Logging preserves what happened — fascinating in hindsight, useless when the client finds the outage first. Monitoring is the active inversion: notifications pushed at you in minutes. Self-hosted n8n operators routinely build enormous log pipelines and still discover a dead order-sync from an angry email two days later. Three layers are what "monitored" actually means.
Layer 1 — Execution failures (the Error Workflow alert)
n8n's built-in mechanism: designate an Error Workflow that fires on every failed execution, route it to Slack or email. Build it, test it with a deliberately failing node, forget it. One warning: the Error Workflow must be rock-solid itself — if it silently breaks, your alert system is a dead load. Monitor the monitor.
Layer 2 — Infrastructure health
Executions only fail if execution happens; the traps below the canvas produce no red rows. In queue mode, that means: living worker count, queue depth, oldest-job age, broker connectivity. A five-line health endpoint plus an external uptime checker covers it. Route these alerts to the infra human, not the workflow owner.
Layer 3 — Absence-of-signal (nobody builds this)
The asymmetric truth: most catastrophic failures never produce an error. The trigger broke before execution; the webhook route died; the workflow went inactive. Nothing happened, so nothing failed. Layer 3 is the dead-man switch: every critical workflow pings an external heartbeat on completion, and the service alerts when the expected heartbeat goes quiet. You are now monitoring the absence of a signal — the exact signature of every silent production killer.
The two-sided defense contract
All three layers see only what happens after deploy. Nothing prevents a broken workflow from shipping — that side is the preflight: sixty seconds with the free n8n workflow audit per export kills five killers before your monitoring has to chase them. Preflight at dusk, monitor through the night.
Top comments (0)