AI & Production · 18 Nov 2026 · 9 min read
n8n Production Checklist: 20 Things to Check Before Going Live
An n8n production checklist covering error handling, environments, credentials, scaling, backups, monitoring and documentation, with the plan limits that matter on n8n Cloud.

A workflow that ran once when someone clicked Execute is a prototype, and the twenty checks below are what we look at before we call one production.
On n8n Cloud the plan itself fixes several limits that matter at go live: Starter ($20 a month, billed annually) caps one execution at 5 minutes and runs 5 executions at once, while Pro ($50 a month) allows 40 minutes and 20 to 50 concurrent executions, which is 8 times the duration and 4 to 10 times the concurrency of Starter¹. Git based version control, multiple environments and queue mode sit in the self hosted Business plan (667 euros a month) and in Enterprise¹, so a team on Starter or Pro has to cover change control some other way, for instance with scheduled JSON exports. The strongest independent figure we found concerns defaults: in an outage probe by the measurement site BenchTruth (30 minutes of 503 responses from the receiving endpoint, five test events per platform, August 2026), a self hosted n8n install on default settings made no automatic retry and sent no notification, and its accepted events were recovered by a manual retry about 14 hours later, each delivered once, while Make retried six times and delivered everything after 79 minutes⁷. The same source recorded no silent failures in 14,161 n8n webhook runs (upper 95% bound 0.03%), so what n8n accepts it delivers, and the gap lies in telling you about everything else. These are one site’s measurements on its own hardware (a 1 GB virtual machine), with only five events per platform in the outage test, and we have not reproduced them. We also have no benchmark tying our twenty checks to fewer incidents; the ordering below reflects our experience on client builds, and more detail on each item follows.
Reliability
A production workflow has to tell a human when it fails, and most of the first six checks serve that single purpose.
Error workflow (1). Attach one to every production workflow; it starts from the Error Trigger node and should name the workflow, the failed node and a link to the execution, then post to a channel someone reads²; the longer treatment is in n8n error handling, and BenchTruth’s default install notified nobody until an Error Trigger workflow was configured⁷.
Retries (2). Turn them on for calls to external APIs and leave them off for writes that are not safe to repeat; per node Retry On Fail ships switched off⁷, so it has to be a decision.
Validation (3). Check required fields and types at the start of the run, and use the Stop and Error node³ for data that no API would reject but that you know is wrong.
Idempotency (4). Upsert on a stable ID and keep a record of processed IDs, so that a replayed execution has no second effect (an email sent twice is the usual symptom when this is missing). BenchTruth lost 2 of 524 new items on a polling workflow after an n8n upgrade (2.42.4) added a 5 second default limit on expression evaluation that its memory starved 1 GB machine exceeded; the workflow marked each item as seen before sending it, which made the loss permanent, and the authors attribute part of the fault to that design⁷.
Limits (5). Split any job that can approach the plan’s execution cap, which is the 5 minute Starter figure above.
Heartbeat (6). A scheduled workflow that alerts when something expected did not happen, such as a day with no leads, catches the case where nothing failed because nothing ran.
Image: Pixabay (free to use)
Change control
Anything that cannot be rebuilt from a repository is a liability on the day an instance is lost.
Version control (7). Use n8n’s source control and environments where your plan includes them⁴, and where it does not, export workflows to JSON on a schedule and commit the files.
Environments (8). Keep test and production apart, at minimum with separate credentials and destinations, since a test run that emails real customers is a common first incident.
Naming (9). System, purpose and trigger in the name, with tags for owner and environment.
Notes (10). One line on every node whose logic is not obvious, Code nodes and expressions first.
Security
Credentials and webhooks are where a workflow touches the outside world, so the checks here are blunt.
Least privilege (11). One credential per purpose with the narrowest scopes, and no personal admin tokens.
Secrets (12). Keys belong in n8n credentials or environment variables, never in a Set node or a note.
