Work snippet
n8n and Make error monitoring with retries and a dead-letter queue
Concept DesignIllustrative integration walkthrough. Business details, figures and records are worked examples.
A five-branch service group runs fourteen automations between n8n and Make. When one fails, the error is sorted by type, retried automatically only if a repeat cannot double-post, otherwise set aside with its full payload, and sent to the person who owns it at the right level of urgency.
The problemAutomations fail silently, and retries sometimes create duplicates.
What changesEach failure is classified, retried safely or parked, and reported to someone who can fix it.
- Area
- Automation reliability
- Tools
- n8n · Make · Airtable · Slack
- Typical scope
- US$1,500-3,000
- Timeline
- About 4 weeks · fixed milestones
The problem
Silent failures leave invoices unposted, and no one knows which run broke or whether re-running it will post everything twice.
- The same failure is retried in a tight loop and hits the rate limit harder
- A partial write leaves a job in one system and not in the other
- Alerts fire 20 times for one outage, then get muted
- Nobody owns the item, and the payload is already gone
What discovery would examine
A read-only review of 14 Make scenarios and n8n workflows for a five-branch service group, plus two weeks of run history and a 45-minute call with the operations lead.
| ID | Finding | Evidence | Risk if left |
|---|---|---|---|
| F1 | Three invoice and bill flows retry in a tight loop on a 429 | Run history: 40-90 repeats in under two minutes | Rate-limit lockout; delayed postings |
| F2 | Two flows have no error path; a failed run leaves no payload | Scenario export: no error handler route | The item cannot be replayed |
| F3 | Slack receives one message per failed item, with no grouping | A Tuesday outage produced 86 messages in 11 minutes | The channel is muted, then silence |
| F4 | A job-create flow is not idempotent on replay | A manual rerun created a second ServiceM8 job | Duplicate jobs and double billing risk |
| F5 | No heartbeat on six overnight flows | A paused Make scheduling switch was noticed at 09:40 | A whole night of work is missed |
| F6 | Credentials for two connections expire with no warning | Vendor email sat in a personal inbox | An outage that looks like a mapping bug |
Approach
- List every automation
Owner, trigger, destination and what a failure costs, written in one inventory.
- Write the rules
Which errors retry, how often, when to park, and who is alerted.
- Build one handler
A shared pattern for Make and for n8n, plus an Airtable queue and heartbeats.
- Roll out and tune
One automation at a time, with a rollback that leaves the original steps in place.
Current state and future state
Current state
Each automation is its own island. A failure is either retried blindly or lost.
Future state
The original steps stay in place. The handler can be removed from one automation without touching the others.
Future-state modules
How it is built
- Module AInventory
Every automation is listed with owner, trigger, destination, last failure and what a silent failure costs.
- Module BClassification
Each error is classed: rate limit, auth, validation, timeout, or unknown.
- Module CRetry policy
Only rate limits and timeouts retry, on a written backoff, with a cap.
- Module DDead-letter queue
Anything else is parked in Airtable with the full payload, the error and an owner.
- Module EIdempotent replay
A successful replay cannot create a second job or invoice from the same payload key.
- Module FAlerts and heartbeats
Critical failures alert at once. A missed heartbeat is treated as a failure of its own.
Field mapping
| Source | Field or signal | Destination | Rule |
|---|---|---|---|
| Make / n8n run | Execution id, scenario or workflow name, timestamp | Airtable: Run key | Becomes the idempotency key |
| Error object | HTTP status, message, node or module name | Airtable: Error type and detail | Mapped to the classification table |
| Original input | Full JSON payload | Airtable: Payload | Stored as received, never trimmed |
| Classification | Retryable or not | Handler path | Retry schedule, or park with no retry |
| Queue row | Status, owner, attempt count | Slack or email | Alert text and link; grouped in a 15 minute window |
| Heartbeat watch | Last successful finish | Slack and email | Alert if the gap exceeds the written maximum |
Worked exhibits
| Error type | Retry? | Schedule | If it still fails |
|---|---|---|---|
| Rate limit (429) | Yes | 5 attempts, 1, 2, 4, 8, 16 min | Parked, High alert |
| Timeout / network | Yes | 3 attempts, 2, 4, 8 min | Parked, High alert |
| Authentication | No | Paused at once | Parked, Critical alert |
| Validation / bad record | No | No retry | Parked, named field, High alert |
| Unknown | No | No retry | Parked, Critical alert |
Exceptions and runbook
| Signal | Threshold | Goes to | First action |
|---|---|---|---|
| Critical failure | Last retry failed on a money or customer flow | Finance email and operations Slack | Open the item, fix the cause, replay once |
| Queue size | More than 10 open, or any older than 24 hours | Operations Slack | Review the same morning |
| Heartbeat missed | Gap above the written maximum | Owner and operations Slack | Check the trigger and the scheduling switch |
| Credential expiry | 14 days before, and on the day | Owner email | Renew and test the connection |
| Handler itself fails | Any | Owner and a nominated second mailbox | Restore the handler, then replay parked items |
How to read the exhibits
- The findings are from the review of this kind of setup, not a claim about a named business.
