n8n Production Mastery

• Demo-ready vs production-ready: the five failures that never show up in testing
• The 6-dimension model — a shared language for workflow reliability
• Reading your own workflow's risk in 10 minutes

Exercise · Score one of your live workflows across the 6 dimensions and find its weakest two.

• Why duplicate triggers happen: retries, at-least-once delivery, redeploys
• Dedup patterns: lookup-before-insert and hash-based replay detection
• Idempotency keys for external API calls
• Making multi-item batches safe to re-run

Exercise · Add a dedup guard to a webhook workflow and prove it survives a double-fire.

• The three tools: Retry on Fail, error output, and the Error Trigger workflow
• Transient vs permanent: what to retry and what to dead-letter
• Exponential backoff with jitter
• Failing loud: alerting by severity

Exercise · Add an Error Trigger workflow + dead-letter table, then replay a failed item.

• Correlation IDs across multi-step workflows
• Checkpoint logging to a Postgres audit table
• Append-only logs and compliance-ready retention

Exercise · Add audit logging, then answer “did we process order X on Tuesday?” with one query.

• Credential store vs hardcoded secrets — and how to find leaks
• Environment variables, external secrets managers, and rotation
• Least-privilege scopes per integration
• Securing webhook endpoints with HMAC validation

Exercise · Move a workflow's hardcoded secrets to the credential store and add HMAC verification.

• Failure alerts that reach where your team actually looks
• Heartbeat / dead-man's-switch monitoring for when n8n itself goes down
• Execution metrics, anomaly detection, and SLA tracking

Exercise · Wire failure alerts + a heartbeat monitor, simulate an outage, and confirm you're notified.

• Queue mode and worker architecture
• Batching, pagination, and respecting rate limits
• PostgreSQL tuning, Redis config, and connection pooling
• Load testing before you need it

Exercise · Convert a single-instance workflow to queue mode and load-test it to find its ceiling.

• Why AI agents fail differently: non-determinism, cost, and rate limits
• Idempotent tool calls and safe retries for LLM steps
• Guardrails: structured outputs, validation, and fallback paths
• Cost and rate-limit controls that won't trip under load

Exercise · Harden a RAG/agent workflow: add output validation, a cost cap, and a graceful fallback.

• Auditing a real workflow end-to-end
• Prioritising fixes by blast radius
• Before/after: proving it now survives failure, load, and replay

Exercise · Submit your hardened workflow against the 6-dimension scorecard — your proof of mastery.