payloads. Grab a real webhook and pipe it through. Ugly real-world data is what breaks flows.

n8n's built-in versioning only tracks the current workflow. For real version control:

• Export JSON weekly. Use the UI Export button or CLI: n8n export:workflow --id=42 --output=./
• Commit to git. Even a private gist works. You get diffs, history, and rollback when an update breaks something.
• Tag stable versions. When a workflow hits production and works: git tag prod-2026-04-18. If a later change breaks things, you know exactly what to revert to.
• Write real commit messages. Not "updated workflow" — write what changed and why: "Added retry on 429 for Shopify API, removed deprecated legacy pagination."

Bonus: store the workflow JSON alongside the application code it talks to. One repo, one history, one truth.