Apr 2026 – Sep 2026
n8n platform reliability
A self-hosted n8n with over 100 workflows, kept healthy, observable and audited on evidence
- Automation
- Platform operations
- Duration
- 5 months
- Company
- Boxette Group
- Role
- Platform owner and operator
Overview
Boxette’s self-hosted n8n runs in queue mode with a worker, Postgres, Redis and Docker on a VPS. The work keeps it healthy and observable (execution history, error alerts, worker capacity) and audits what is actually failing, with evidence.
Tools
- Docker
- PostgreSQL
- Redis
- Nginx Proxy Manager
Failures nobody could see
Over 100 workflows run on one server, some every 5 seconds. Execution history vanished within hours, some workflows showed success while failing inside, and data errors were blamed on workflows without proof.
Proof before fixes
When the e-signing workflow was said to assign the wrong branch, 1,443 tracking codes were checked against the parcel system and the stored execution data was decoded. That proved the workflow correct: the source system reassigns branches later, so a daily updater workflow now fixes the sheet. Lost history turned out to come from an execution cap that filled in about 14 hours; after measuring the disk cost per execution, success logging was switched off for the five noisiest workflows and the cap was raised.
Capacity, alerts and audits
Worker concurrency was raised for the evening peak, and a shared Telegram error workflow and retries were attached across the critical workflows. In September 2026, Claude Code subagents (collector, infrastructure, root cause, verifier) ran a 7-day read-only audit under my direction. They went through executions, Postgres, Docker, nginx and Redis and ranked the root causes by priority, each backed by execution evidence.
Result
Execution writes fell from about 57 a minute to about 2.4, roughly 24 times fewer, which stretched the history from about 14 hours to about 3 months. In the branch check, 1,253 of 1,443 codes (86.8%) matched and the other 190 had changed later in the source system, with 0 workflow bugs. The 7-day audit covered 52,190 executions with 153 recorded errors and 0 crashes, and the verifier confirmed all 7 top findings.