- The rules table is the contract: rate limits retry, bad records do not.
- The week of figures shows the handler at work: most first-try failures recover, and nothing is created twice.
Test plan
| ID | Check | Method | PASS when | FAIL when |
|---|---|---|---|---|
| T1 | Rate limit | Force a 429 in a copy | Five attempts on the schedule, then success or park | Fewer attempts, or no park |
| T2 | Authentication | Revoke a test token | No retry, paused, Critical alert at once | Retries, or no alert |
| T3 | Validation | Send a bad record | Parked with the field named | Retried or lost |
| T4 | Idempotency | Replay a successful item | No second record in the destination | A duplicate is created |
| T5 | Heartbeat | Switch off a test automation | Alert inside the maximum gap | No alert |
| T6 | Alert grouping | Fail 20 items at once | One grouped message per window | 20 separate messages |
| T7 | Replay after a fix | Fix the cause and replay | Item created once, queue row closed | Item stuck or duplicated |
| T8 | Handler failure | Break the handler in a copy | Second-level email sent | Silence |
Scope
| Area | Included |
|---|---|
| Discovery | Read-only review of every Make scenario and n8n workflow, owners and failure history, written findings |
| Inventory | Every automation with owner, trigger, steps and failure cost |
| Shared error handler | One pattern for Make and one for n8n, with types and retries |
| Dead-letter queue | Airtable table, views, owner and status, with replay steps |
| Alerts and heartbeats | Slack and email by severity, grouped; missed-run watches |
| Test plan | The PASS and FAIL checks in this write-up, run in a copy before go-live |
| Handover and support | Written SOPs, a recorded walkthrough, files you own, 14 days of fixes, and an optional monthly hour block |
| Not included | Rebuilding the automations themselves, a 24-hour on-call roster, or vendor subscriptions |
Rollout, handover and monthly support
| Milestone | Deliverable | Timing | Share of fee |
|---|---|---|---|
| M1 Discovery and rules | Inventory, findings and error rules in writing | Week 1 | 25% |
| M2 Handler and queue | Handler and queue built and tested in a copy | Week 2 | 30% |
| M3 Rollout | Handler live on all automations, heartbeats on | Week 3 | 30% |
| M4 Handover | Alert tuning, runbook, SOPs, 14 days of fixes | Week 4 | 15% |
Handover
- Written operating procedures for the happy path and each exception in the runbook
- A recorded walkthrough of the architecture, the queue and how to replay an item
- A credentials and ownership list, with every connection in your name
- Fourteen days of fixes after go-live, for defects in the agreed scope
Monthly support
A set block of hours each month, at US$35 per hour, agreed in writing before it starts. Typical use of the block: watching the queue, renewing credentials, and adjusting retry windows after a vendor change
- Review of the exception queue and the weekly digest
- Small mapping or rule changes, agreed before they are made
- A short written note on what was done and what is still open
Typical scope for this kind of project: US$1,500-3,000
About 4 weeks · fixed milestones. The systems health check (US$150, fixed) comes first and confirms the scope. Extra hourly work is US$35 per hour, agreed before it starts. After handover, a set block of hours each month is available at the same rate, agreed in writing, to watch the flows, handle changes and work the exception queue.
Discuss an engagement.
Outline the operational problem through the contact form. I respond with a recommended starting point, including whether the US$150 systems health check is the appropriate first step.