Webhooks (13). Require a header token or a signature check, because a public URL with no check accepts anything sent to it.
Data minimisation (14). Pass only the fields a step needs and prune stored execution data on self hosted installs, where n8n enables pruning by default and deletes finished executions older than 336 hours (14 days) or beyond 10,000 executions, though never annotated ones⁵; Starter keeps seven days of execution logs¹, which bounds exposure on Cloud.
Personal data (15). Know which workflows touch it, where it is stored and who can open the executions.
Self hosted infrastructure
This section applies only if you run the instance yourself; on Cloud, n8n carries most of it, and the trade is discussed in n8n self hosted vs cloud and, for cost, n8n pricing in 2026.
Backups (16). Back up the database and the encryption key, and restore once into a scratch instance; n8n stores credentials encrypted, and a restored database or credential export cannot be decrypted without the key from the config file or a custom N8N_ENCRYPTION_KEY⁸; in our experience the restore is the step teams skip.
Queue mode (17). Use it when one instance cannot keep up with execution volume⁶; on a 1 GB machine a burst of 10 events once stalled BenchTruth’s n8n for roughly 8 seconds and one request was refused at the door, which a sender that checks response codes can retry⁷.
Updates (18). Pin versions, read the release notes and upgrade in staging first; the 2.42.4 change above altered a default timeout, and a staging run is, in our view, where that kind of difference should surface.
Monitoring (19). Watch CPU, memory and disk, plus an external check that the instance answers.
People
Owner and runbook (20). One paragraph per workflow covering what it does, its trigger, who owns it, what to do when it fails and how to switch it off; for anything that moves money or contacts customers, add a short recorded walkthrough. This is also the item that decides whether a freelancer’s handover works (compare options in freelancer vs agency vs DIY).
Extra checks when a model makes decisions
A language model adds failure modes that the twenty checks do not cover. Decisions touching money, customers or contracts need a confidence threshold and a review queue (see human in the loop AI agents); ask for structured output against a fixed JSON schema so the next step never parses free text; cap tokens, retries and loop depth; keep a set of real examples and rerun it whenever the prompt or model changes (monitoring and testing AI automations); and log inputs and outputs with privacy in mind.
Testing before go live
Six tests find most of what the checks miss, and each is cheap to run: send empty and malformed input and see whether validation stops it; send the same event twice and look for a duplicate write; disable a credential and watch for the alert; push a large or slow payload through; restart the instance mid run and see whether work is lost or repeated; hand the runbook to a colleague and see whether they can operate the workflow unaided. We have not measured how many defects each test catches, so treat the list as ours and not as established practice.
Scoring yourself
Fewer than 9 boxes ticked is a prototype and should stay away from customers and money; 9 to 15 suits internal and low risk flows if you have a plan to close the gaps; 16 or more is where we are comfortable with revenue critical work. These bands are our own rule of thumb and we have not validated them against incident data.
The list is free to copy into a ticket template or a pull request checklist, and if you would like a second pair of eyes on a specific workflow, the free 30 minute audit starts with the first three sections above and takes one workflow of yours as the example.
- n8n plans and pricing, checked 5 November 2026; plan limits change, so confirm at n8n.io/pricing before relying on a figure.
- n8n docs, Error Trigger node.
- n8n docs, Stop and Error node.
- n8n docs, source control and environments; availability depends on plan.
- n8n docs, manage execution data; defaults as read on 18 November 2026.
- n8n docs, queue mode.
- BenchTruth, Silent Failure Rate, measured, read 18 November 2026. An independent site’s own methodology: webhook workflows with both endpoints under its control, n8n self hosted on a 1 GB virtual machine, Make and Zapier on paid plans, an outage probe of five events per platform; the figures describe those plans, hardware and windows, say nothing about n8n Cloud, and the site notes that its n8n tail latency (p95 9.72 s against a 1.07 s median) mostly reflects its own small machine.
- n8n docs, back up and restore.